Spaces:
Running
Running
added comparison of different repetition detection methods
Browse files- eval_modules/calc_repetitions.py +65 -0
- eval_modules/calc_repetitions_v1.py +929 -0
- eval_modules/calc_repetitions_v2.py +1087 -0
- eval_modules/calc_repetitions_v3.py +1095 -0
- eval_modules/calc_repetitions_v4.py +1296 -0
- eval_modules/calc_repetitions_v5.py +1383 -0
- notebooks/00_Repetition_Algorithms_Comparison.ipynb +0 -0
- notebooks/04_RAPGeT_v2.ipynb +0 -0
eval_modules/calc_repetitions.py
CHANGED
@@ -7,6 +7,9 @@ import matplotlib.pyplot as plt
|
|
7 |
import matplotlib.ticker as mtick
|
8 |
import seaborn as sns
|
9 |
import nltk
|
|
|
|
|
|
|
10 |
|
11 |
print(f"loading: {__file__}")
|
12 |
|
@@ -1316,3 +1319,65 @@ def load_ms_marco_result(csv_result_files, force_recalculate=False):
|
|
1316 |
print(f"Error: {e}")
|
1317 |
|
1318 |
return result
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
7 |
import matplotlib.ticker as mtick
|
8 |
import seaborn as sns
|
9 |
import nltk
|
10 |
+
import evaluate
|
11 |
+
|
12 |
+
meteor = evaluate.load("meteor")
|
13 |
|
14 |
print(f"loading: {__file__}")
|
15 |
|
|
|
1319 |
print(f"Error: {e}")
|
1320 |
|
1321 |
return result
|
1322 |
+
|
1323 |
+
|
1324 |
+
def load_ms_marco_result_v2(csv_result_files, force_recalculate=False):
|
1325 |
+
model_name_exts = {
|
1326 |
+
"true": "(RAG - Chat Template)",
|
1327 |
+
"false": "(RAG - Generic Prompt)",
|
1328 |
+
"rag": "(Non-RAG)",
|
1329 |
+
}
|
1330 |
+
|
1331 |
+
result = {}
|
1332 |
+
for csv_result_file in csv_result_files:
|
1333 |
+
try:
|
1334 |
+
df = pd.read_csv(csv_result_file)
|
1335 |
+
|
1336 |
+
parts = re.split(r"[_\.]", csv_result_file)
|
1337 |
+
model_name = f'{df["model"][0]}{model_name_exts[parts[-2]]}'
|
1338 |
+
|
1339 |
+
print(f"\tmodel_name: {model_name}")
|
1340 |
+
dfs = [
|
1341 |
+
load_for_repetition_penalty_ms_macro(
|
1342 |
+
csv_result_file,
|
1343 |
+
repetition_penalty,
|
1344 |
+
force_recalculate=force_recalculate,
|
1345 |
+
)
|
1346 |
+
for repetition_penalty in df["repetition_penalty"]
|
1347 |
+
]
|
1348 |
+
|
1349 |
+
answer_lens = []
|
1350 |
+
for df_rpp in dfs:
|
1351 |
+
df_rpp["answer_len"] = df_rpp["answer"].apply(
|
1352 |
+
lambda x: len(x) if isinstance(x, str) else 0
|
1353 |
+
)
|
1354 |
+
answer_lens.append(df_rpp["answer_len"].mean())
|
1355 |
+
df["answer_len"] = answer_lens
|
1356 |
+
|
1357 |
+
meteor_scores = []
|
1358 |
+
for df_rpp in dfs:
|
1359 |
+
meteor_score = meteor.compute(
|
1360 |
+
predictions=df_rpp["answer"], references=df_rpp["ground_truth"]
|
1361 |
+
)["meteor"]
|
1362 |
+
meteor_scores.append(meteor_score)
|
1363 |
+
df["meteor_scores"] = meteor_scores
|
1364 |
+
|
1365 |
+
result[model_name] = {
|
1366 |
+
"df_overall": df,
|
1367 |
+
"df_list_repetition_penalty": dfs,
|
1368 |
+
"file": csv_result_file,
|
1369 |
+
}
|
1370 |
+
newline_score, repetition_score, perf, rap = calc_rap_scores(
|
1371 |
+
result[model_name],
|
1372 |
+
precision="meteor_scores",
|
1373 |
+
recall="meteor_scores",
|
1374 |
+
)
|
1375 |
+
df["newline_score"] = newline_score
|
1376 |
+
df["repetition_score"] = repetition_score
|
1377 |
+
df["total_repetitions"] = df["newline_score"] + df["repetition_score"]
|
1378 |
+
df["perf"] = perf
|
1379 |
+
df["rap"] = rap
|
1380 |
+
except Exception as e:
|
1381 |
+
print(f"Error: {e}")
|
1382 |
+
|
1383 |
+
return result
|
eval_modules/calc_repetitions_v1.py
ADDED
@@ -0,0 +1,929 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
1 |
+
import re
|
2 |
+
import math
|
3 |
+
import pandas as pd
|
4 |
+
import numpy as np
|
5 |
+
import matplotlib.pyplot as plt
|
6 |
+
import matplotlib.ticker as mtick
|
7 |
+
import seaborn as sns
|
8 |
+
|
9 |
+
# final version
|
10 |
+
pattern_abnormal_newlines = re.compile(r"\n{5,}")
|
11 |
+
pattern_text_repetitions = re.compile(r"\b(\w.+?)\b(\1+)", re.M | re.DOTALL)
|
12 |
+
exception_pattern = re.compile(r"(\w+\.)\1")
|
13 |
+
|
14 |
+
|
15 |
+
# final version for repetition detection
|
16 |
+
def detect_repetitions(
|
17 |
+
text, debug=False, pattern_text_repetitions=pattern_text_repetitions
|
18 |
+
):
|
19 |
+
subtotals = [0, 0]
|
20 |
+
|
21 |
+
if isinstance(text, str):
|
22 |
+
patterns = [pattern_abnormal_newlines, pattern_text_repetitions]
|
23 |
+
for i, pattern in enumerate(patterns):
|
24 |
+
if debug:
|
25 |
+
print(
|
26 |
+
f"----detect {'abnormal newlines' if i == 0 else 'text repetitions'}----"
|
27 |
+
)
|
28 |
+
matches = pattern.finditer(text)
|
29 |
+
for match in matches:
|
30 |
+
if debug:
|
31 |
+
print(match)
|
32 |
+
for groupNum in range(0, len(match.groups())):
|
33 |
+
groupNum = groupNum + 1
|
34 |
+
print(
|
35 |
+
"Group {groupNum} found at {start}-{end}: `{group}`".format(
|
36 |
+
groupNum=groupNum,
|
37 |
+
start=match.start(groupNum),
|
38 |
+
end=match.end(groupNum),
|
39 |
+
group=match.group(groupNum),
|
40 |
+
)
|
41 |
+
)
|
42 |
+
|
43 |
+
if exception_pattern.match(match[0]):
|
44 |
+
if debug:
|
45 |
+
print("ignored: ", match[0])
|
46 |
+
continue
|
47 |
+
|
48 |
+
start, end = match.span()
|
49 |
+
subtotals[i] += end - start
|
50 |
+
|
51 |
+
result = (subtotals[0], subtotals[1], subtotals[0] + subtotals[1])
|
52 |
+
|
53 |
+
if debug:
|
54 |
+
print(result)
|
55 |
+
return result
|
56 |
+
|
57 |
+
|
58 |
+
def detect_abnormal_newlines(text, debug=False):
|
59 |
+
return detect_repetitions(text, debug=debug)[0]
|
60 |
+
|
61 |
+
|
62 |
+
def detect_text_repetitions(text, debug=False):
|
63 |
+
return detect_repetitions(text, debug=debug)[1]
|
64 |
+
|
65 |
+
|
66 |
+
def detect_scores(text, debug=False):
|
67 |
+
newline_score, repetition_score, total_repetitions = detect_repetitions(
|
68 |
+
text, debug=debug
|
69 |
+
)
|
70 |
+
return pd.Series([newline_score, repetition_score, total_repetitions])
|
71 |
+
|
72 |
+
|
73 |
+
def load_with_newline_and_repetition_scores(result_file, force_recalculate=False):
|
74 |
+
print(f"loading result file: {result_file}")
|
75 |
+
df = pd.read_csv(result_file, comment="#", on_bad_lines="warn")
|
76 |
+
|
77 |
+
if (
|
78 |
+
force_recalculate
|
79 |
+
or "newline_score" not in df.columns
|
80 |
+
or "repetition_score" not in df.columns
|
81 |
+
or "total_repetitions" not in df.columns
|
82 |
+
):
|
83 |
+
df[["newline_score", "repetition_score", "total_repetitions"]] = df[
|
84 |
+
"answer"
|
85 |
+
].apply(detect_scores)
|
86 |
+
df.to_csv(result_file, index=False)
|
87 |
+
|
88 |
+
return df
|
89 |
+
|
90 |
+
|
91 |
+
def replace_last(source_string, old_string, new_string):
|
92 |
+
head, _sep, tail = source_string.rpartition(old_string)
|
93 |
+
return head + new_string + tail
|
94 |
+
|
95 |
+
|
96 |
+
def load_for_repetition_penalty(
|
97 |
+
csv_result_file, repetition_penalty, force_recalculate=False
|
98 |
+
):
|
99 |
+
result_file = replace_last(
|
100 |
+
csv_result_file, ".csv", f"_RP_{repetition_penalty:.3f}.csv"
|
101 |
+
)
|
102 |
+
return load_with_newline_and_repetition_scores(
|
103 |
+
result_file, force_recalculate=force_recalculate
|
104 |
+
)
|
105 |
+
|
106 |
+
|
107 |
+
def calc_adjusted_performance(f, r):
|
108 |
+
return f / math.log10(10 + r)
|
109 |
+
|
110 |
+
|
111 |
+
def calculate_adjusted_performance(row):
|
112 |
+
r = row["total_repetitions"]
|
113 |
+
adjusted_precision = calc_adjusted_performance(row["precision"], r)
|
114 |
+
adjusted_recall = calc_adjusted_performance(row["recall"], r)
|
115 |
+
return pd.Series([adjusted_precision, adjusted_recall])
|
116 |
+
|
117 |
+
|
118 |
+
def load_performance_df(csv_result_file, repetition_penalty):
|
119 |
+
result_file = replace_last(
|
120 |
+
csv_result_file, ".csv", f"_RP_{repetition_penalty:.3f}-t2_evaluated.json"
|
121 |
+
)
|
122 |
+
result_file = result_file.replace("/results/", "/eval/")
|
123 |
+
print(f"loading json file: {result_file}")
|
124 |
+
df = pd.read_json(result_file)
|
125 |
+
|
126 |
+
return df
|
127 |
+
|
128 |
+
|
129 |
+
def calculate_performance_score(
|
130 |
+
csv_result_file, repetition_penalty, force_recalculate=False
|
131 |
+
):
|
132 |
+
result_file = replace_last(
|
133 |
+
csv_result_file, ".csv", f"_RP_{repetition_penalty:.3f}.csv"
|
134 |
+
)
|
135 |
+
print(f"loading result file: {result_file}")
|
136 |
+
df = pd.read_csv(result_file, comment="#", on_bad_lines="warn")
|
137 |
+
|
138 |
+
if force_recalculate or "f2" in df.columns or "f1" not in df.columns:
|
139 |
+
df.drop(
|
140 |
+
columns=[
|
141 |
+
"precision",
|
142 |
+
"recall",
|
143 |
+
"f1",
|
144 |
+
"f2",
|
145 |
+
"entities_in_answer",
|
146 |
+
"entities_in_question",
|
147 |
+
],
|
148 |
+
errors="ignore",
|
149 |
+
inplace=True,
|
150 |
+
)
|
151 |
+
perf_df = load_performance_df(csv_result_file, repetition_penalty)
|
152 |
+
filtered_df = perf_df[perf_df["id"].isin(df["id"])]
|
153 |
+
perf_df = filtered_df.reset_index(drop=True)
|
154 |
+
print(f"perf_df len: {len(perf_df)}")
|
155 |
+
# print(perf_df.head())
|
156 |
+
|
157 |
+
df["eval_gemini_1.0_pro"] = perf_df["eval_gemini_1.0_pro"]
|
158 |
+
|
159 |
+
df["precision"] = perf_df["score"].apply(lambda x: x[0])
|
160 |
+
df["recall"] = perf_df["score"].apply(lambda x: x[1])
|
161 |
+
|
162 |
+
df[["adjusted_precision", "adjusted_recall"]] = df.apply(
|
163 |
+
calculate_adjusted_performance, axis=1
|
164 |
+
)
|
165 |
+
|
166 |
+
df.to_csv(result_file, index=False)
|
167 |
+
print(f"performance scores saved to result file: {result_file}")
|
168 |
+
|
169 |
+
print(f"df len: {len(df)}")
|
170 |
+
|
171 |
+
return df
|
172 |
+
|
173 |
+
|
174 |
+
def adjust_perf_scores_with_repetition_penalty(result, precision, recall):
|
175 |
+
newline_score = [
|
176 |
+
df["newline_score"].mean() for df in result["df_list_repetition_penalty"]
|
177 |
+
]
|
178 |
+
print(f"newline_score: {newline_score}")
|
179 |
+
|
180 |
+
repetition_score = [
|
181 |
+
df["repetition_score"].mean() for df in result["df_list_repetition_penalty"]
|
182 |
+
]
|
183 |
+
print(f"repetition_score: {repetition_score}")
|
184 |
+
|
185 |
+
precision = [
|
186 |
+
f / math.log10(10 + n + r)
|
187 |
+
for f, n, r in zip(precision, newline_score, repetition_score)
|
188 |
+
]
|
189 |
+
recall = [
|
190 |
+
f / math.log10(10 + n + r)
|
191 |
+
for f, n, r in zip(recall, newline_score, repetition_score)
|
192 |
+
]
|
193 |
+
|
194 |
+
return precision, recall
|
195 |
+
|
196 |
+
|
197 |
+
def plot_performance_scores(
|
198 |
+
result,
|
199 |
+
models=None,
|
200 |
+
title="Performance",
|
201 |
+
):
|
202 |
+
|
203 |
+
if models is None:
|
204 |
+
models = result.keys()
|
205 |
+
for model in models:
|
206 |
+
print(f"model: {model}")
|
207 |
+
df = result[model]["df_overall"]
|
208 |
+
|
209 |
+
# Calculate the statistics
|
210 |
+
precision = [
|
211 |
+
df["precision"].mean() for df in result[model]["df_list_repetition_penalty"]
|
212 |
+
]
|
213 |
+
recall = [
|
214 |
+
df["recall"].mean() for df in result[model]["df_list_repetition_penalty"]
|
215 |
+
]
|
216 |
+
f1 = [2 * (p * r) / (p + r) for p, r in zip(precision, recall)]
|
217 |
+
best_f1 = max(f1)
|
218 |
+
best_f1_index = f1.index(best_f1)
|
219 |
+
|
220 |
+
precision, recall = adjust_perf_scores_with_repetition_penalty(
|
221 |
+
result[model], precision, recall
|
222 |
+
)
|
223 |
+
afrp = [2 * (p * r) / (p + r) for p, r in zip(precision, recall)]
|
224 |
+
|
225 |
+
# f1 = [df["f1"].mean() for df in result[model]["df_list_repetition_penalty"]]
|
226 |
+
best_afrp = max(afrp)
|
227 |
+
best_afrp_index = afrp.index(best_afrp)
|
228 |
+
|
229 |
+
adjusted_precision = [
|
230 |
+
df["adjusted_precision"].mean()
|
231 |
+
for df in result[model]["df_list_repetition_penalty"]
|
232 |
+
]
|
233 |
+
adjusted_recall = [
|
234 |
+
df["adjusted_recall"].mean()
|
235 |
+
for df in result[model]["df_list_repetition_penalty"]
|
236 |
+
]
|
237 |
+
afrp2 = [
|
238 |
+
2 * (p * r) / (p + r) for p, r in zip(adjusted_precision, adjusted_recall)
|
239 |
+
]
|
240 |
+
best_afrp2 = max(afrp2)
|
241 |
+
best_afrp2_index = afrp2.index(best_afrp2)
|
242 |
+
|
243 |
+
repetition_penalties = list(df["repetition_penalty"])
|
244 |
+
|
245 |
+
# line plot for precision, recall, f1
|
246 |
+
plt.figure(figsize=(10, 6))
|
247 |
+
|
248 |
+
plt.axvspan(
|
249 |
+
repetition_penalties[best_f1_index] - 0.01,
|
250 |
+
repetition_penalties[best_f1_index] + 0.01,
|
251 |
+
alpha=0.5,
|
252 |
+
edgecolor="none",
|
253 |
+
facecolor="blue",
|
254 |
+
)
|
255 |
+
|
256 |
+
plt.axvspan(
|
257 |
+
repetition_penalties[best_afrp2_index] - 0.01,
|
258 |
+
repetition_penalties[best_afrp2_index] + 0.01,
|
259 |
+
alpha=0.5,
|
260 |
+
edgecolor="none",
|
261 |
+
facecolor="green",
|
262 |
+
)
|
263 |
+
|
264 |
+
plt.axvspan(
|
265 |
+
repetition_penalties[best_afrp_index] - 0.01,
|
266 |
+
repetition_penalties[best_afrp_index] + 0.01,
|
267 |
+
alpha=0.5,
|
268 |
+
edgecolor="none",
|
269 |
+
facecolor="orange",
|
270 |
+
)
|
271 |
+
|
272 |
+
plt.plot(repetition_penalties, f1, label="F1", marker="D", color="blue")
|
273 |
+
plt.plot(
|
274 |
+
repetition_penalties,
|
275 |
+
afrp2,
|
276 |
+
label="Per-question RF Adjusted F1",
|
277 |
+
marker="s",
|
278 |
+
color="green",
|
279 |
+
)
|
280 |
+
plt.plot(
|
281 |
+
repetition_penalties,
|
282 |
+
afrp,
|
283 |
+
label="Overall RF Adjusted F1",
|
284 |
+
marker="o",
|
285 |
+
color="orange",
|
286 |
+
)
|
287 |
+
plt.xlabel("Repetition Penalties")
|
288 |
+
plt.ylabel("Score")
|
289 |
+
plt.xlim(0.99, 1.31)
|
290 |
+
# y in percentage
|
291 |
+
plt.gca().yaxis.set_major_formatter(mtick.PercentFormatter(1.0))
|
292 |
+
plt.title(f"{model} {title}")
|
293 |
+
plt.legend(bbox_to_anchor=(1.0, 0.5), loc="center left")
|
294 |
+
|
295 |
+
plt.show()
|
296 |
+
|
297 |
+
|
298 |
+
def plot_best_afrp(
|
299 |
+
result,
|
300 |
+
models=None,
|
301 |
+
title="Models with Best Repetition Factor Adjusted F1",
|
302 |
+
ref_result=None,
|
303 |
+
):
|
304 |
+
# Initialize lists to store the statistics
|
305 |
+
model_names = []
|
306 |
+
best_f1 = []
|
307 |
+
best_afrp = []
|
308 |
+
best_repetition_penalty = []
|
309 |
+
|
310 |
+
if models is None:
|
311 |
+
models = result.keys()
|
312 |
+
for model in models:
|
313 |
+
print(f"model: {model}")
|
314 |
+
df = result[model]["df_overall"]
|
315 |
+
|
316 |
+
# Calculate the statistics
|
317 |
+
precision = [
|
318 |
+
df["precision"].mean() for df in result[model]["df_list_repetition_penalty"]
|
319 |
+
]
|
320 |
+
recall = [
|
321 |
+
df["recall"].mean() for df in result[model]["df_list_repetition_penalty"]
|
322 |
+
]
|
323 |
+
# f1 = [df["f1"].mean() for df in result[model]["df_list_repetition_penalty"]]
|
324 |
+
f1 = [2 * (p * r) / (p + r) for p, r in zip(precision, recall)]
|
325 |
+
|
326 |
+
newline_score = [
|
327 |
+
df["newline_score"].mean()
|
328 |
+
for df in result[model]["df_list_repetition_penalty"]
|
329 |
+
]
|
330 |
+
print(f"newline_score: {newline_score}")
|
331 |
+
|
332 |
+
repetition_score = [
|
333 |
+
df["repetition_score"].mean()
|
334 |
+
for df in result[model]["df_list_repetition_penalty"]
|
335 |
+
]
|
336 |
+
print(f"repetition_score: {repetition_score}")
|
337 |
+
|
338 |
+
afrp = [
|
339 |
+
f / math.log10(10 + n + r)
|
340 |
+
for f, n, r in zip(f1, newline_score, repetition_score)
|
341 |
+
]
|
342 |
+
|
343 |
+
best_afrp.append(max(afrp))
|
344 |
+
best_afrp_index = afrp.index(best_afrp[-1])
|
345 |
+
best_repetition_penalty.append(df["repetition_penalty"][best_afrp_index])
|
346 |
+
|
347 |
+
best_f1.append(f1[best_afrp_index])
|
348 |
+
|
349 |
+
print(
|
350 |
+
f"best repetition penalty: {best_repetition_penalty[-1]}, best afrp: {best_afrp[-1]}, f1: {best_f1[-1]}"
|
351 |
+
)
|
352 |
+
|
353 |
+
df = result[model]["df_list_repetition_penalty"][best_afrp_index]
|
354 |
+
|
355 |
+
model_names.append(
|
356 |
+
f"{model} (RP={best_repetition_penalty[-1]})"
|
357 |
+
) # Add the model name to the list
|
358 |
+
|
359 |
+
if ref_result is not None:
|
360 |
+
print("ref_result:", ref_result)
|
361 |
+
for model in ref_result.keys():
|
362 |
+
model_names.append(model)
|
363 |
+
df = pd.read_csv(ref_result[model])
|
364 |
+
# df = df[df["id"].isin(wikidata_df["id"])]
|
365 |
+
|
366 |
+
p = df["precision"].mean()
|
367 |
+
r = df["recall"].mean()
|
368 |
+
|
369 |
+
f1 = 2 * p * r / (p + r) if p + r > 0 else 0
|
370 |
+
best_f1.append(f1)
|
371 |
+
best_afrp.append(f1)
|
372 |
+
|
373 |
+
print("model_names:", model_names)
|
374 |
+
print("best_f1:", best_f1)
|
375 |
+
print("best_afrp:", best_afrp)
|
376 |
+
|
377 |
+
# Create a DataFrame with the statistics
|
378 |
+
data = pd.DataFrame(
|
379 |
+
{
|
380 |
+
"Model": model_names,
|
381 |
+
"Repetition Factor Adjusted F1": best_afrp,
|
382 |
+
"F1": best_f1,
|
383 |
+
}
|
384 |
+
)
|
385 |
+
|
386 |
+
# Melt the DataFrame to a long format
|
387 |
+
data_melted = data.melt(id_vars="Model", var_name="Metric", value_name="Score")
|
388 |
+
|
389 |
+
# Pivot the DataFrame to a wide format
|
390 |
+
data_pivoted = data_melted.pivot(index="Metric", columns="Model", values="Score")
|
391 |
+
|
392 |
+
# make sure the columns are following the order of the models
|
393 |
+
data_pivoted = data_pivoted[model_names]
|
394 |
+
|
395 |
+
# make sure three groups in the order of precision, recall, f1
|
396 |
+
data_pivoted = data_pivoted.reindex(["Repetition Factor Adjusted F1", "F1"])
|
397 |
+
|
398 |
+
# Plot the statistics
|
399 |
+
plt.figure(figsize=(10, 6))
|
400 |
+
ax = data_pivoted.plot(kind="bar", ax=plt.gca(), width=0.9)
|
401 |
+
plt.title(title)
|
402 |
+
plt.legend(bbox_to_anchor=(1.0, 0.5), loc="center left")
|
403 |
+
|
404 |
+
# Set the rotation of the x-axis labels to 0 degrees
|
405 |
+
plt.xticks(rotation=0)
|
406 |
+
|
407 |
+
# Format the y-axis to display as percentage
|
408 |
+
ax.yaxis.set_major_formatter(mtick.PercentFormatter(1.0))
|
409 |
+
|
410 |
+
# get the max value of the y-axis
|
411 |
+
a1 = max(best_afrp)
|
412 |
+
a2 = max(best_f1)
|
413 |
+
|
414 |
+
max_value = max([a1, a2]) * 1.12
|
415 |
+
print("max_value:", max_value)
|
416 |
+
|
417 |
+
# Set the y-axis limit up to 70%
|
418 |
+
ax.set_ylim(0, max_value)
|
419 |
+
|
420 |
+
# Add the values above each bar
|
421 |
+
for p in ax.patches:
|
422 |
+
ax.annotate(
|
423 |
+
f"{p.get_height() * 100:.1f}",
|
424 |
+
(p.get_x() + p.get_width() / 2.0, p.get_height()),
|
425 |
+
ha="center",
|
426 |
+
va="bottom",
|
427 |
+
xytext=(0, 10),
|
428 |
+
textcoords="offset points",
|
429 |
+
rotation=90,
|
430 |
+
)
|
431 |
+
|
432 |
+
plt.show()
|
433 |
+
|
434 |
+
|
435 |
+
def plot_best_performance(
|
436 |
+
result,
|
437 |
+
models=None,
|
438 |
+
title="Models with Best F1 Score",
|
439 |
+
adjusted_f1=False,
|
440 |
+
ref_result=None,
|
441 |
+
):
|
442 |
+
# Initialize lists to store the statistics
|
443 |
+
model_names = []
|
444 |
+
best_precision = []
|
445 |
+
best_recall = []
|
446 |
+
best_f1 = []
|
447 |
+
best_repetition_penalty = []
|
448 |
+
|
449 |
+
if models is None:
|
450 |
+
models = result.keys()
|
451 |
+
for model in models:
|
452 |
+
print(f"model: {model}")
|
453 |
+
df = result[model]["df_overall"]
|
454 |
+
|
455 |
+
# Calculate the statistics
|
456 |
+
precision = [
|
457 |
+
df["precision"].mean() for df in result[model]["df_list_repetition_penalty"]
|
458 |
+
]
|
459 |
+
recall = [
|
460 |
+
df["recall"].mean() for df in result[model]["df_list_repetition_penalty"]
|
461 |
+
]
|
462 |
+
|
463 |
+
if adjusted_f1:
|
464 |
+
precision, recall = adjust_perf_scores_with_repetition_penalty(
|
465 |
+
result[model], precision, recall
|
466 |
+
)
|
467 |
+
|
468 |
+
# f1 = [df["f1"].mean() for df in result[model]["df_list_repetition_penalty"]]
|
469 |
+
f1 = [2 * (p * r) / (p + r) for p, r in zip(precision, recall)]
|
470 |
+
|
471 |
+
best_f1.append(max(f1))
|
472 |
+
best_f1_index = f1.index(best_f1[-1])
|
473 |
+
best_repetition_penalty.append(df["repetition_penalty"][best_f1_index])
|
474 |
+
|
475 |
+
best_precision.append(precision[best_f1_index])
|
476 |
+
best_recall.append(recall[best_f1_index])
|
477 |
+
|
478 |
+
print(
|
479 |
+
f"best repetition penalty: {best_repetition_penalty[-1]}, best f1: {best_f1[-1]}, precision: {best_precision[-1]}, recall: {best_recall[-1]}"
|
480 |
+
)
|
481 |
+
|
482 |
+
df = result[model]["df_list_repetition_penalty"][best_f1_index]
|
483 |
+
|
484 |
+
model_names.append(
|
485 |
+
f"{model} (RP={best_repetition_penalty[-1]})"
|
486 |
+
) # Add the model name to the list
|
487 |
+
|
488 |
+
# print sum for columns: newline_score, repetition_score
|
489 |
+
print(
|
490 |
+
f"newline_score: {df['newline_score'].sum()}, repetition_score: {df['repetition_score'].sum()}"
|
491 |
+
)
|
492 |
+
|
493 |
+
if ref_result is not None:
|
494 |
+
print("ref_result:", ref_result)
|
495 |
+
for model in ref_result.keys():
|
496 |
+
model_names.append(model)
|
497 |
+
df = pd.read_csv(ref_result[model])
|
498 |
+
# df = df[df["id"].isin(wikidata_df["id"])]
|
499 |
+
|
500 |
+
best_precision.append(df["precision"].mean())
|
501 |
+
best_recall.append(df["recall"].mean())
|
502 |
+
f1 = (
|
503 |
+
2
|
504 |
+
* (best_precision[-1] * best_recall[-1])
|
505 |
+
/ (best_precision[-1] + best_recall[-1])
|
506 |
+
)
|
507 |
+
# best_f1.append(df["f1"].mean())
|
508 |
+
best_f1.append(f1)
|
509 |
+
|
510 |
+
# Create a DataFrame with the statistics
|
511 |
+
data = (
|
512 |
+
pd.DataFrame(
|
513 |
+
{
|
514 |
+
"Model": model_names,
|
515 |
+
"Adjusted Precision with RP": best_precision,
|
516 |
+
"Adjusted Recall with RP": best_recall,
|
517 |
+
"Adjusted F1 with RP": best_f1,
|
518 |
+
}
|
519 |
+
)
|
520 |
+
if adjusted_f1
|
521 |
+
else pd.DataFrame(
|
522 |
+
{
|
523 |
+
"Model": model_names,
|
524 |
+
"Precision": best_precision,
|
525 |
+
"Recall": best_recall,
|
526 |
+
"F1": best_f1,
|
527 |
+
}
|
528 |
+
)
|
529 |
+
)
|
530 |
+
columns = list(data.columns)
|
531 |
+
|
532 |
+
# Melt the DataFrame to a long format
|
533 |
+
data_melted = data.melt(id_vars="Model", var_name="Metric", value_name="Score")
|
534 |
+
|
535 |
+
# Pivot the DataFrame to a wide format
|
536 |
+
data_pivoted = data_melted.pivot(index="Metric", columns="Model", values="Score")
|
537 |
+
|
538 |
+
# make sure the columns are following the order of the models
|
539 |
+
data_pivoted = data_pivoted[model_names]
|
540 |
+
|
541 |
+
# make sure three groups in the order of precision, recall, f1
|
542 |
+
data_pivoted = data_pivoted.reindex(columns[1:])
|
543 |
+
|
544 |
+
# Plot the statistics
|
545 |
+
plt.figure(figsize=(10, 6))
|
546 |
+
ax = data_pivoted.plot(kind="bar", ax=plt.gca(), width=0.9)
|
547 |
+
plt.title(title)
|
548 |
+
plt.legend(bbox_to_anchor=(1.0, 0.5), loc="center left")
|
549 |
+
|
550 |
+
# Set the rotation of the x-axis labels to 0 degrees
|
551 |
+
plt.xticks(rotation=0)
|
552 |
+
|
553 |
+
# Format the y-axis to display as percentage
|
554 |
+
ax.yaxis.set_major_formatter(mtick.PercentFormatter(1.0))
|
555 |
+
|
556 |
+
# get the max value of the y-axis
|
557 |
+
a1 = max(best_precision)
|
558 |
+
a2 = max(best_recall)
|
559 |
+
a3 = max(best_f1)
|
560 |
+
|
561 |
+
max_value = max([a1, a2, a3]) * 1.12
|
562 |
+
print("max_value:", max_value)
|
563 |
+
|
564 |
+
# Set the y-axis limit up to 70%
|
565 |
+
ax.set_ylim(0, max_value)
|
566 |
+
|
567 |
+
# Add the values above each bar
|
568 |
+
for p in ax.patches:
|
569 |
+
ax.annotate(
|
570 |
+
f"{p.get_height() * 100:.1f}",
|
571 |
+
(p.get_x() + p.get_width() / 2.0, p.get_height()),
|
572 |
+
ha="center",
|
573 |
+
va="bottom",
|
574 |
+
xytext=(0, 10),
|
575 |
+
textcoords="offset points",
|
576 |
+
rotation=90,
|
577 |
+
)
|
578 |
+
|
579 |
+
plt.show()
|
580 |
+
|
581 |
+
|
582 |
+
def plot_best_performance_ms_macro(
|
583 |
+
result,
|
584 |
+
models=None,
|
585 |
+
title="Models with Best Repetition Factor Adjusted Performance",
|
586 |
+
ref_result=None,
|
587 |
+
skip_generic_prompt=False,
|
588 |
+
include_adjusted_performance=True,
|
589 |
+
):
|
590 |
+
# Initialize lists to store the statistics
|
591 |
+
model_names = []
|
592 |
+
best_f1 = []
|
593 |
+
best_afrp = []
|
594 |
+
best_repetition_penalty = []
|
595 |
+
best_bleu1 = []
|
596 |
+
best_rougeL = []
|
597 |
+
|
598 |
+
if models is None:
|
599 |
+
models = result.keys()
|
600 |
+
for model in models:
|
601 |
+
if skip_generic_prompt and "generic prompt" in model:
|
602 |
+
continue
|
603 |
+
print(f"model: {model}")
|
604 |
+
df = result[model]["df_overall"]
|
605 |
+
|
606 |
+
# Calculate the statistics
|
607 |
+
bleu1 = [x for x in df["bleu1"]]
|
608 |
+
rougeL = [x for x in df["rougeL"]]
|
609 |
+
f1 = [2 * (p * r) / (p + r) for p, r in zip(bleu1, rougeL)]
|
610 |
+
|
611 |
+
newline_score = [
|
612 |
+
df["newline_score"].mean()
|
613 |
+
for df in result[model]["df_list_repetition_penalty"]
|
614 |
+
]
|
615 |
+
print(f"newline_score: {newline_score}")
|
616 |
+
|
617 |
+
repetition_score = [
|
618 |
+
df["repetition_score"].mean()
|
619 |
+
for df in result[model]["df_list_repetition_penalty"]
|
620 |
+
]
|
621 |
+
print(f"repetition_score: {repetition_score}")
|
622 |
+
|
623 |
+
afrp = [
|
624 |
+
f / math.log10(10 + n + r)
|
625 |
+
for f, n, r in zip(f1, newline_score, repetition_score)
|
626 |
+
]
|
627 |
+
|
628 |
+
best_afrp.append(max(afrp if include_adjusted_performance else f1))
|
629 |
+
best_afrp_index = (
|
630 |
+
afrp.index(best_afrp[-1])
|
631 |
+
if include_adjusted_performance
|
632 |
+
else f1.index(best_afrp[-1])
|
633 |
+
)
|
634 |
+
best_repetition_penalty.append(df["repetition_penalty"][best_afrp_index])
|
635 |
+
|
636 |
+
best_f1.append(f1[best_afrp_index])
|
637 |
+
best_bleu1.append(bleu1[best_afrp_index])
|
638 |
+
best_rougeL.append(rougeL[best_afrp_index])
|
639 |
+
|
640 |
+
print(
|
641 |
+
f"best repetition penalty: {best_repetition_penalty[-1]}, best afrp: {best_afrp[-1]}, f1: {best_f1[-1]}"
|
642 |
+
)
|
643 |
+
|
644 |
+
df = result[model]["df_list_repetition_penalty"][best_afrp_index]
|
645 |
+
|
646 |
+
model_names.append(
|
647 |
+
f"{model} (RP={best_repetition_penalty[-1]})"
|
648 |
+
) # Add the model name to the list
|
649 |
+
|
650 |
+
if ref_result is not None:
|
651 |
+
print("ref_result:", ref_result)
|
652 |
+
for model in ref_result.keys():
|
653 |
+
model_names.append(model)
|
654 |
+
df = pd.read_csv(ref_result[model], comment="#", on_bad_lines="warn")
|
655 |
+
# df = df[df["id"].isin(wikidata_df["id"])]
|
656 |
+
|
657 |
+
p = df["bleu1"][0]
|
658 |
+
best_bleu1.append(p)
|
659 |
+
|
660 |
+
r = df["rougeL"][0]
|
661 |
+
best_rougeL.append(r)
|
662 |
+
|
663 |
+
f1 = 2 * p * r / (p + r) if p + r > 0 else 0
|
664 |
+
best_f1.append(f1)
|
665 |
+
best_afrp.append(f1)
|
666 |
+
|
667 |
+
print("model_names:", model_names)
|
668 |
+
print("best_f1:", best_f1)
|
669 |
+
print("best_afrp:", best_afrp)
|
670 |
+
|
671 |
+
# Create a DataFrame with the statistics
|
672 |
+
data = (
|
673 |
+
pd.DataFrame(
|
674 |
+
{
|
675 |
+
"Model": model_names,
|
676 |
+
"Repetition Factor Adjusted Perf Score": best_afrp,
|
677 |
+
"Overall Perf Score": best_f1,
|
678 |
+
}
|
679 |
+
)
|
680 |
+
if include_adjusted_performance
|
681 |
+
else pd.DataFrame(
|
682 |
+
{
|
683 |
+
"Model": model_names,
|
684 |
+
"Bleu-1": best_bleu1,
|
685 |
+
"Rouge-L": best_rougeL,
|
686 |
+
"Overall Perf Score": best_f1,
|
687 |
+
}
|
688 |
+
)
|
689 |
+
)
|
690 |
+
|
691 |
+
# Melt the DataFrame to a long format
|
692 |
+
data_melted = data.melt(id_vars="Model", var_name="Metric", value_name="Score")
|
693 |
+
|
694 |
+
# Pivot the DataFrame to a wide format
|
695 |
+
data_pivoted = data_melted.pivot(index="Metric", columns="Model", values="Score")
|
696 |
+
|
697 |
+
# make sure the columns are following the order of the models
|
698 |
+
data_pivoted = data_pivoted[model_names]
|
699 |
+
|
700 |
+
columns = list(data.columns)
|
701 |
+
data_pivoted = data_pivoted.reindex(columns[1:])
|
702 |
+
|
703 |
+
# Plot the statistics
|
704 |
+
plt.figure(figsize=(10, 6))
|
705 |
+
ax = data_pivoted.plot(kind="bar", ax=plt.gca(), width=0.9)
|
706 |
+
plt.title(title)
|
707 |
+
plt.legend(bbox_to_anchor=(1.0, 0.5), loc="center left")
|
708 |
+
|
709 |
+
# Set the rotation of the x-axis labels to 0 degrees
|
710 |
+
plt.xticks(rotation=0)
|
711 |
+
|
712 |
+
# Format the y-axis to display as percentage
|
713 |
+
ax.yaxis.set_major_formatter(mtick.PercentFormatter(1.0))
|
714 |
+
|
715 |
+
# get the max value of the y-axis
|
716 |
+
a1 = max(best_afrp)
|
717 |
+
a2 = max(best_f1)
|
718 |
+
a3 = max(best_bleu1)
|
719 |
+
a4 = max(best_rougeL)
|
720 |
+
|
721 |
+
max_value = (
|
722 |
+
max([a1, a2] if include_adjusted_performance else [a1, a2, a3, a4]) * 1.12
|
723 |
+
)
|
724 |
+
print("max_value:", max_value)
|
725 |
+
|
726 |
+
# Set the y-axis limit up to 70%
|
727 |
+
ax.set_ylim(0, max_value)
|
728 |
+
|
729 |
+
# Add the values above each bar
|
730 |
+
for p in ax.patches:
|
731 |
+
ax.annotate(
|
732 |
+
f"{p.get_height() * 100:.1f}",
|
733 |
+
(p.get_x() + p.get_width() / 2.0, p.get_height()),
|
734 |
+
ha="center",
|
735 |
+
va="bottom",
|
736 |
+
xytext=(0, 10),
|
737 |
+
textcoords="offset points",
|
738 |
+
rotation=90,
|
739 |
+
)
|
740 |
+
|
741 |
+
plt.show()
|
742 |
+
|
743 |
+
|
744 |
+
non_rag_csv_result_files = [
|
745 |
+
"./data/results/Phi-3-mini-128k-instruct_wd_non_rag_batch_16.csv", # Phi-3-mini-128k-instruct(batch size:16)
|
746 |
+
# "./data/results/Tune_2024-04-12_17-14-28.csv", # Orca-2-7b
|
747 |
+
"./data/results/Tune_2024-04-09_09-19-22.csv", # Llama-2-7b-chat-hf
|
748 |
+
# "./data/results/Tune_2024-04-15_12-43-48.csv", # Llama-2-7b-chat-hf(cwq)
|
749 |
+
"./data/results/Meta-Llama-3-8B-Instruct_wd_non_rag.csv", # Meta-Llama-3-8B-Instruct
|
750 |
+
"./data/results/Meta-Llama-3-8B-Instruct_wd_1_non_rag.csv", # Meta-Llama-3-8B-Instruct
|
751 |
+
"./data/results/Tune_2024-04-16_12-24-27.csv.csv", # Mistral-7B-Instruct-v0.2
|
752 |
+
"./data/results/gemma-1.1-2b-it_wd_non_rag.csv", # gemma-1.1-2b-it
|
753 |
+
# "./data/results/Tune_2024-04-17_04-23-15.csv", # gemma-1.1-2b-it(cwq)
|
754 |
+
"./data/results/gemma-1.1-7b-it_wd_non_rag.csv", # gemma-1.1-2b-it
|
755 |
+
# "./data/results/Tune_2024-04-18_21-56-52.csv", # gemma-1.1-7b-it
|
756 |
+
# "./data/results/Tune_2024-04-19_08-14-49.csv", # gemma-1.1-7b-it(cwq)
|
757 |
+
# "./data/results/Tune_2024-04-17_23-52-04.csv", # Orca-2-13b
|
758 |
+
"./data/results/Tune_2024-04-10_16-53-38.csv", # Llama-2-13b-chat-hf
|
759 |
+
"./data/results/Llama-2-70b-chat-hf_wd_non_rag.csv", # Llama-2-70b-chat-hf
|
760 |
+
"./data/results/Meta-Llama-3-70B-Instruct_wd_non_rag.csv", # Meta-Llama-3-70B-Instruct
|
761 |
+
# "./data/results/llama-3-70b-instruct-awq_wd_non_rag.csv", # Llama-3-70b-instruct-awq
|
762 |
+
]
|
763 |
+
|
764 |
+
rag_csv_result_files = [
|
765 |
+
"./data/results/Phi-3-mini-128k-instruct_wd_rag_batch_4.csv", # Phi-3-mini-128k-instruct(batch size:16)
|
766 |
+
"./data/results/Phi-3-mini-128k-instruct_wd_true.csv", # Phi-3-mini-128k-instruct(batch size:16)
|
767 |
+
# "./data/results/Tune_2024-03-19_19-13-36.csv", # Orca-2-7b
|
768 |
+
"./data/results/Tune_2024-03-20_15-35-37.csv", # Llama-2-7b-chat-hf
|
769 |
+
"./data/results/Llama-2-7b-chat-hf_wd_true.csv", # Llama-2-7b-chat-hf(true)
|
770 |
+
# "./data/results/Tune_2024-04-15_14-52-31.csv", # Llama-2-7b-chat-hf(cwq)
|
771 |
+
"./data/results/Meta-Llama-3-8B-Instruct_wd.csv", # Meta-Llama-3-8b-instruct
|
772 |
+
"./data/results/Meta-Llama-3-8B-Instruct_wd_true.csv", # Meta-Llama-3-8b-instruct(true)
|
773 |
+
"./data/results/Tune_2024-03-29_11-28-20.csv", # Mistral-7B-Instruct-v0.2
|
774 |
+
"./data/results/Mistral-7B-Instruct-v0.2_wd_true.csv", # Mistral-7B-Instruct-v0.2(true)
|
775 |
+
"./data/results/gemma-1.1-2b-it_wd.csv", # gemma-1.1-2b-it
|
776 |
+
"./data/results/gemma-1.1-2b-it_wd_true.csv", # gemma-1.1-7b-it(true)
|
777 |
+
# "./data/results/Tune_2024-04-20_13-12-43.csv", # gemma-1.1-2b-it
|
778 |
+
# "./data/results/Tune_2024-04-16_06-48-32.csv", # gemma-1.1-2b-it(cwq)
|
779 |
+
"./data/results/gemma-1.1-7b-it_wd.csv", # gemma-1.1-7b-it
|
780 |
+
"./data/results/gemma-1.1-7b-it_wd_true.csv", # gemma-1.1-7b-it(true)
|
781 |
+
# "./data/results/Tune_2024-04-18_13-18-38.csv", # gemma-1.1-7b-it
|
782 |
+
# "./data/results/Tune_2024-04-19_04-26-33.csv", # gemma-1.1-7b-it(cwq)
|
783 |
+
# "./data/results/Orca-2-13b_wd.csv", # Orca-2-13b
|
784 |
+
# "./data/results/Tune_2024-03-22_09-28-56.csv", # Orca-2-13b
|
785 |
+
"./data/results/Tune_2024-03-25_23-32-57.csv", # Llama-2-13b-chat-hf
|
786 |
+
"./data/results/Llama-2-13b-chat-hf_wd_true.csv", # Llama-2-13b-chat-hf(true)
|
787 |
+
"./data/results/Llama-2-70b-chat-hf_wd.csv", # Llama-2-70b-chat-hf
|
788 |
+
"./data/results/Llama-2-70b-chat-hf_wd_true.csv", # Llama-2-70b-chat-hf
|
789 |
+
"./data/results/Meta-Llama-3-70B-Instruct_wd.csv", # Meta-Llama-3-70B-Instruct
|
790 |
+
"./data/results/Meta-Llama-3-70B-Instruct_wd_true.csv", # Meta-Llama-3-70B-Instruct(true)
|
791 |
+
]
|
792 |
+
|
793 |
+
df_ms_macro = pd.read_json("./data/datasets/ms_macro.json")
|
794 |
+
|
795 |
+
|
796 |
+
def load_for_repetition_penalty_ms_macro(
|
797 |
+
csv_result_file, repetition_penalty, force_recalculate=False
|
798 |
+
):
|
799 |
+
result_file = replace_last(
|
800 |
+
csv_result_file, ".csv", f"_RP_{repetition_penalty:.3f}.csv"
|
801 |
+
)
|
802 |
+
df = load_with_newline_and_repetition_scores(
|
803 |
+
result_file, force_recalculate=force_recalculate
|
804 |
+
)
|
805 |
+
|
806 |
+
if df["ground_truth"][0] != df_ms_macro["wellFormedAnswers"][0]:
|
807 |
+
df["ground_truth"] = df_ms_macro["wellFormedAnswers"]
|
808 |
+
print("ground_truth updated for:", result_file)
|
809 |
+
df.to_csv(result_file, index=False)
|
810 |
+
return df
|
811 |
+
|
812 |
+
|
813 |
+
# MS MACRO
|
814 |
+
def plot_performance_scores_ms_macro(
|
815 |
+
result,
|
816 |
+
models=None,
|
817 |
+
title="Performance",
|
818 |
+
):
|
819 |
+
|
820 |
+
if models is None:
|
821 |
+
models = result.keys()
|
822 |
+
for model in models:
|
823 |
+
print(f"model: {model}")
|
824 |
+
df = result[model]["df_overall"]
|
825 |
+
# print(result[model]["df_list_repetition_penalty"][0].describe())
|
826 |
+
|
827 |
+
# Calculate the statistics
|
828 |
+
bleu1 = list(df["bleu1"])
|
829 |
+
rougeL = list(df["rougeL"])
|
830 |
+
f1 = [2 * (p * r) / (p + r) for p, r in zip(bleu1, rougeL)]
|
831 |
+
best_f1 = max(f1)
|
832 |
+
best_f1_index = f1.index(best_f1)
|
833 |
+
|
834 |
+
bleu1, rougeL = adjust_perf_scores_with_repetition_penalty(
|
835 |
+
result[model], bleu1, rougeL
|
836 |
+
)
|
837 |
+
afrp = [2 * (p * r) / (p + r) for p, r in zip(bleu1, rougeL)]
|
838 |
+
|
839 |
+
# f1 = [df["f1"].mean() for df in result[model]["df_list_repetition_penalty"]]
|
840 |
+
best_afrp = max(afrp)
|
841 |
+
best_afrp_index = afrp.index(best_afrp)
|
842 |
+
|
843 |
+
repetition_penalties = list(df["repetition_penalty"])
|
844 |
+
|
845 |
+
# line plot for precision, recall, f1
|
846 |
+
plt.figure(figsize=(10, 6))
|
847 |
+
|
848 |
+
plt.axvspan(
|
849 |
+
repetition_penalties[best_f1_index] - 0.01,
|
850 |
+
repetition_penalties[best_f1_index] + 0.01,
|
851 |
+
alpha=0.5,
|
852 |
+
edgecolor="none",
|
853 |
+
facecolor="blue",
|
854 |
+
)
|
855 |
+
|
856 |
+
plt.axvspan(
|
857 |
+
repetition_penalties[best_afrp_index] - 0.01,
|
858 |
+
repetition_penalties[best_afrp_index] + 0.01,
|
859 |
+
alpha=0.5,
|
860 |
+
edgecolor="none",
|
861 |
+
facecolor="orange",
|
862 |
+
)
|
863 |
+
|
864 |
+
plt.plot(
|
865 |
+
repetition_penalties,
|
866 |
+
f1,
|
867 |
+
label="Overall Perf Score",
|
868 |
+
marker="D",
|
869 |
+
color="blue",
|
870 |
+
)
|
871 |
+
plt.plot(
|
872 |
+
repetition_penalties,
|
873 |
+
afrp,
|
874 |
+
label="RF Adjusted Perf Score",
|
875 |
+
marker="o",
|
876 |
+
color="orange",
|
877 |
+
)
|
878 |
+
|
879 |
+
plt.xlabel("Repetition Penalties")
|
880 |
+
plt.ylabel("Score")
|
881 |
+
plt.xlim(0.99, 1.31)
|
882 |
+
# y in percentage
|
883 |
+
plt.gca().yaxis.set_major_formatter(mtick.PercentFormatter(1.0))
|
884 |
+
plt.title(f"{model} {title}")
|
885 |
+
plt.legend(bbox_to_anchor=(1.0, 0.5), loc="center left")
|
886 |
+
|
887 |
+
plt.show()
|
888 |
+
|
889 |
+
|
890 |
+
def plot_repetition_factors(result, groups):
|
891 |
+
for group in groups:
|
892 |
+
# Plot the statistics
|
893 |
+
plt.figure(figsize=(10, 6))
|
894 |
+
|
895 |
+
max_value = 0
|
896 |
+
for model in result.keys():
|
897 |
+
if not group in model.lower():
|
898 |
+
continue
|
899 |
+
print(f"model: {model}")
|
900 |
+
df = result[model]["df_overall"]
|
901 |
+
repetition_panelties = [
|
902 |
+
repetition_penalty for repetition_penalty in df["repetition_penalty"]
|
903 |
+
]
|
904 |
+
|
905 |
+
mean_score = [
|
906 |
+
math.log10(10 + df["total_repetitions"].mean())
|
907 |
+
for df in result[model]["df_list_repetition_penalty"]
|
908 |
+
]
|
909 |
+
|
910 |
+
sns.lineplot(x=repetition_panelties, y=mean_score, label=model)
|
911 |
+
|
912 |
+
new_max = max(mean_score)
|
913 |
+
if new_max > max_value:
|
914 |
+
max_value = new_max
|
915 |
+
|
916 |
+
max_value = max_value * 1.05
|
917 |
+
if max_value < 1.5:
|
918 |
+
max_value = 1.5
|
919 |
+
# set ylimit
|
920 |
+
plt.ylim(1, max_value)
|
921 |
+
|
922 |
+
# show grid
|
923 |
+
plt.grid(True)
|
924 |
+
plt.xlabel("Repetition Penalties")
|
925 |
+
plt.ylabel("Repetition Factors")
|
926 |
+
plt.title("Repetition Factors vs Repetition Penalties")
|
927 |
+
plt.legend()
|
928 |
+
|
929 |
+
plt.show()
|
eval_modules/calc_repetitions_v2.py
ADDED
@@ -0,0 +1,1087 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
1 |
+
import os
|
2 |
+
import re
|
3 |
+
import math
|
4 |
+
import pandas as pd
|
5 |
+
import numpy as np
|
6 |
+
import matplotlib.pyplot as plt
|
7 |
+
import matplotlib.ticker as mtick
|
8 |
+
import seaborn as sns
|
9 |
+
import nltk
|
10 |
+
|
11 |
+
# final version
|
12 |
+
pattern_abnormal_newlines = re.compile(r"\n{5,}")
|
13 |
+
pattern_text_repetitions = re.compile(r"(.{5,}?)(\1+)", re.M | re.DOTALL)
|
14 |
+
exception_patterns = [re.compile(r"(\w+\.?)\1")]
|
15 |
+
|
16 |
+
|
17 |
+
# final version for repetition detection
|
18 |
+
def detect_repetitions(
|
19 |
+
text, debug=False, pattern_text_repetitions=pattern_text_repetitions
|
20 |
+
):
|
21 |
+
subtotals = [0, 0]
|
22 |
+
|
23 |
+
if isinstance(text, str):
|
24 |
+
patterns = [pattern_abnormal_newlines, pattern_text_repetitions]
|
25 |
+
for i, pattern in enumerate(patterns):
|
26 |
+
if debug:
|
27 |
+
print(
|
28 |
+
f"----detect {'abnormal newlines' if i == 0 else 'text repetitions'}----"
|
29 |
+
)
|
30 |
+
matches = pattern.finditer(text)
|
31 |
+
for match in matches:
|
32 |
+
if i > 0:
|
33 |
+
ignored = False
|
34 |
+
for exception_pattern in exception_patterns:
|
35 |
+
if exception_pattern.match(match[0]):
|
36 |
+
if debug:
|
37 |
+
print("ignored: ", match[0])
|
38 |
+
ignored = True
|
39 |
+
break
|
40 |
+
if ignored:
|
41 |
+
continue
|
42 |
+
|
43 |
+
if debug:
|
44 |
+
print(match)
|
45 |
+
for groupNum in range(0, len(match.groups())):
|
46 |
+
groupNum = groupNum + 1
|
47 |
+
print(
|
48 |
+
"Group {groupNum} found at {start}-{end}: `{group}`".format(
|
49 |
+
groupNum=groupNum,
|
50 |
+
start=match.start(groupNum),
|
51 |
+
end=match.end(groupNum),
|
52 |
+
group=match.group(groupNum),
|
53 |
+
)
|
54 |
+
)
|
55 |
+
|
56 |
+
start, end = match.span()
|
57 |
+
subtotals[i] += end - start
|
58 |
+
|
59 |
+
if i == 0 and subtotals[i] > 0:
|
60 |
+
text = pattern.sub("", text)
|
61 |
+
if debug:
|
62 |
+
print(f"removed abnormal newlines: {subtotals[i]}")
|
63 |
+
|
64 |
+
result = (subtotals[0], subtotals[1], subtotals[0] + subtotals[1])
|
65 |
+
|
66 |
+
if debug:
|
67 |
+
print(result)
|
68 |
+
return result
|
69 |
+
|
70 |
+
|
71 |
+
def detect_abnormal_newlines(text, debug=False):
|
72 |
+
return detect_repetitions(text, debug=debug)[0]
|
73 |
+
|
74 |
+
|
75 |
+
def detect_text_repetitions(text, debug=False):
|
76 |
+
return detect_repetitions(text, debug=debug)[1]
|
77 |
+
|
78 |
+
|
79 |
+
def detect_scores(text, debug=False):
|
80 |
+
newline_score, repetition_score, total_repetitions = detect_repetitions(
|
81 |
+
text, debug=debug
|
82 |
+
)
|
83 |
+
return pd.Series([newline_score, repetition_score, total_repetitions])
|
84 |
+
|
85 |
+
|
86 |
+
def load_with_newline_and_repetition_scores(result_file, force_recalculate=False):
|
87 |
+
print(f"loading result file: {result_file}")
|
88 |
+
df = pd.read_csv(result_file, comment="#", on_bad_lines="warn")
|
89 |
+
|
90 |
+
if (
|
91 |
+
force_recalculate
|
92 |
+
or "newline_score" not in df.columns
|
93 |
+
or "repetition_score" not in df.columns
|
94 |
+
or "total_repetitions" not in df.columns
|
95 |
+
):
|
96 |
+
df[["newline_score", "repetition_score", "total_repetitions"]] = df[
|
97 |
+
"answer"
|
98 |
+
].apply(detect_scores)
|
99 |
+
df.to_csv(result_file, index=False)
|
100 |
+
|
101 |
+
return df
|
102 |
+
|
103 |
+
|
104 |
+
def replace_last(source_string, old_string, new_string):
|
105 |
+
head, _sep, tail = source_string.rpartition(old_string)
|
106 |
+
return head + new_string + tail
|
107 |
+
|
108 |
+
|
109 |
+
def load_for_repetition_penalty(
|
110 |
+
csv_result_file, repetition_penalty, force_recalculate=False
|
111 |
+
):
|
112 |
+
result_file = replace_last(
|
113 |
+
csv_result_file, ".csv", f"_RP_{repetition_penalty:.3f}.csv"
|
114 |
+
)
|
115 |
+
return load_with_newline_and_repetition_scores(
|
116 |
+
result_file, force_recalculate=force_recalculate
|
117 |
+
)
|
118 |
+
|
119 |
+
|
120 |
+
def calc_adjusted_performance(f, r):
|
121 |
+
return f / math.log10(10 + r)
|
122 |
+
|
123 |
+
|
124 |
+
def calculate_adjusted_performance(row):
|
125 |
+
r = row["total_repetitions"]
|
126 |
+
adjusted_precision = calc_adjusted_performance(row["precision"], r)
|
127 |
+
adjusted_recall = calc_adjusted_performance(row["recall"], r)
|
128 |
+
return pd.Series([adjusted_precision, adjusted_recall])
|
129 |
+
|
130 |
+
|
131 |
+
def load_performance_df(csv_result_file, repetition_penalty):
|
132 |
+
result_file = replace_last(
|
133 |
+
csv_result_file, ".csv", f"_RP_{repetition_penalty:.3f}-t2_evaluated.json"
|
134 |
+
)
|
135 |
+
result_file = result_file.replace("/results/", "/eval/")
|
136 |
+
print(f"loading json file: {result_file}")
|
137 |
+
df = pd.read_json(result_file)
|
138 |
+
|
139 |
+
return df
|
140 |
+
|
141 |
+
|
142 |
+
def calculate_performance_score_v1(
|
143 |
+
csv_result_file, repetition_penalty, force_recalculate=False
|
144 |
+
):
|
145 |
+
result_file = replace_last(
|
146 |
+
csv_result_file, ".csv", f"_RP_{repetition_penalty:.3f}.csv"
|
147 |
+
)
|
148 |
+
print(f"loading result file: {result_file}")
|
149 |
+
df = pd.read_csv(result_file, comment="#", on_bad_lines="warn")
|
150 |
+
|
151 |
+
if force_recalculate or "f2" in df.columns or "f1" not in df.columns:
|
152 |
+
df.drop(
|
153 |
+
columns=[
|
154 |
+
"precision",
|
155 |
+
"recall",
|
156 |
+
"f1",
|
157 |
+
"f2",
|
158 |
+
"entities_in_answer",
|
159 |
+
"entities_in_question",
|
160 |
+
],
|
161 |
+
errors="ignore",
|
162 |
+
inplace=True,
|
163 |
+
)
|
164 |
+
perf_df = load_performance_df(csv_result_file, repetition_penalty)
|
165 |
+
filtered_df = perf_df[perf_df["id"].isin(df["id"])]
|
166 |
+
perf_df = filtered_df.reset_index(drop=True)
|
167 |
+
print(f"perf_df len: {len(perf_df)}")
|
168 |
+
# print(perf_df.head())
|
169 |
+
|
170 |
+
df["eval_gemini_1.0_pro"] = perf_df["eval_gemini_1.0_pro"]
|
171 |
+
|
172 |
+
df["precision"] = perf_df["score"].apply(lambda x: x[0])
|
173 |
+
df["recall"] = perf_df["score"].apply(lambda x: x[1])
|
174 |
+
df["f1"] = perf_df["score"].apply(lambda x: x[2])
|
175 |
+
|
176 |
+
df[["adjusted_precision", "adjusted_recall"]] = df.apply(
|
177 |
+
calculate_adjusted_performance, axis=1
|
178 |
+
)
|
179 |
+
|
180 |
+
df.to_csv(result_file, index=False)
|
181 |
+
print(f"performance scores saved to result file: {result_file}")
|
182 |
+
|
183 |
+
print(f"df len: {len(df)}")
|
184 |
+
|
185 |
+
return df
|
186 |
+
|
187 |
+
|
188 |
+
ref_df = pd.read_csv(
|
189 |
+
"./data/results/gpt-3.5-turbo_non_rag.csv", comment="#", on_bad_lines="warn"
|
190 |
+
)
|
191 |
+
|
192 |
+
|
193 |
+
def calculate_performance_score(
|
194 |
+
csv_result_file, repetition_penalty, force_recalculate=False
|
195 |
+
):
|
196 |
+
result_file = replace_last(
|
197 |
+
csv_result_file, ".csv", f"_RP_{repetition_penalty:.3f}.csv"
|
198 |
+
)
|
199 |
+
|
200 |
+
re_creating = False
|
201 |
+
if os.path.exists(result_file):
|
202 |
+
print(f"loading result file: {result_file}")
|
203 |
+
df = pd.read_csv(result_file, comment="#", on_bad_lines="warn")
|
204 |
+
else:
|
205 |
+
print(f"re-creating result file: {result_file}")
|
206 |
+
df = pd.DataFrame()
|
207 |
+
force_recalculate = True
|
208 |
+
|
209 |
+
if force_recalculate or "f2" in df.columns or "f1" not in df.columns:
|
210 |
+
df.drop(
|
211 |
+
columns=[
|
212 |
+
"precision",
|
213 |
+
"recall",
|
214 |
+
"f1",
|
215 |
+
"f2",
|
216 |
+
"entities_in_answer",
|
217 |
+
"entities_in_question",
|
218 |
+
"word_count",
|
219 |
+
],
|
220 |
+
errors="ignore",
|
221 |
+
inplace=True,
|
222 |
+
)
|
223 |
+
perf_df = load_performance_df(csv_result_file, repetition_penalty)
|
224 |
+
filtered_df = perf_df[perf_df["id"].isin(ref_df["id"])]
|
225 |
+
perf_df = filtered_df.reset_index(drop=True)
|
226 |
+
print(f"perf_df len: {len(perf_df)}")
|
227 |
+
|
228 |
+
if len(perf_df) != len(ref_df):
|
229 |
+
print(f"error: len(perf_df) != {len(ref_df)}")
|
230 |
+
missing_ids = [
|
231 |
+
id for id in ref_df["id"].unique() if id not in perf_df["id"].unique()
|
232 |
+
]
|
233 |
+
print(f"missing_ids: {missing_ids}")
|
234 |
+
|
235 |
+
# print(perf_df.head())
|
236 |
+
|
237 |
+
df["id"] = perf_df["id"]
|
238 |
+
df["question"] = perf_df["question"]
|
239 |
+
df["answer"] = perf_df["pred_answer"]
|
240 |
+
df["word_count"] = df["answer"].apply(
|
241 |
+
lambda x: len(nltk.word_tokenize(x)) if isinstance(x, str) else 0
|
242 |
+
)
|
243 |
+
df["ground_truth"] = perf_df["ground_truth"]
|
244 |
+
df[["newline_score", "repetition_score", "total_repetitions"]] = df[
|
245 |
+
"answer"
|
246 |
+
].apply(detect_scores)
|
247 |
+
|
248 |
+
df["eval_gemini_1.0_pro"] = perf_df["eval_gemini_1.0_pro"]
|
249 |
+
df["precision"] = perf_df["score"].apply(lambda x: x[0])
|
250 |
+
df["recall"] = perf_df["score"].apply(lambda x: x[1])
|
251 |
+
df["f1"] = perf_df["score"].apply(lambda x: x[2])
|
252 |
+
|
253 |
+
df[["adjusted_precision", "adjusted_recall"]] = df.apply(
|
254 |
+
calculate_adjusted_performance, axis=1
|
255 |
+
)
|
256 |
+
|
257 |
+
df.to_csv(result_file, index=False)
|
258 |
+
print(f"performance scores saved to result file: {result_file}")
|
259 |
+
|
260 |
+
print(f"df len: {len(df)}")
|
261 |
+
|
262 |
+
return df
|
263 |
+
|
264 |
+
|
265 |
+
def adjust_perf_scores_with_repetition_penalty(result, precision, recall):
|
266 |
+
newline_score = [
|
267 |
+
df["newline_score"].mean() for df in result["df_list_repetition_penalty"]
|
268 |
+
]
|
269 |
+
print(f"newline_score: {newline_score}")
|
270 |
+
|
271 |
+
repetition_score = [
|
272 |
+
df["repetition_score"].mean() for df in result["df_list_repetition_penalty"]
|
273 |
+
]
|
274 |
+
print(f"repetition_score: {repetition_score}")
|
275 |
+
|
276 |
+
precision = [
|
277 |
+
f / math.log10(10 + n + r)
|
278 |
+
for f, n, r in zip(precision, newline_score, repetition_score)
|
279 |
+
]
|
280 |
+
recall = [
|
281 |
+
f / math.log10(10 + n + r)
|
282 |
+
for f, n, r in zip(recall, newline_score, repetition_score)
|
283 |
+
]
|
284 |
+
|
285 |
+
return precision, recall
|
286 |
+
|
287 |
+
|
288 |
+
def plot_performance_scores(
|
289 |
+
result,
|
290 |
+
models=None,
|
291 |
+
title="Performance",
|
292 |
+
):
|
293 |
+
|
294 |
+
if models is None:
|
295 |
+
models = result.keys()
|
296 |
+
for model in models:
|
297 |
+
print(f"model: {model}")
|
298 |
+
df = result[model]["df_overall"]
|
299 |
+
|
300 |
+
# Calculate the statistics
|
301 |
+
precision = [
|
302 |
+
df["precision"].mean() for df in result[model]["df_list_repetition_penalty"]
|
303 |
+
]
|
304 |
+
recall = [
|
305 |
+
df["recall"].mean() for df in result[model]["df_list_repetition_penalty"]
|
306 |
+
]
|
307 |
+
f1 = [2 * (p * r) / (p + r) for p, r in zip(precision, recall)]
|
308 |
+
best_f1 = max(f1)
|
309 |
+
best_f1_index = f1.index(best_f1)
|
310 |
+
|
311 |
+
precision, recall = adjust_perf_scores_with_repetition_penalty(
|
312 |
+
result[model], precision, recall
|
313 |
+
)
|
314 |
+
afrp = [2 * (p * r) / (p + r) for p, r in zip(precision, recall)]
|
315 |
+
|
316 |
+
# f1 = [df["f1"].mean() for df in result[model]["df_list_repetition_penalty"]]
|
317 |
+
best_afrp = max(afrp)
|
318 |
+
best_afrp_index = afrp.index(best_afrp)
|
319 |
+
|
320 |
+
adjusted_precision = [
|
321 |
+
df["adjusted_precision"].mean()
|
322 |
+
for df in result[model]["df_list_repetition_penalty"]
|
323 |
+
]
|
324 |
+
adjusted_recall = [
|
325 |
+
df["adjusted_recall"].mean()
|
326 |
+
for df in result[model]["df_list_repetition_penalty"]
|
327 |
+
]
|
328 |
+
afrp2 = [
|
329 |
+
2 * (p * r) / (p + r) for p, r in zip(adjusted_precision, adjusted_recall)
|
330 |
+
]
|
331 |
+
best_afrp2 = max(afrp2)
|
332 |
+
best_afrp2_index = afrp2.index(best_afrp2)
|
333 |
+
|
334 |
+
repetition_penalties = list(df["repetition_penalty"])
|
335 |
+
|
336 |
+
# line plot for precision, recall, f1
|
337 |
+
plt.figure(figsize=(10, 6))
|
338 |
+
|
339 |
+
plt.axvspan(
|
340 |
+
repetition_penalties[best_f1_index] - 0.01,
|
341 |
+
repetition_penalties[best_f1_index] + 0.01,
|
342 |
+
alpha=0.5,
|
343 |
+
edgecolor="none",
|
344 |
+
facecolor="blue",
|
345 |
+
)
|
346 |
+
|
347 |
+
# plt.axvspan(
|
348 |
+
# repetition_penalties[best_afrp2_index] - 0.01,
|
349 |
+
# repetition_penalties[best_afrp2_index] + 0.01,
|
350 |
+
# alpha=0.5,
|
351 |
+
# edgecolor="none",
|
352 |
+
# facecolor="green",
|
353 |
+
# )
|
354 |
+
|
355 |
+
plt.axvspan(
|
356 |
+
repetition_penalties[best_afrp_index] - 0.01,
|
357 |
+
repetition_penalties[best_afrp_index] + 0.01,
|
358 |
+
alpha=0.5,
|
359 |
+
edgecolor="none",
|
360 |
+
facecolor="orange",
|
361 |
+
)
|
362 |
+
|
363 |
+
plt.plot(repetition_penalties, f1, label="F1", marker="D", color="blue")
|
364 |
+
# plt.plot(
|
365 |
+
# repetition_penalties,
|
366 |
+
# afrp2,
|
367 |
+
# label="Per-question RF Adjusted F1",
|
368 |
+
# marker="s",
|
369 |
+
# color="green",
|
370 |
+
# )
|
371 |
+
plt.plot(
|
372 |
+
repetition_penalties,
|
373 |
+
afrp,
|
374 |
+
label="RF Adjusted F1",
|
375 |
+
marker="o",
|
376 |
+
color="orange",
|
377 |
+
)
|
378 |
+
plt.xlabel("Repetition Penalties")
|
379 |
+
plt.ylabel("Score")
|
380 |
+
plt.xlim(0.99, 1.31)
|
381 |
+
# y in percentage
|
382 |
+
plt.gca().yaxis.set_major_formatter(mtick.PercentFormatter(1.0))
|
383 |
+
plt.title(f"{model} {title}")
|
384 |
+
plt.legend(bbox_to_anchor=(1.0, 0.5), loc="center left")
|
385 |
+
|
386 |
+
plt.show()
|
387 |
+
|
388 |
+
|
389 |
+
def plot_best_afrp(
|
390 |
+
result,
|
391 |
+
models=None,
|
392 |
+
title="Models with Best Repetition Factor Adjusted F1",
|
393 |
+
ref_result=None,
|
394 |
+
):
|
395 |
+
# Initialize lists to store the statistics
|
396 |
+
model_names = []
|
397 |
+
best_f1 = []
|
398 |
+
best_afrp = []
|
399 |
+
best_repetition_penalty = []
|
400 |
+
|
401 |
+
if models is None:
|
402 |
+
models = result.keys()
|
403 |
+
for model in models:
|
404 |
+
print(f"model: {model}")
|
405 |
+
df = result[model]["df_overall"]
|
406 |
+
|
407 |
+
# Calculate the statistics
|
408 |
+
precision = [
|
409 |
+
df["precision"].mean() for df in result[model]["df_list_repetition_penalty"]
|
410 |
+
]
|
411 |
+
recall = [
|
412 |
+
df["recall"].mean() for df in result[model]["df_list_repetition_penalty"]
|
413 |
+
]
|
414 |
+
# f1 = [df["f1"].mean() for df in result[model]["df_list_repetition_penalty"]]
|
415 |
+
f1 = [2 * (p * r) / (p + r) for p, r in zip(precision, recall)]
|
416 |
+
|
417 |
+
newline_score = [
|
418 |
+
df["newline_score"].mean()
|
419 |
+
for df in result[model]["df_list_repetition_penalty"]
|
420 |
+
]
|
421 |
+
print(f"newline_score: {newline_score}")
|
422 |
+
|
423 |
+
repetition_score = [
|
424 |
+
df["repetition_score"].mean()
|
425 |
+
for df in result[model]["df_list_repetition_penalty"]
|
426 |
+
]
|
427 |
+
print(f"repetition_score: {repetition_score}")
|
428 |
+
|
429 |
+
afrp = [
|
430 |
+
f / math.log10(10 + n + r)
|
431 |
+
for f, n, r in zip(f1, newline_score, repetition_score)
|
432 |
+
]
|
433 |
+
|
434 |
+
best_afrp.append(max(afrp))
|
435 |
+
best_afrp_index = afrp.index(best_afrp[-1])
|
436 |
+
best_repetition_penalty.append(df["repetition_penalty"][best_afrp_index])
|
437 |
+
|
438 |
+
best_f1.append(f1[best_afrp_index])
|
439 |
+
|
440 |
+
print(
|
441 |
+
f"best repetition penalty: {best_repetition_penalty[-1]}, best afrp: {best_afrp[-1]}, f1: {best_f1[-1]}"
|
442 |
+
)
|
443 |
+
|
444 |
+
df = result[model]["df_list_repetition_penalty"][best_afrp_index]
|
445 |
+
|
446 |
+
model_names.append(
|
447 |
+
f"{model} (RP={best_repetition_penalty[-1]})"
|
448 |
+
) # Add the model name to the list
|
449 |
+
|
450 |
+
if ref_result is not None:
|
451 |
+
print("ref_result:", ref_result)
|
452 |
+
for model in ref_result.keys():
|
453 |
+
model_names.append(model)
|
454 |
+
df = pd.read_csv(ref_result[model])
|
455 |
+
# df = df[df["id"].isin(wikidata_df["id"])]
|
456 |
+
|
457 |
+
p = df["precision"].mean()
|
458 |
+
r = df["recall"].mean()
|
459 |
+
|
460 |
+
f1 = 2 * p * r / (p + r) if p + r > 0 else 0
|
461 |
+
best_f1.append(f1)
|
462 |
+
best_afrp.append(f1)
|
463 |
+
|
464 |
+
print("model_names:", model_names)
|
465 |
+
print("best_f1:", best_f1)
|
466 |
+
print("best_afrp:", best_afrp)
|
467 |
+
|
468 |
+
# Create a DataFrame with the statistics
|
469 |
+
data = pd.DataFrame(
|
470 |
+
{
|
471 |
+
"Model": model_names,
|
472 |
+
"Repetition Factor Adjusted F1": best_afrp,
|
473 |
+
"F1": best_f1,
|
474 |
+
}
|
475 |
+
)
|
476 |
+
|
477 |
+
# Melt the DataFrame to a long format
|
478 |
+
data_melted = data.melt(id_vars="Model", var_name="Metric", value_name="Score")
|
479 |
+
|
480 |
+
# Pivot the DataFrame to a wide format
|
481 |
+
data_pivoted = data_melted.pivot(index="Metric", columns="Model", values="Score")
|
482 |
+
|
483 |
+
# make sure the columns are following the order of the models
|
484 |
+
data_pivoted = data_pivoted[model_names]
|
485 |
+
|
486 |
+
# make sure three groups in the order of precision, recall, f1
|
487 |
+
data_pivoted = data_pivoted.reindex(["Repetition Factor Adjusted F1", "F1"])
|
488 |
+
|
489 |
+
# Plot the statistics
|
490 |
+
plt.figure(figsize=(15, 6))
|
491 |
+
ax = data_pivoted.plot(kind="bar", ax=plt.gca(), width=0.9)
|
492 |
+
plt.title(title)
|
493 |
+
plt.legend(bbox_to_anchor=(1.0, 0.5), loc="center left")
|
494 |
+
|
495 |
+
# Set the rotation of the x-axis labels to 0 degrees
|
496 |
+
plt.xticks(rotation=0)
|
497 |
+
|
498 |
+
# Format the y-axis to display as percentage
|
499 |
+
ax.yaxis.set_major_formatter(mtick.PercentFormatter(1.0))
|
500 |
+
|
501 |
+
# get the max value of the y-axis
|
502 |
+
a1 = max(best_afrp)
|
503 |
+
a2 = max(best_f1)
|
504 |
+
|
505 |
+
max_value = max([a1, a2]) * 1.12
|
506 |
+
print("max_value:", max_value)
|
507 |
+
|
508 |
+
# Set the y-axis limit up to 70%
|
509 |
+
ax.set_ylim(0, max_value)
|
510 |
+
|
511 |
+
# Add the values above each bar
|
512 |
+
for p in ax.patches:
|
513 |
+
ax.annotate(
|
514 |
+
f"{p.get_height() * 100:.1f}",
|
515 |
+
(p.get_x() + p.get_width() / 2.0, p.get_height()),
|
516 |
+
ha="center",
|
517 |
+
va="bottom",
|
518 |
+
xytext=(0, 10),
|
519 |
+
textcoords="offset points",
|
520 |
+
rotation=90,
|
521 |
+
)
|
522 |
+
|
523 |
+
plt.show()
|
524 |
+
|
525 |
+
|
526 |
+
def plot_best_performance(
|
527 |
+
result,
|
528 |
+
models=None,
|
529 |
+
title="Models with Best F1 Score",
|
530 |
+
adjusted_f1=False,
|
531 |
+
ref_result=None,
|
532 |
+
):
|
533 |
+
# Initialize lists to store the statistics
|
534 |
+
model_names = []
|
535 |
+
best_precision = []
|
536 |
+
best_recall = []
|
537 |
+
best_f1 = []
|
538 |
+
best_repetition_penalty = []
|
539 |
+
|
540 |
+
if models is None:
|
541 |
+
models = result.keys()
|
542 |
+
for model in models:
|
543 |
+
print(f"model: {model}")
|
544 |
+
df = result[model]["df_overall"]
|
545 |
+
|
546 |
+
# Calculate the statistics
|
547 |
+
precision = [
|
548 |
+
df["precision"].mean() for df in result[model]["df_list_repetition_penalty"]
|
549 |
+
]
|
550 |
+
recall = [
|
551 |
+
df["recall"].mean() for df in result[model]["df_list_repetition_penalty"]
|
552 |
+
]
|
553 |
+
|
554 |
+
if adjusted_f1:
|
555 |
+
precision, recall = adjust_perf_scores_with_repetition_penalty(
|
556 |
+
result[model], precision, recall
|
557 |
+
)
|
558 |
+
|
559 |
+
# f1 = [df["f1"].mean() for df in result[model]["df_list_repetition_penalty"]]
|
560 |
+
f1 = [2 * (p * r) / (p + r) for p, r in zip(precision, recall)]
|
561 |
+
|
562 |
+
best_f1.append(max(f1))
|
563 |
+
best_f1_index = f1.index(best_f1[-1])
|
564 |
+
best_repetition_penalty.append(df["repetition_penalty"][best_f1_index])
|
565 |
+
|
566 |
+
best_precision.append(precision[best_f1_index])
|
567 |
+
best_recall.append(recall[best_f1_index])
|
568 |
+
|
569 |
+
print(
|
570 |
+
f"best repetition penalty: {best_repetition_penalty[-1]}, best f1: {best_f1[-1]}, precision: {best_precision[-1]}, recall: {best_recall[-1]}"
|
571 |
+
)
|
572 |
+
|
573 |
+
df = result[model]["df_list_repetition_penalty"][best_f1_index]
|
574 |
+
|
575 |
+
model_names.append(
|
576 |
+
f"{model} (RP={best_repetition_penalty[-1]})"
|
577 |
+
) # Add the model name to the list
|
578 |
+
|
579 |
+
# print sum for columns: newline_score, repetition_score
|
580 |
+
print(
|
581 |
+
f"newline_score: {df['newline_score'].sum()}, repetition_score: {df['repetition_score'].sum()}"
|
582 |
+
)
|
583 |
+
|
584 |
+
if ref_result is not None:
|
585 |
+
print("ref_result:", ref_result)
|
586 |
+
for model in ref_result.keys():
|
587 |
+
model_names.append(model)
|
588 |
+
df = pd.read_csv(ref_result[model])
|
589 |
+
# df = df[df["id"].isin(wikidata_df["id"])]
|
590 |
+
|
591 |
+
best_precision.append(df["precision"].mean())
|
592 |
+
best_recall.append(df["recall"].mean())
|
593 |
+
f1 = (
|
594 |
+
2
|
595 |
+
* (best_precision[-1] * best_recall[-1])
|
596 |
+
/ (best_precision[-1] + best_recall[-1])
|
597 |
+
)
|
598 |
+
# best_f1.append(df["f1"].mean())
|
599 |
+
best_f1.append(f1)
|
600 |
+
|
601 |
+
# Create a DataFrame with the statistics
|
602 |
+
data = (
|
603 |
+
pd.DataFrame(
|
604 |
+
{
|
605 |
+
"Model": model_names,
|
606 |
+
"Adjusted Precision with RP": best_precision,
|
607 |
+
"Adjusted Recall with RP": best_recall,
|
608 |
+
"Adjusted F1 with RP": best_f1,
|
609 |
+
}
|
610 |
+
)
|
611 |
+
if adjusted_f1
|
612 |
+
else pd.DataFrame(
|
613 |
+
{
|
614 |
+
"Model": model_names,
|
615 |
+
"Precision": best_precision,
|
616 |
+
"Recall": best_recall,
|
617 |
+
"F1": best_f1,
|
618 |
+
}
|
619 |
+
)
|
620 |
+
)
|
621 |
+
columns = list(data.columns)
|
622 |
+
|
623 |
+
# Melt the DataFrame to a long format
|
624 |
+
data_melted = data.melt(id_vars="Model", var_name="Metric", value_name="Score")
|
625 |
+
|
626 |
+
# Pivot the DataFrame to a wide format
|
627 |
+
data_pivoted = data_melted.pivot(index="Metric", columns="Model", values="Score")
|
628 |
+
|
629 |
+
# make sure the columns are following the order of the models
|
630 |
+
data_pivoted = data_pivoted[model_names]
|
631 |
+
|
632 |
+
# make sure three groups in the order of precision, recall, f1
|
633 |
+
data_pivoted = data_pivoted.reindex(columns[1:])
|
634 |
+
|
635 |
+
# Plot the statistics
|
636 |
+
plt.figure(figsize=(10, 6))
|
637 |
+
ax = data_pivoted.plot(kind="bar", ax=plt.gca(), width=0.9)
|
638 |
+
plt.title(title)
|
639 |
+
plt.legend(bbox_to_anchor=(1.0, 0.5), loc="center left")
|
640 |
+
|
641 |
+
# Set the rotation of the x-axis labels to 0 degrees
|
642 |
+
plt.xticks(rotation=0)
|
643 |
+
|
644 |
+
# Format the y-axis to display as percentage
|
645 |
+
ax.yaxis.set_major_formatter(mtick.PercentFormatter(1.0))
|
646 |
+
|
647 |
+
# get the max value of the y-axis
|
648 |
+
a1 = max(best_precision)
|
649 |
+
a2 = max(best_recall)
|
650 |
+
a3 = max(best_f1)
|
651 |
+
|
652 |
+
max_value = max([a1, a2, a3]) * 1.12
|
653 |
+
print("max_value:", max_value)
|
654 |
+
|
655 |
+
# Set the y-axis limit up to 70%
|
656 |
+
ax.set_ylim(0, max_value)
|
657 |
+
|
658 |
+
# Add the values above each bar
|
659 |
+
for p in ax.patches:
|
660 |
+
ax.annotate(
|
661 |
+
f"{p.get_height() * 100:.1f}",
|
662 |
+
(p.get_x() + p.get_width() / 2.0, p.get_height()),
|
663 |
+
ha="center",
|
664 |
+
va="bottom",
|
665 |
+
xytext=(0, 10),
|
666 |
+
textcoords="offset points",
|
667 |
+
rotation=90,
|
668 |
+
)
|
669 |
+
|
670 |
+
plt.show()
|
671 |
+
|
672 |
+
|
673 |
+
def plot_best_performance_ms_macro(
|
674 |
+
result,
|
675 |
+
models=None,
|
676 |
+
title="Models with Best Repetition Factor Adjusted Performance",
|
677 |
+
ref_result=None,
|
678 |
+
skip_generic_prompt=False,
|
679 |
+
include_adjusted_performance=True,
|
680 |
+
):
|
681 |
+
# Initialize lists to store the statistics
|
682 |
+
model_names = []
|
683 |
+
best_f1 = []
|
684 |
+
best_afrp = []
|
685 |
+
best_repetition_penalty = []
|
686 |
+
best_bleu1 = []
|
687 |
+
best_rougeL = []
|
688 |
+
|
689 |
+
if models is None:
|
690 |
+
models = result.keys()
|
691 |
+
for model in models:
|
692 |
+
if skip_generic_prompt and "generic prompt" in model:
|
693 |
+
continue
|
694 |
+
print(f"model: {model}")
|
695 |
+
df = result[model]["df_overall"]
|
696 |
+
|
697 |
+
# Calculate the statistics
|
698 |
+
bleu1 = [x for x in df["bleu1"]]
|
699 |
+
rougeL = [x for x in df["rougeL"]]
|
700 |
+
f1 = [2 * (p * r) / (p + r) for p, r in zip(bleu1, rougeL)]
|
701 |
+
|
702 |
+
newline_score = [
|
703 |
+
df["newline_score"].mean()
|
704 |
+
for df in result[model]["df_list_repetition_penalty"]
|
705 |
+
]
|
706 |
+
print(f"newline_score: {newline_score}")
|
707 |
+
|
708 |
+
repetition_score = [
|
709 |
+
df["repetition_score"].mean()
|
710 |
+
for df in result[model]["df_list_repetition_penalty"]
|
711 |
+
]
|
712 |
+
print(f"repetition_score: {repetition_score}")
|
713 |
+
|
714 |
+
afrp = [
|
715 |
+
f / math.log10(10 + n + r)
|
716 |
+
for f, n, r in zip(f1, newline_score, repetition_score)
|
717 |
+
]
|
718 |
+
|
719 |
+
best_afrp.append(max(afrp if include_adjusted_performance else f1))
|
720 |
+
best_afrp_index = (
|
721 |
+
afrp.index(best_afrp[-1])
|
722 |
+
if include_adjusted_performance
|
723 |
+
else f1.index(best_afrp[-1])
|
724 |
+
)
|
725 |
+
best_repetition_penalty.append(df["repetition_penalty"][best_afrp_index])
|
726 |
+
|
727 |
+
best_f1.append(f1[best_afrp_index])
|
728 |
+
best_bleu1.append(bleu1[best_afrp_index])
|
729 |
+
best_rougeL.append(rougeL[best_afrp_index])
|
730 |
+
|
731 |
+
print(
|
732 |
+
f"best repetition penalty: {best_repetition_penalty[-1]}, best afrp: {best_afrp[-1]}, f1: {best_f1[-1]}"
|
733 |
+
)
|
734 |
+
|
735 |
+
df = result[model]["df_list_repetition_penalty"][best_afrp_index]
|
736 |
+
|
737 |
+
model_names.append(
|
738 |
+
f"{model} (RP={best_repetition_penalty[-1]})"
|
739 |
+
) # Add the model name to the list
|
740 |
+
|
741 |
+
if ref_result is not None:
|
742 |
+
print("ref_result:", ref_result)
|
743 |
+
for model in ref_result.keys():
|
744 |
+
model_names.append(model)
|
745 |
+
df = pd.read_csv(ref_result[model], comment="#", on_bad_lines="warn")
|
746 |
+
# df = df[df["id"].isin(wikidata_df["id"])]
|
747 |
+
|
748 |
+
p = df["bleu1"][0]
|
749 |
+
best_bleu1.append(p)
|
750 |
+
|
751 |
+
r = df["rougeL"][0]
|
752 |
+
best_rougeL.append(r)
|
753 |
+
|
754 |
+
f1 = 2 * p * r / (p + r) if p + r > 0 else 0
|
755 |
+
best_f1.append(f1)
|
756 |
+
best_afrp.append(f1)
|
757 |
+
|
758 |
+
print("model_names:", model_names)
|
759 |
+
print("best_f1:", best_f1)
|
760 |
+
print("best_afrp:", best_afrp)
|
761 |
+
|
762 |
+
# Create a DataFrame with the statistics
|
763 |
+
data = (
|
764 |
+
pd.DataFrame(
|
765 |
+
{
|
766 |
+
"Model": model_names,
|
767 |
+
"Repetition Factor Adjusted Perf Score": best_afrp,
|
768 |
+
"Overall Perf Score": best_f1,
|
769 |
+
}
|
770 |
+
)
|
771 |
+
if include_adjusted_performance
|
772 |
+
else pd.DataFrame(
|
773 |
+
{
|
774 |
+
"Model": model_names,
|
775 |
+
"Bleu-1": best_bleu1,
|
776 |
+
"Rouge-L": best_rougeL,
|
777 |
+
"Overall Perf Score": best_f1,
|
778 |
+
}
|
779 |
+
)
|
780 |
+
)
|
781 |
+
|
782 |
+
# Melt the DataFrame to a long format
|
783 |
+
data_melted = data.melt(id_vars="Model", var_name="Metric", value_name="Score")
|
784 |
+
|
785 |
+
# Pivot the DataFrame to a wide format
|
786 |
+
data_pivoted = data_melted.pivot(index="Metric", columns="Model", values="Score")
|
787 |
+
|
788 |
+
# make sure the columns are following the order of the models
|
789 |
+
data_pivoted = data_pivoted[model_names]
|
790 |
+
|
791 |
+
columns = list(data.columns)
|
792 |
+
data_pivoted = data_pivoted.reindex(columns[1:])
|
793 |
+
|
794 |
+
# Plot the statistics
|
795 |
+
plt.figure(figsize=(10, 6))
|
796 |
+
ax = data_pivoted.plot(kind="bar", ax=plt.gca(), width=0.9)
|
797 |
+
plt.title(title)
|
798 |
+
plt.legend(bbox_to_anchor=(1.0, 0.5), loc="center left")
|
799 |
+
|
800 |
+
# Set the rotation of the x-axis labels to 0 degrees
|
801 |
+
plt.xticks(rotation=0)
|
802 |
+
|
803 |
+
# Format the y-axis to display as percentage
|
804 |
+
ax.yaxis.set_major_formatter(mtick.PercentFormatter(1.0))
|
805 |
+
|
806 |
+
# get the max value of the y-axis
|
807 |
+
a1 = max(best_afrp)
|
808 |
+
a2 = max(best_f1)
|
809 |
+
a3 = max(best_bleu1)
|
810 |
+
a4 = max(best_rougeL)
|
811 |
+
|
812 |
+
max_value = (
|
813 |
+
max([a1, a2] if include_adjusted_performance else [a1, a2, a3, a4]) * 1.12
|
814 |
+
)
|
815 |
+
print("max_value:", max_value)
|
816 |
+
|
817 |
+
# Set the y-axis limit up to 70%
|
818 |
+
ax.set_ylim(0, max_value)
|
819 |
+
|
820 |
+
# Add the values above each bar
|
821 |
+
for p in ax.patches:
|
822 |
+
ax.annotate(
|
823 |
+
f"{p.get_height() * 100:.1f}",
|
824 |
+
(p.get_x() + p.get_width() / 2.0, p.get_height()),
|
825 |
+
ha="center",
|
826 |
+
va="bottom",
|
827 |
+
xytext=(0, 10),
|
828 |
+
textcoords="offset points",
|
829 |
+
rotation=90,
|
830 |
+
)
|
831 |
+
|
832 |
+
plt.show()
|
833 |
+
|
834 |
+
|
835 |
+
all_open_source_models = [
|
836 |
+
"gemma-1.1-2b-it",
|
837 |
+
"Phi-3-mini-128k-instruct",
|
838 |
+
"gemma-1.1-7b-it",
|
839 |
+
"Llama-2-7b-chat-hf",
|
840 |
+
"Mistral-7B-Instruct-v0.2",
|
841 |
+
"Meta-Llama-3-8B-Instruct",
|
842 |
+
"Llama-2-13b-chat-hf",
|
843 |
+
"Llama-2-70b-chat-hf",
|
844 |
+
"Meta-Llama-3-70B-Instruct",
|
845 |
+
]
|
846 |
+
|
847 |
+
|
848 |
+
non_rag_csv_result_files = [
|
849 |
+
"./data/results/gemma-1.1-2b-it_wd_non_rag.csv", # gemma-1.1-2b-it
|
850 |
+
"./data/results/Phi-3-mini-128k-instruct_wd_non_rag_batch_16.csv", # Phi-3-mini-128k-instruct(batch size:16)
|
851 |
+
"./data/results/gemma-1.1-7b-it_wd_non_rag.csv", # gemma-1.1-7b-it
|
852 |
+
"./data/results/Tune_2024-04-09_09-19-22.csv", # Llama-2-7b-chat-hf
|
853 |
+
"./data/results/Tune_2024-04-16_12-24-27.csv.csv", # Mistral-7B-Instruct-v0.2
|
854 |
+
"./data/results/Meta-Llama-3-8B-Instruct_wd_non_rag.csv", # Meta-Llama-3-8B-Instruct
|
855 |
+
"./data/results/Meta-Llama-3-8B-Instruct_wd_1_non_rag.csv", # Meta-Llama-3-8B-Instruct
|
856 |
+
"./data/results/Tune_2024-04-10_16-53-38.csv", # Llama-2-13b-chat-hf
|
857 |
+
"./data/results/Llama-2-70b-chat-hf_wd_non_rag.csv", # Llama-2-70b-chat-hf
|
858 |
+
"./data/results/Meta-Llama-3-70B-Instruct_wd_non_rag.csv", # Meta-Llama-3-70B-Instruct
|
859 |
+
]
|
860 |
+
|
861 |
+
rag_csv_result_files = [
|
862 |
+
"./data/results/gemma-1.1-2b-it_wd.csv", # gemma-1.1-2b-it
|
863 |
+
"./data/results/gemma-1.1-2b-it_wd_true.csv", # gemma-1.1-2b-it(true)
|
864 |
+
"./data/results/Phi-3-mini-128k-instruct_wd_rag_batch_4.csv", # Phi-3-mini-128k-instruct(batch size:16)
|
865 |
+
"./data/results/Phi-3-mini-128k-instruct_wd_true.csv", # Phi-3-mini-128k-instruct(batch size:16)
|
866 |
+
"./data/results/gemma-1.1-7b-it_wd.csv", # gemma-1.1-7b-it
|
867 |
+
"./data/results/gemma-1.1-7b-it_wd_true.csv", # gemma-1.1-7b-it(true)
|
868 |
+
"./data/results/Tune_2024-03-20_15-35-37.csv", # Llama-2-7b-chat-hf
|
869 |
+
"./data/results/Llama-2-7b-chat-hf_wd_true.csv", # Llama-2-7b-chat-hf(true)
|
870 |
+
"./data/results/Tune_2024-03-29_11-28-20.csv", # Mistral-7B-Instruct-v0.2
|
871 |
+
"./data/results/Mistral-7B-Instruct-v0.2_wd_true.csv", # Mistral-7B-Instruct-v0.2(true)
|
872 |
+
"./data/results/Meta-Llama-3-8B-Instruct_wd.csv", # Meta-Llama-3-8b-instruct
|
873 |
+
"./data/results/Meta-Llama-3-8B-Instruct_wd_true.csv", # Meta-Llama-3-8b-instruct(true)
|
874 |
+
"./data/results/Tune_2024-03-25_23-32-57.csv", # Llama-2-13b-chat-hf
|
875 |
+
"./data/results/Llama-2-13b-chat-hf_wd_true.csv", # Llama-2-13b-chat-hf(true)
|
876 |
+
"./data/results/Llama-2-70b-chat-hf_wd.csv", # Llama-2-70b-chat-hf
|
877 |
+
"./data/results/Llama-2-70b-chat-hf_wd_true.csv", # Llama-2-70b-chat-hf
|
878 |
+
"./data/results/Meta-Llama-3-70B-Instruct_wd.csv", # Meta-Llama-3-70B-Instruct
|
879 |
+
"./data/results/Meta-Llama-3-70B-Instruct_wd_true.csv", # Meta-Llama-3-70B-Instruct(true)
|
880 |
+
]
|
881 |
+
|
882 |
+
df_ms_macro = pd.read_json("./data/datasets/ms_macro.json")
|
883 |
+
|
884 |
+
|
885 |
+
def load_for_repetition_penalty_ms_macro(
|
886 |
+
csv_result_file, repetition_penalty, force_recalculate=False
|
887 |
+
):
|
888 |
+
result_file = replace_last(
|
889 |
+
csv_result_file, ".csv", f"_RP_{repetition_penalty:.3f}.csv"
|
890 |
+
)
|
891 |
+
df = load_with_newline_and_repetition_scores(
|
892 |
+
result_file, force_recalculate=force_recalculate
|
893 |
+
)
|
894 |
+
|
895 |
+
if len(df) != len(df_ms_macro):
|
896 |
+
print(f"error: len(df) != {len(df_ms_macro)}")
|
897 |
+
missing_ids = [
|
898 |
+
id for id in df_ms_macro["id"].unique() if id not in df["id"].unique()
|
899 |
+
]
|
900 |
+
print(f"missing_ids: {missing_ids}")
|
901 |
+
|
902 |
+
if df["ground_truth"][0] != str(df_ms_macro["wellFormedAnswers"][0]):
|
903 |
+
df["ground_truth"] = df_ms_macro["wellFormedAnswers"]
|
904 |
+
print("ground_truth updated for:", result_file)
|
905 |
+
df.to_csv(result_file, index=False)
|
906 |
+
return df
|
907 |
+
|
908 |
+
|
909 |
+
# MS MACRO
|
910 |
+
def plot_performance_scores_ms_macro(
|
911 |
+
result,
|
912 |
+
models=None,
|
913 |
+
title="Performance",
|
914 |
+
):
|
915 |
+
|
916 |
+
if models is None:
|
917 |
+
models = result.keys()
|
918 |
+
for model in models:
|
919 |
+
print(f"model: {model}")
|
920 |
+
df = result[model]["df_overall"]
|
921 |
+
# print(result[model]["df_list_repetition_penalty"][0].describe())
|
922 |
+
|
923 |
+
# Calculate the statistics
|
924 |
+
bleu1 = list(df["bleu1"])
|
925 |
+
rougeL = list(df["rougeL"])
|
926 |
+
f1 = [2 * (p * r) / (p + r) for p, r in zip(bleu1, rougeL)]
|
927 |
+
best_f1 = max(f1)
|
928 |
+
best_f1_index = f1.index(best_f1)
|
929 |
+
|
930 |
+
bleu1, rougeL = adjust_perf_scores_with_repetition_penalty(
|
931 |
+
result[model], bleu1, rougeL
|
932 |
+
)
|
933 |
+
afrp = [2 * (p * r) / (p + r) for p, r in zip(bleu1, rougeL)]
|
934 |
+
|
935 |
+
# f1 = [df["f1"].mean() for df in result[model]["df_list_repetition_penalty"]]
|
936 |
+
best_afrp = max(afrp)
|
937 |
+
best_afrp_index = afrp.index(best_afrp)
|
938 |
+
|
939 |
+
repetition_penalties = list(df["repetition_penalty"])
|
940 |
+
|
941 |
+
# line plot for precision, recall, f1
|
942 |
+
plt.figure(figsize=(10, 6))
|
943 |
+
|
944 |
+
plt.axvspan(
|
945 |
+
repetition_penalties[best_f1_index] - 0.01,
|
946 |
+
repetition_penalties[best_f1_index] + 0.01,
|
947 |
+
alpha=0.5,
|
948 |
+
edgecolor="none",
|
949 |
+
facecolor="blue",
|
950 |
+
)
|
951 |
+
|
952 |
+
plt.axvspan(
|
953 |
+
repetition_penalties[best_afrp_index] - 0.01,
|
954 |
+
repetition_penalties[best_afrp_index] + 0.01,
|
955 |
+
alpha=0.5,
|
956 |
+
edgecolor="none",
|
957 |
+
facecolor="orange",
|
958 |
+
)
|
959 |
+
|
960 |
+
plt.plot(
|
961 |
+
repetition_penalties,
|
962 |
+
f1,
|
963 |
+
label="Overall Perf Score",
|
964 |
+
marker="D",
|
965 |
+
color="blue",
|
966 |
+
)
|
967 |
+
plt.plot(
|
968 |
+
repetition_penalties,
|
969 |
+
afrp,
|
970 |
+
label="RF Adjusted Perf Score",
|
971 |
+
marker="o",
|
972 |
+
color="orange",
|
973 |
+
)
|
974 |
+
|
975 |
+
plt.xlabel("Repetition Penalties")
|
976 |
+
plt.ylabel("Score")
|
977 |
+
plt.xlim(0.99, 1.31)
|
978 |
+
# y in percentage
|
979 |
+
plt.gca().yaxis.set_major_formatter(mtick.PercentFormatter(1.0))
|
980 |
+
plt.title(f"{model} {title}")
|
981 |
+
plt.legend(bbox_to_anchor=(1.0, 0.5), loc="center left")
|
982 |
+
|
983 |
+
plt.show()
|
984 |
+
|
985 |
+
|
986 |
+
def plot_repetition_factors(result, groups):
|
987 |
+
for group in groups:
|
988 |
+
# Plot the statistics
|
989 |
+
plt.figure(figsize=(10, 6))
|
990 |
+
|
991 |
+
max_value = 0
|
992 |
+
for model in result.keys():
|
993 |
+
if not group in model.lower():
|
994 |
+
continue
|
995 |
+
print(f"model: {model}")
|
996 |
+
df = result[model]["df_overall"]
|
997 |
+
repetition_panelties = [
|
998 |
+
repetition_penalty for repetition_penalty in df["repetition_penalty"]
|
999 |
+
]
|
1000 |
+
|
1001 |
+
mean_score = [
|
1002 |
+
math.log10(10 + df["total_repetitions"].mean())
|
1003 |
+
for df in result[model]["df_list_repetition_penalty"]
|
1004 |
+
]
|
1005 |
+
|
1006 |
+
sns.lineplot(x=repetition_panelties, y=mean_score, label=model)
|
1007 |
+
|
1008 |
+
new_max = max(mean_score)
|
1009 |
+
if new_max > max_value:
|
1010 |
+
max_value = new_max
|
1011 |
+
|
1012 |
+
max_value = max_value * 1.05
|
1013 |
+
if max_value < 1.5:
|
1014 |
+
max_value = 1.5
|
1015 |
+
# set ylimit
|
1016 |
+
plt.ylim(1, max_value)
|
1017 |
+
|
1018 |
+
# show grid
|
1019 |
+
plt.grid(True)
|
1020 |
+
plt.xlabel("Repetition Penalties")
|
1021 |
+
plt.ylabel("Repetition Factors")
|
1022 |
+
plt.title("Repetition Factors vs Repetition Penalties")
|
1023 |
+
plt.legend()
|
1024 |
+
|
1025 |
+
plt.show()
|
1026 |
+
|
1027 |
+
|
1028 |
+
def plot_repetition_factors_by_group(result, group_filter=None):
|
1029 |
+
markers = ["D", "o", "s", "x"]
|
1030 |
+
colors = ["blue", "orange", "green", "red"]
|
1031 |
+
|
1032 |
+
# Plot the statistics
|
1033 |
+
plt.figure(figsize=(10, 6))
|
1034 |
+
index = 0
|
1035 |
+
max_value = 0
|
1036 |
+
|
1037 |
+
for model in result.keys():
|
1038 |
+
if group_filter is not None and group_filter not in model:
|
1039 |
+
continue
|
1040 |
+
|
1041 |
+
print(f"model: {model}")
|
1042 |
+
|
1043 |
+
df = result[model]["df_overall"]
|
1044 |
+
repetition_panelties = [
|
1045 |
+
repetition_penalty for repetition_penalty in df["repetition_penalty"]
|
1046 |
+
]
|
1047 |
+
|
1048 |
+
# Calculate the statistics
|
1049 |
+
mean_score = [
|
1050 |
+
math.log10(10 + df["total_repetitions"].mean())
|
1051 |
+
for df in result[model]["df_list_repetition_penalty"]
|
1052 |
+
]
|
1053 |
+
if len(mean_score) != len(repetition_panelties):
|
1054 |
+
print(
|
1055 |
+
f"model: {model} has different length of repetition penalties and mean score"
|
1056 |
+
)
|
1057 |
+
print("repetition_panelties:", len(repetition_panelties))
|
1058 |
+
print("mean_score:", len(mean_score))
|
1059 |
+
continue
|
1060 |
+
|
1061 |
+
new_max = max(mean_score)
|
1062 |
+
if new_max > max_value:
|
1063 |
+
max_value = new_max
|
1064 |
+
|
1065 |
+
sns.lineplot(
|
1066 |
+
x=repetition_panelties,
|
1067 |
+
y=mean_score,
|
1068 |
+
label=model,
|
1069 |
+
marker=markers[index],
|
1070 |
+
color=colors[index],
|
1071 |
+
)
|
1072 |
+
|
1073 |
+
index += 1
|
1074 |
+
|
1075 |
+
max_value = max_value * 1.05
|
1076 |
+
if max_value < 1.5:
|
1077 |
+
max_value = 1.5
|
1078 |
+
# set ylimit
|
1079 |
+
plt.ylim(1, max_value)
|
1080 |
+
max_value = 0
|
1081 |
+
|
1082 |
+
plt.xlabel("Repetition Penalties")
|
1083 |
+
plt.ylabel("Repetition Factors")
|
1084 |
+
plt.title("Repetition Factors vs Repetition Penalties")
|
1085 |
+
plt.legend(bbox_to_anchor=(1.0, 0.5), loc="center left")
|
1086 |
+
|
1087 |
+
plt.show()
|
eval_modules/calc_repetitions_v3.py
ADDED
@@ -0,0 +1,1095 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
1 |
+
import os
|
2 |
+
import re
|
3 |
+
import math
|
4 |
+
import pandas as pd
|
5 |
+
import numpy as np
|
6 |
+
import matplotlib.pyplot as plt
|
7 |
+
import matplotlib.ticker as mtick
|
8 |
+
import seaborn as sns
|
9 |
+
import nltk
|
10 |
+
|
11 |
+
# final version
|
12 |
+
pattern_abnormal_newlines = re.compile(r"\n{5,}")
|
13 |
+
pattern_text_repetitions = re.compile(r"(.{5}.*)\s*((\1)\s*)+", re.M | re.DOTALL)
|
14 |
+
exception_patterns = [
|
15 |
+
re.compile(r"(\w+\.?)(\1)+$"),
|
16 |
+
re.compile(r"\W*(wink|nudge|Virginia)\W*((\1)\W*)+$"),
|
17 |
+
re.compile(r"\s+$"),
|
18 |
+
]
|
19 |
+
|
20 |
+
|
21 |
+
# final version for repetition detection
|
22 |
+
def detect_repetitions(
|
23 |
+
text, debug=False, pattern_text_repetitions=pattern_text_repetitions
|
24 |
+
):
|
25 |
+
subtotals = [0, 0]
|
26 |
+
|
27 |
+
if isinstance(text, str):
|
28 |
+
patterns = [pattern_abnormal_newlines, pattern_text_repetitions]
|
29 |
+
for i, pattern in enumerate(patterns):
|
30 |
+
if debug:
|
31 |
+
print(
|
32 |
+
f"----detect {'abnormal newlines' if i == 0 else 'text repetitions'}----"
|
33 |
+
)
|
34 |
+
matches = pattern.finditer(text)
|
35 |
+
for match in matches:
|
36 |
+
if i > 0:
|
37 |
+
ignored = False
|
38 |
+
for exception_pattern in exception_patterns:
|
39 |
+
match_ex = exception_pattern.match(match[0])
|
40 |
+
if match_ex:
|
41 |
+
if debug:
|
42 |
+
print("ignored: ", match[0])
|
43 |
+
print("exception: ", match_ex)
|
44 |
+
ignored = True
|
45 |
+
break
|
46 |
+
if ignored:
|
47 |
+
continue
|
48 |
+
|
49 |
+
if debug:
|
50 |
+
print(match)
|
51 |
+
for groupNum in range(0, len(match.groups())):
|
52 |
+
groupNum = groupNum + 1
|
53 |
+
print(
|
54 |
+
"Group {groupNum} found at {start}-{end}: `{group}`".format(
|
55 |
+
groupNum=groupNum,
|
56 |
+
start=match.start(groupNum),
|
57 |
+
end=match.end(groupNum),
|
58 |
+
group=match.group(groupNum),
|
59 |
+
)
|
60 |
+
)
|
61 |
+
|
62 |
+
start, end = match.span()
|
63 |
+
subtotals[i] += end - start
|
64 |
+
|
65 |
+
if i == 0:
|
66 |
+
text = text.strip()
|
67 |
+
if subtotals[i] > 0:
|
68 |
+
text = pattern.sub("", text)
|
69 |
+
if debug:
|
70 |
+
print(f"removed abnormal newlines: {subtotals[i]}")
|
71 |
+
|
72 |
+
result = (subtotals[0], subtotals[1], subtotals[0] + subtotals[1])
|
73 |
+
|
74 |
+
if debug:
|
75 |
+
print(result)
|
76 |
+
return result
|
77 |
+
|
78 |
+
|
79 |
+
def detect_abnormal_newlines(text, debug=False):
|
80 |
+
return detect_repetitions(text, debug=debug)[0]
|
81 |
+
|
82 |
+
|
83 |
+
def detect_text_repetitions(text, debug=False):
|
84 |
+
return detect_repetitions(text, debug=debug)[1]
|
85 |
+
|
86 |
+
|
87 |
+
def detect_scores(text, debug=False):
|
88 |
+
newline_score, repetition_score, total_repetitions = detect_repetitions(
|
89 |
+
text, debug=debug
|
90 |
+
)
|
91 |
+
return pd.Series([newline_score, repetition_score, total_repetitions])
|
92 |
+
|
93 |
+
|
94 |
+
def load_with_newline_and_repetition_scores(result_file, force_recalculate=False):
|
95 |
+
print(f"loading result file: {result_file}")
|
96 |
+
df = pd.read_csv(result_file, comment="#", on_bad_lines="warn")
|
97 |
+
|
98 |
+
if (
|
99 |
+
force_recalculate
|
100 |
+
or "newline_score" not in df.columns
|
101 |
+
or "repetition_score" not in df.columns
|
102 |
+
or "total_repetitions" not in df.columns
|
103 |
+
):
|
104 |
+
df[["newline_score", "repetition_score", "total_repetitions"]] = df[
|
105 |
+
"answer"
|
106 |
+
].apply(detect_scores)
|
107 |
+
df.to_csv(result_file, index=False)
|
108 |
+
|
109 |
+
return df
|
110 |
+
|
111 |
+
|
112 |
+
def replace_last(source_string, old_string, new_string):
|
113 |
+
head, _sep, tail = source_string.rpartition(old_string)
|
114 |
+
return head + new_string + tail
|
115 |
+
|
116 |
+
|
117 |
+
def load_for_repetition_penalty(
|
118 |
+
csv_result_file, repetition_penalty, force_recalculate=False
|
119 |
+
):
|
120 |
+
result_file = replace_last(
|
121 |
+
csv_result_file, ".csv", f"_RP_{repetition_penalty:.3f}.csv"
|
122 |
+
)
|
123 |
+
return load_with_newline_and_repetition_scores(
|
124 |
+
result_file, force_recalculate=force_recalculate
|
125 |
+
)
|
126 |
+
|
127 |
+
|
128 |
+
def calc_adjusted_performance(f, r):
|
129 |
+
return f / math.log10(10 + r)
|
130 |
+
|
131 |
+
|
132 |
+
def calculate_adjusted_performance(row):
|
133 |
+
r = row["total_repetitions"]
|
134 |
+
adjusted_precision = calc_adjusted_performance(row["precision"], r)
|
135 |
+
adjusted_recall = calc_adjusted_performance(row["recall"], r)
|
136 |
+
return pd.Series([adjusted_precision, adjusted_recall])
|
137 |
+
|
138 |
+
|
139 |
+
def load_performance_df(csv_result_file, repetition_penalty):
|
140 |
+
result_file = replace_last(
|
141 |
+
csv_result_file, ".csv", f"_RP_{repetition_penalty:.3f}-t2_evaluated.json"
|
142 |
+
)
|
143 |
+
result_file = result_file.replace("/results/", "/eval/")
|
144 |
+
print(f"loading json file: {result_file}")
|
145 |
+
df = pd.read_json(result_file)
|
146 |
+
|
147 |
+
return df
|
148 |
+
|
149 |
+
|
150 |
+
def calculate_performance_score_v1(
|
151 |
+
csv_result_file, repetition_penalty, force_recalculate=False
|
152 |
+
):
|
153 |
+
result_file = replace_last(
|
154 |
+
csv_result_file, ".csv", f"_RP_{repetition_penalty:.3f}.csv"
|
155 |
+
)
|
156 |
+
print(f"loading result file: {result_file}")
|
157 |
+
df = pd.read_csv(result_file, comment="#", on_bad_lines="warn")
|
158 |
+
|
159 |
+
if force_recalculate or "f2" in df.columns or "f1" not in df.columns:
|
160 |
+
df.drop(
|
161 |
+
columns=[
|
162 |
+
"precision",
|
163 |
+
"recall",
|
164 |
+
"f1",
|
165 |
+
"f2",
|
166 |
+
"entities_in_answer",
|
167 |
+
"entities_in_question",
|
168 |
+
],
|
169 |
+
errors="ignore",
|
170 |
+
inplace=True,
|
171 |
+
)
|
172 |
+
perf_df = load_performance_df(csv_result_file, repetition_penalty)
|
173 |
+
filtered_df = perf_df[perf_df["id"].isin(df["id"])]
|
174 |
+
perf_df = filtered_df.reset_index(drop=True)
|
175 |
+
print(f"perf_df len: {len(perf_df)}")
|
176 |
+
# print(perf_df.head())
|
177 |
+
|
178 |
+
df["eval_gemini_1.0_pro"] = perf_df["eval_gemini_1.0_pro"]
|
179 |
+
|
180 |
+
df["precision"] = perf_df["score"].apply(lambda x: x[0])
|
181 |
+
df["recall"] = perf_df["score"].apply(lambda x: x[1])
|
182 |
+
df["f1"] = perf_df["score"].apply(lambda x: x[2])
|
183 |
+
|
184 |
+
df[["adjusted_precision", "adjusted_recall"]] = df.apply(
|
185 |
+
calculate_adjusted_performance, axis=1
|
186 |
+
)
|
187 |
+
|
188 |
+
df.to_csv(result_file, index=False)
|
189 |
+
print(f"performance scores saved to result file: {result_file}")
|
190 |
+
|
191 |
+
print(f"df len: {len(df)}")
|
192 |
+
|
193 |
+
return df
|
194 |
+
|
195 |
+
|
196 |
+
ref_df = pd.read_csv(
|
197 |
+
"./data/results/gpt-3.5-turbo_non_rag.csv", comment="#", on_bad_lines="warn"
|
198 |
+
)
|
199 |
+
|
200 |
+
|
201 |
+
def calculate_performance_score(
|
202 |
+
csv_result_file, repetition_penalty, force_recalculate=False
|
203 |
+
):
|
204 |
+
result_file = replace_last(
|
205 |
+
csv_result_file, ".csv", f"_RP_{repetition_penalty:.3f}.csv"
|
206 |
+
)
|
207 |
+
|
208 |
+
re_creating = False
|
209 |
+
if os.path.exists(result_file):
|
210 |
+
print(f"loading result file: {result_file}")
|
211 |
+
df = pd.read_csv(result_file, comment="#", on_bad_lines="warn")
|
212 |
+
else:
|
213 |
+
print(f"re-creating result file: {result_file}")
|
214 |
+
df = pd.DataFrame()
|
215 |
+
force_recalculate = True
|
216 |
+
|
217 |
+
if force_recalculate or "f2" in df.columns or "f1" not in df.columns:
|
218 |
+
df.drop(
|
219 |
+
columns=[
|
220 |
+
"precision",
|
221 |
+
"recall",
|
222 |
+
"f1",
|
223 |
+
"f2",
|
224 |
+
"entities_in_answer",
|
225 |
+
"entities_in_question",
|
226 |
+
"word_count",
|
227 |
+
],
|
228 |
+
errors="ignore",
|
229 |
+
inplace=True,
|
230 |
+
)
|
231 |
+
perf_df = load_performance_df(csv_result_file, repetition_penalty)
|
232 |
+
filtered_df = perf_df[perf_df["id"].isin(ref_df["id"])]
|
233 |
+
perf_df = filtered_df.reset_index(drop=True)
|
234 |
+
print(f"perf_df len: {len(perf_df)}")
|
235 |
+
|
236 |
+
if len(perf_df) != len(ref_df):
|
237 |
+
print(f"error: len(perf_df) != {len(ref_df)}")
|
238 |
+
missing_ids = [
|
239 |
+
id for id in ref_df["id"].unique() if id not in perf_df["id"].unique()
|
240 |
+
]
|
241 |
+
print(f"missing_ids: {missing_ids}")
|
242 |
+
|
243 |
+
# print(perf_df.head())
|
244 |
+
|
245 |
+
df["id"] = perf_df["id"]
|
246 |
+
df["question"] = perf_df["question"]
|
247 |
+
df["answer"] = perf_df["pred_answer"]
|
248 |
+
df["word_count"] = df["answer"].apply(
|
249 |
+
lambda x: len(nltk.word_tokenize(x)) if isinstance(x, str) else 0
|
250 |
+
)
|
251 |
+
df["ground_truth"] = perf_df["ground_truth"]
|
252 |
+
df[["newline_score", "repetition_score", "total_repetitions"]] = df[
|
253 |
+
"answer"
|
254 |
+
].apply(detect_scores)
|
255 |
+
|
256 |
+
df["eval_gemini_1.0_pro"] = perf_df["eval_gemini_1.0_pro"]
|
257 |
+
df["precision"] = perf_df["score"].apply(lambda x: x[0])
|
258 |
+
df["recall"] = perf_df["score"].apply(lambda x: x[1])
|
259 |
+
df["f1"] = perf_df["score"].apply(lambda x: x[2])
|
260 |
+
|
261 |
+
df[["adjusted_precision", "adjusted_recall"]] = df.apply(
|
262 |
+
calculate_adjusted_performance, axis=1
|
263 |
+
)
|
264 |
+
|
265 |
+
df.to_csv(result_file, index=False)
|
266 |
+
print(f"performance scores saved to result file: {result_file}")
|
267 |
+
|
268 |
+
print(f"df len: {len(df)}")
|
269 |
+
|
270 |
+
return df
|
271 |
+
|
272 |
+
|
273 |
+
def adjust_perf_scores_with_repetition_penalty(result, precision, recall):
|
274 |
+
newline_score = [
|
275 |
+
df["newline_score"].mean() for df in result["df_list_repetition_penalty"]
|
276 |
+
]
|
277 |
+
print(f"newline_score: {newline_score}")
|
278 |
+
|
279 |
+
repetition_score = [
|
280 |
+
df["repetition_score"].mean() for df in result["df_list_repetition_penalty"]
|
281 |
+
]
|
282 |
+
print(f"repetition_score: {repetition_score}")
|
283 |
+
|
284 |
+
precision = [
|
285 |
+
f / math.log10(10 + n + r)
|
286 |
+
for f, n, r in zip(precision, newline_score, repetition_score)
|
287 |
+
]
|
288 |
+
recall = [
|
289 |
+
f / math.log10(10 + n + r)
|
290 |
+
for f, n, r in zip(recall, newline_score, repetition_score)
|
291 |
+
]
|
292 |
+
|
293 |
+
return precision, recall
|
294 |
+
|
295 |
+
|
296 |
+
def plot_performance_scores(
|
297 |
+
result,
|
298 |
+
models=None,
|
299 |
+
title="Performance",
|
300 |
+
):
|
301 |
+
|
302 |
+
if models is None:
|
303 |
+
models = result.keys()
|
304 |
+
for model in models:
|
305 |
+
print(f"model: {model}")
|
306 |
+
df = result[model]["df_overall"]
|
307 |
+
|
308 |
+
# Calculate the statistics
|
309 |
+
precision = [
|
310 |
+
df["precision"].mean() for df in result[model]["df_list_repetition_penalty"]
|
311 |
+
]
|
312 |
+
recall = [
|
313 |
+
df["recall"].mean() for df in result[model]["df_list_repetition_penalty"]
|
314 |
+
]
|
315 |
+
f1 = [2 * (p * r) / (p + r) for p, r in zip(precision, recall)]
|
316 |
+
best_f1 = max(f1)
|
317 |
+
best_f1_index = f1.index(best_f1)
|
318 |
+
|
319 |
+
precision, recall = adjust_perf_scores_with_repetition_penalty(
|
320 |
+
result[model], precision, recall
|
321 |
+
)
|
322 |
+
afrp = [2 * (p * r) / (p + r) for p, r in zip(precision, recall)]
|
323 |
+
|
324 |
+
# f1 = [df["f1"].mean() for df in result[model]["df_list_repetition_penalty"]]
|
325 |
+
best_afrp = max(afrp)
|
326 |
+
best_afrp_index = afrp.index(best_afrp)
|
327 |
+
|
328 |
+
adjusted_precision = [
|
329 |
+
df["adjusted_precision"].mean()
|
330 |
+
for df in result[model]["df_list_repetition_penalty"]
|
331 |
+
]
|
332 |
+
adjusted_recall = [
|
333 |
+
df["adjusted_recall"].mean()
|
334 |
+
for df in result[model]["df_list_repetition_penalty"]
|
335 |
+
]
|
336 |
+
afrp2 = [
|
337 |
+
2 * (p * r) / (p + r) for p, r in zip(adjusted_precision, adjusted_recall)
|
338 |
+
]
|
339 |
+
best_afrp2 = max(afrp2)
|
340 |
+
best_afrp2_index = afrp2.index(best_afrp2)
|
341 |
+
|
342 |
+
repetition_penalties = list(df["repetition_penalty"])
|
343 |
+
|
344 |
+
# line plot for precision, recall, f1
|
345 |
+
plt.figure(figsize=(10, 6))
|
346 |
+
|
347 |
+
plt.axvspan(
|
348 |
+
repetition_penalties[best_f1_index] - 0.01,
|
349 |
+
repetition_penalties[best_f1_index] + 0.01,
|
350 |
+
alpha=0.5,
|
351 |
+
edgecolor="none",
|
352 |
+
facecolor="blue",
|
353 |
+
)
|
354 |
+
|
355 |
+
# plt.axvspan(
|
356 |
+
# repetition_penalties[best_afrp2_index] - 0.01,
|
357 |
+
# repetition_penalties[best_afrp2_index] + 0.01,
|
358 |
+
# alpha=0.5,
|
359 |
+
# edgecolor="none",
|
360 |
+
# facecolor="green",
|
361 |
+
# )
|
362 |
+
|
363 |
+
plt.axvspan(
|
364 |
+
repetition_penalties[best_afrp_index] - 0.01,
|
365 |
+
repetition_penalties[best_afrp_index] + 0.01,
|
366 |
+
alpha=0.5,
|
367 |
+
edgecolor="none",
|
368 |
+
facecolor="orange",
|
369 |
+
)
|
370 |
+
|
371 |
+
plt.plot(repetition_penalties, f1, label="F1", marker="D", color="blue")
|
372 |
+
# plt.plot(
|
373 |
+
# repetition_penalties,
|
374 |
+
# afrp2,
|
375 |
+
# label="Per-question RF Adjusted F1",
|
376 |
+
# marker="s",
|
377 |
+
# color="green",
|
378 |
+
# )
|
379 |
+
plt.plot(
|
380 |
+
repetition_penalties,
|
381 |
+
afrp,
|
382 |
+
label="RF Adjusted F1",
|
383 |
+
marker="o",
|
384 |
+
color="orange",
|
385 |
+
)
|
386 |
+
plt.xlabel("Repetition Penalties")
|
387 |
+
plt.ylabel("Score")
|
388 |
+
plt.xlim(0.99, 1.31)
|
389 |
+
# y in percentage
|
390 |
+
plt.gca().yaxis.set_major_formatter(mtick.PercentFormatter(1.0))
|
391 |
+
plt.title(f"{model} {title}")
|
392 |
+
plt.legend(bbox_to_anchor=(1.0, 0.5), loc="center left")
|
393 |
+
|
394 |
+
plt.show()
|
395 |
+
|
396 |
+
|
397 |
+
def plot_best_afrp(
|
398 |
+
result,
|
399 |
+
models=None,
|
400 |
+
title="Models with Best Repetition Factor Adjusted F1",
|
401 |
+
ref_result=None,
|
402 |
+
):
|
403 |
+
# Initialize lists to store the statistics
|
404 |
+
model_names = []
|
405 |
+
best_f1 = []
|
406 |
+
best_afrp = []
|
407 |
+
best_repetition_penalty = []
|
408 |
+
|
409 |
+
if models is None:
|
410 |
+
models = result.keys()
|
411 |
+
for model in models:
|
412 |
+
print(f"model: {model}")
|
413 |
+
df = result[model]["df_overall"]
|
414 |
+
|
415 |
+
# Calculate the statistics
|
416 |
+
precision = [
|
417 |
+
df["precision"].mean() for df in result[model]["df_list_repetition_penalty"]
|
418 |
+
]
|
419 |
+
recall = [
|
420 |
+
df["recall"].mean() for df in result[model]["df_list_repetition_penalty"]
|
421 |
+
]
|
422 |
+
# f1 = [df["f1"].mean() for df in result[model]["df_list_repetition_penalty"]]
|
423 |
+
f1 = [2 * (p * r) / (p + r) for p, r in zip(precision, recall)]
|
424 |
+
|
425 |
+
newline_score = [
|
426 |
+
df["newline_score"].mean()
|
427 |
+
for df in result[model]["df_list_repetition_penalty"]
|
428 |
+
]
|
429 |
+
print(f"newline_score: {newline_score}")
|
430 |
+
|
431 |
+
repetition_score = [
|
432 |
+
df["repetition_score"].mean()
|
433 |
+
for df in result[model]["df_list_repetition_penalty"]
|
434 |
+
]
|
435 |
+
print(f"repetition_score: {repetition_score}")
|
436 |
+
|
437 |
+
afrp = [
|
438 |
+
f / math.log10(10 + n + r)
|
439 |
+
for f, n, r in zip(f1, newline_score, repetition_score)
|
440 |
+
]
|
441 |
+
|
442 |
+
best_afrp.append(max(afrp))
|
443 |
+
best_afrp_index = afrp.index(best_afrp[-1])
|
444 |
+
best_repetition_penalty.append(df["repetition_penalty"][best_afrp_index])
|
445 |
+
|
446 |
+
best_f1.append(f1[best_afrp_index])
|
447 |
+
|
448 |
+
print(
|
449 |
+
f"best repetition penalty: {best_repetition_penalty[-1]}, best afrp: {best_afrp[-1]}, f1: {best_f1[-1]}"
|
450 |
+
)
|
451 |
+
|
452 |
+
df = result[model]["df_list_repetition_penalty"][best_afrp_index]
|
453 |
+
|
454 |
+
model_names.append(
|
455 |
+
f"{model} (RP={best_repetition_penalty[-1]})"
|
456 |
+
) # Add the model name to the list
|
457 |
+
|
458 |
+
if ref_result is not None:
|
459 |
+
print("ref_result:", ref_result)
|
460 |
+
for model in ref_result.keys():
|
461 |
+
model_names.append(model)
|
462 |
+
df = pd.read_csv(ref_result[model])
|
463 |
+
# df = df[df["id"].isin(wikidata_df["id"])]
|
464 |
+
|
465 |
+
p = df["precision"].mean()
|
466 |
+
r = df["recall"].mean()
|
467 |
+
|
468 |
+
f1 = 2 * p * r / (p + r) if p + r > 0 else 0
|
469 |
+
best_f1.append(f1)
|
470 |
+
best_afrp.append(f1)
|
471 |
+
|
472 |
+
print("model_names:", model_names)
|
473 |
+
print("best_f1:", best_f1)
|
474 |
+
print("best_afrp:", best_afrp)
|
475 |
+
|
476 |
+
# Create a DataFrame with the statistics
|
477 |
+
data = pd.DataFrame(
|
478 |
+
{
|
479 |
+
"Model": model_names,
|
480 |
+
"Repetition Factor Adjusted F1": best_afrp,
|
481 |
+
"F1": best_f1,
|
482 |
+
}
|
483 |
+
)
|
484 |
+
|
485 |
+
# Melt the DataFrame to a long format
|
486 |
+
data_melted = data.melt(id_vars="Model", var_name="Metric", value_name="Score")
|
487 |
+
|
488 |
+
# Pivot the DataFrame to a wide format
|
489 |
+
data_pivoted = data_melted.pivot(index="Metric", columns="Model", values="Score")
|
490 |
+
|
491 |
+
# make sure the columns are following the order of the models
|
492 |
+
data_pivoted = data_pivoted[model_names]
|
493 |
+
|
494 |
+
# make sure three groups in the order of precision, recall, f1
|
495 |
+
data_pivoted = data_pivoted.reindex(["Repetition Factor Adjusted F1", "F1"])
|
496 |
+
|
497 |
+
# Plot the statistics
|
498 |
+
plt.figure(figsize=(15, 6))
|
499 |
+
ax = data_pivoted.plot(kind="bar", ax=plt.gca(), width=0.9)
|
500 |
+
plt.title(title)
|
501 |
+
plt.legend(bbox_to_anchor=(1.0, 0.5), loc="center left")
|
502 |
+
|
503 |
+
# Set the rotation of the x-axis labels to 0 degrees
|
504 |
+
plt.xticks(rotation=0)
|
505 |
+
|
506 |
+
# Format the y-axis to display as percentage
|
507 |
+
ax.yaxis.set_major_formatter(mtick.PercentFormatter(1.0))
|
508 |
+
|
509 |
+
# get the max value of the y-axis
|
510 |
+
a1 = max(best_afrp)
|
511 |
+
a2 = max(best_f1)
|
512 |
+
|
513 |
+
max_value = max([a1, a2]) * 1.12
|
514 |
+
print("max_value:", max_value)
|
515 |
+
|
516 |
+
# Set the y-axis limit up to 70%
|
517 |
+
ax.set_ylim(0, max_value)
|
518 |
+
|
519 |
+
# Add the values above each bar
|
520 |
+
for p in ax.patches:
|
521 |
+
ax.annotate(
|
522 |
+
f"{p.get_height() * 100:.1f}",
|
523 |
+
(p.get_x() + p.get_width() / 2.0, p.get_height()),
|
524 |
+
ha="center",
|
525 |
+
va="bottom",
|
526 |
+
xytext=(0, 10),
|
527 |
+
textcoords="offset points",
|
528 |
+
rotation=90,
|
529 |
+
)
|
530 |
+
|
531 |
+
plt.show()
|
532 |
+
|
533 |
+
|
534 |
+
def plot_best_performance(
|
535 |
+
result,
|
536 |
+
models=None,
|
537 |
+
title="Models with Best F1 Score",
|
538 |
+
adjusted_f1=False,
|
539 |
+
ref_result=None,
|
540 |
+
):
|
541 |
+
# Initialize lists to store the statistics
|
542 |
+
model_names = []
|
543 |
+
best_precision = []
|
544 |
+
best_recall = []
|
545 |
+
best_f1 = []
|
546 |
+
best_repetition_penalty = []
|
547 |
+
|
548 |
+
if models is None:
|
549 |
+
models = result.keys()
|
550 |
+
for model in models:
|
551 |
+
print(f"model: {model}")
|
552 |
+
df = result[model]["df_overall"]
|
553 |
+
|
554 |
+
# Calculate the statistics
|
555 |
+
precision = [
|
556 |
+
df["precision"].mean() for df in result[model]["df_list_repetition_penalty"]
|
557 |
+
]
|
558 |
+
recall = [
|
559 |
+
df["recall"].mean() for df in result[model]["df_list_repetition_penalty"]
|
560 |
+
]
|
561 |
+
|
562 |
+
if adjusted_f1:
|
563 |
+
precision, recall = adjust_perf_scores_with_repetition_penalty(
|
564 |
+
result[model], precision, recall
|
565 |
+
)
|
566 |
+
|
567 |
+
# f1 = [df["f1"].mean() for df in result[model]["df_list_repetition_penalty"]]
|
568 |
+
f1 = [2 * (p * r) / (p + r) for p, r in zip(precision, recall)]
|
569 |
+
|
570 |
+
best_f1.append(max(f1))
|
571 |
+
best_f1_index = f1.index(best_f1[-1])
|
572 |
+
best_repetition_penalty.append(df["repetition_penalty"][best_f1_index])
|
573 |
+
|
574 |
+
best_precision.append(precision[best_f1_index])
|
575 |
+
best_recall.append(recall[best_f1_index])
|
576 |
+
|
577 |
+
print(
|
578 |
+
f"best repetition penalty: {best_repetition_penalty[-1]}, best f1: {best_f1[-1]}, precision: {best_precision[-1]}, recall: {best_recall[-1]}"
|
579 |
+
)
|
580 |
+
|
581 |
+
df = result[model]["df_list_repetition_penalty"][best_f1_index]
|
582 |
+
|
583 |
+
model_names.append(
|
584 |
+
f"{model} (RP={best_repetition_penalty[-1]})"
|
585 |
+
) # Add the model name to the list
|
586 |
+
|
587 |
+
# print sum for columns: newline_score, repetition_score
|
588 |
+
print(
|
589 |
+
f"newline_score: {df['newline_score'].sum()}, repetition_score: {df['repetition_score'].sum()}"
|
590 |
+
)
|
591 |
+
|
592 |
+
if ref_result is not None:
|
593 |
+
print("ref_result:", ref_result)
|
594 |
+
for model in ref_result.keys():
|
595 |
+
model_names.append(model)
|
596 |
+
df = pd.read_csv(ref_result[model])
|
597 |
+
# df = df[df["id"].isin(wikidata_df["id"])]
|
598 |
+
|
599 |
+
best_precision.append(df["precision"].mean())
|
600 |
+
best_recall.append(df["recall"].mean())
|
601 |
+
f1 = (
|
602 |
+
2
|
603 |
+
* (best_precision[-1] * best_recall[-1])
|
604 |
+
/ (best_precision[-1] + best_recall[-1])
|
605 |
+
)
|
606 |
+
# best_f1.append(df["f1"].mean())
|
607 |
+
best_f1.append(f1)
|
608 |
+
|
609 |
+
# Create a DataFrame with the statistics
|
610 |
+
data = (
|
611 |
+
pd.DataFrame(
|
612 |
+
{
|
613 |
+
"Model": model_names,
|
614 |
+
"Adjusted Precision with RP": best_precision,
|
615 |
+
"Adjusted Recall with RP": best_recall,
|
616 |
+
"Adjusted F1 with RP": best_f1,
|
617 |
+
}
|
618 |
+
)
|
619 |
+
if adjusted_f1
|
620 |
+
else pd.DataFrame(
|
621 |
+
{
|
622 |
+
"Model": model_names,
|
623 |
+
"Precision": best_precision,
|
624 |
+
"Recall": best_recall,
|
625 |
+
"F1": best_f1,
|
626 |
+
}
|
627 |
+
)
|
628 |
+
)
|
629 |
+
columns = list(data.columns)
|
630 |
+
|
631 |
+
# Melt the DataFrame to a long format
|
632 |
+
data_melted = data.melt(id_vars="Model", var_name="Metric", value_name="Score")
|
633 |
+
|
634 |
+
# Pivot the DataFrame to a wide format
|
635 |
+
data_pivoted = data_melted.pivot(index="Metric", columns="Model", values="Score")
|
636 |
+
|
637 |
+
# make sure the columns are following the order of the models
|
638 |
+
data_pivoted = data_pivoted[model_names]
|
639 |
+
|
640 |
+
# make sure three groups in the order of precision, recall, f1
|
641 |
+
data_pivoted = data_pivoted.reindex(columns[1:])
|
642 |
+
|
643 |
+
# Plot the statistics
|
644 |
+
plt.figure(figsize=(10, 6))
|
645 |
+
ax = data_pivoted.plot(kind="bar", ax=plt.gca(), width=0.9)
|
646 |
+
plt.title(title)
|
647 |
+
plt.legend(bbox_to_anchor=(1.0, 0.5), loc="center left")
|
648 |
+
|
649 |
+
# Set the rotation of the x-axis labels to 0 degrees
|
650 |
+
plt.xticks(rotation=0)
|
651 |
+
|
652 |
+
# Format the y-axis to display as percentage
|
653 |
+
ax.yaxis.set_major_formatter(mtick.PercentFormatter(1.0))
|
654 |
+
|
655 |
+
# get the max value of the y-axis
|
656 |
+
a1 = max(best_precision)
|
657 |
+
a2 = max(best_recall)
|
658 |
+
a3 = max(best_f1)
|
659 |
+
|
660 |
+
max_value = max([a1, a2, a3]) * 1.12
|
661 |
+
print("max_value:", max_value)
|
662 |
+
|
663 |
+
# Set the y-axis limit up to 70%
|
664 |
+
ax.set_ylim(0, max_value)
|
665 |
+
|
666 |
+
# Add the values above each bar
|
667 |
+
for p in ax.patches:
|
668 |
+
ax.annotate(
|
669 |
+
f"{p.get_height() * 100:.1f}",
|
670 |
+
(p.get_x() + p.get_width() / 2.0, p.get_height()),
|
671 |
+
ha="center",
|
672 |
+
va="bottom",
|
673 |
+
xytext=(0, 10),
|
674 |
+
textcoords="offset points",
|
675 |
+
rotation=90,
|
676 |
+
)
|
677 |
+
|
678 |
+
plt.show()
|
679 |
+
|
680 |
+
|
681 |
+
def plot_best_performance_ms_macro(
|
682 |
+
result,
|
683 |
+
models=None,
|
684 |
+
title="Models with Best Repetition Factor Adjusted Performance",
|
685 |
+
ref_result=None,
|
686 |
+
skip_generic_prompt=False,
|
687 |
+
include_adjusted_performance=True,
|
688 |
+
):
|
689 |
+
# Initialize lists to store the statistics
|
690 |
+
model_names = []
|
691 |
+
best_f1 = []
|
692 |
+
best_afrp = []
|
693 |
+
best_repetition_penalty = []
|
694 |
+
best_bleu1 = []
|
695 |
+
best_rougeL = []
|
696 |
+
|
697 |
+
if models is None:
|
698 |
+
models = result.keys()
|
699 |
+
for model in models:
|
700 |
+
if skip_generic_prompt and "generic prompt" in model:
|
701 |
+
continue
|
702 |
+
print(f"model: {model}")
|
703 |
+
df = result[model]["df_overall"]
|
704 |
+
|
705 |
+
# Calculate the statistics
|
706 |
+
bleu1 = [x for x in df["bleu1"]]
|
707 |
+
rougeL = [x for x in df["rougeL"]]
|
708 |
+
f1 = [2 * (p * r) / (p + r) for p, r in zip(bleu1, rougeL)]
|
709 |
+
|
710 |
+
newline_score = [
|
711 |
+
df["newline_score"].mean()
|
712 |
+
for df in result[model]["df_list_repetition_penalty"]
|
713 |
+
]
|
714 |
+
print(f"newline_score: {newline_score}")
|
715 |
+
|
716 |
+
repetition_score = [
|
717 |
+
df["repetition_score"].mean()
|
718 |
+
for df in result[model]["df_list_repetition_penalty"]
|
719 |
+
]
|
720 |
+
print(f"repetition_score: {repetition_score}")
|
721 |
+
|
722 |
+
afrp = [
|
723 |
+
f / math.log10(10 + n + r)
|
724 |
+
for f, n, r in zip(f1, newline_score, repetition_score)
|
725 |
+
]
|
726 |
+
|
727 |
+
best_afrp.append(max(afrp if include_adjusted_performance else f1))
|
728 |
+
best_afrp_index = (
|
729 |
+
afrp.index(best_afrp[-1])
|
730 |
+
if include_adjusted_performance
|
731 |
+
else f1.index(best_afrp[-1])
|
732 |
+
)
|
733 |
+
best_repetition_penalty.append(df["repetition_penalty"][best_afrp_index])
|
734 |
+
|
735 |
+
best_f1.append(f1[best_afrp_index])
|
736 |
+
best_bleu1.append(bleu1[best_afrp_index])
|
737 |
+
best_rougeL.append(rougeL[best_afrp_index])
|
738 |
+
|
739 |
+
print(
|
740 |
+
f"best repetition penalty: {best_repetition_penalty[-1]}, best afrp: {best_afrp[-1]}, f1: {best_f1[-1]}"
|
741 |
+
)
|
742 |
+
|
743 |
+
df = result[model]["df_list_repetition_penalty"][best_afrp_index]
|
744 |
+
|
745 |
+
model_names.append(
|
746 |
+
f"{model} (RP={best_repetition_penalty[-1]})"
|
747 |
+
) # Add the model name to the list
|
748 |
+
|
749 |
+
if ref_result is not None:
|
750 |
+
print("ref_result:", ref_result)
|
751 |
+
for model in ref_result.keys():
|
752 |
+
model_names.append(model)
|
753 |
+
df = pd.read_csv(ref_result[model], comment="#", on_bad_lines="warn")
|
754 |
+
# df = df[df["id"].isin(wikidata_df["id"])]
|
755 |
+
|
756 |
+
p = df["bleu1"][0]
|
757 |
+
best_bleu1.append(p)
|
758 |
+
|
759 |
+
r = df["rougeL"][0]
|
760 |
+
best_rougeL.append(r)
|
761 |
+
|
762 |
+
f1 = 2 * p * r / (p + r) if p + r > 0 else 0
|
763 |
+
best_f1.append(f1)
|
764 |
+
best_afrp.append(f1)
|
765 |
+
|
766 |
+
print("model_names:", model_names)
|
767 |
+
print("best_f1:", best_f1)
|
768 |
+
print("best_afrp:", best_afrp)
|
769 |
+
|
770 |
+
# Create a DataFrame with the statistics
|
771 |
+
data = (
|
772 |
+
pd.DataFrame(
|
773 |
+
{
|
774 |
+
"Model": model_names,
|
775 |
+
"Repetition Factor Adjusted Perf Score": best_afrp,
|
776 |
+
"Overall Perf Score": best_f1,
|
777 |
+
}
|
778 |
+
)
|
779 |
+
if include_adjusted_performance
|
780 |
+
else pd.DataFrame(
|
781 |
+
{
|
782 |
+
"Model": model_names,
|
783 |
+
"Bleu-1": best_bleu1,
|
784 |
+
"Rouge-L": best_rougeL,
|
785 |
+
"Overall Perf Score": best_f1,
|
786 |
+
}
|
787 |
+
)
|
788 |
+
)
|
789 |
+
|
790 |
+
# Melt the DataFrame to a long format
|
791 |
+
data_melted = data.melt(id_vars="Model", var_name="Metric", value_name="Score")
|
792 |
+
|
793 |
+
# Pivot the DataFrame to a wide format
|
794 |
+
data_pivoted = data_melted.pivot(index="Metric", columns="Model", values="Score")
|
795 |
+
|
796 |
+
# make sure the columns are following the order of the models
|
797 |
+
data_pivoted = data_pivoted[model_names]
|
798 |
+
|
799 |
+
columns = list(data.columns)
|
800 |
+
data_pivoted = data_pivoted.reindex(columns[1:])
|
801 |
+
|
802 |
+
# Plot the statistics
|
803 |
+
plt.figure(figsize=(10, 6))
|
804 |
+
ax = data_pivoted.plot(kind="bar", ax=plt.gca(), width=0.9)
|
805 |
+
plt.title(title)
|
806 |
+
plt.legend(bbox_to_anchor=(1.0, 0.5), loc="center left")
|
807 |
+
|
808 |
+
# Set the rotation of the x-axis labels to 0 degrees
|
809 |
+
plt.xticks(rotation=0)
|
810 |
+
|
811 |
+
# Format the y-axis to display as percentage
|
812 |
+
ax.yaxis.set_major_formatter(mtick.PercentFormatter(1.0))
|
813 |
+
|
814 |
+
# get the max value of the y-axis
|
815 |
+
a1 = max(best_afrp)
|
816 |
+
a2 = max(best_f1)
|
817 |
+
a3 = max(best_bleu1)
|
818 |
+
a4 = max(best_rougeL)
|
819 |
+
|
820 |
+
max_value = (
|
821 |
+
max([a1, a2] if include_adjusted_performance else [a1, a2, a3, a4]) * 1.12
|
822 |
+
)
|
823 |
+
print("max_value:", max_value)
|
824 |
+
|
825 |
+
# Set the y-axis limit up to 70%
|
826 |
+
ax.set_ylim(0, max_value)
|
827 |
+
|
828 |
+
# Add the values above each bar
|
829 |
+
for p in ax.patches:
|
830 |
+
ax.annotate(
|
831 |
+
f"{p.get_height() * 100:.1f}",
|
832 |
+
(p.get_x() + p.get_width() / 2.0, p.get_height()),
|
833 |
+
ha="center",
|
834 |
+
va="bottom",
|
835 |
+
xytext=(0, 10),
|
836 |
+
textcoords="offset points",
|
837 |
+
rotation=90,
|
838 |
+
)
|
839 |
+
|
840 |
+
plt.show()
|
841 |
+
|
842 |
+
|
843 |
+
all_open_source_models = [
|
844 |
+
"gemma-1.1-2b-it",
|
845 |
+
"Phi-3-mini-128k-instruct",
|
846 |
+
"gemma-1.1-7b-it",
|
847 |
+
"Llama-2-7b-chat-hf",
|
848 |
+
"Mistral-7B-Instruct-v0.2",
|
849 |
+
"Meta-Llama-3-8B-Instruct",
|
850 |
+
"Llama-2-13b-chat-hf",
|
851 |
+
"Llama-2-70b-chat-hf",
|
852 |
+
"Meta-Llama-3-70B-Instruct",
|
853 |
+
]
|
854 |
+
|
855 |
+
|
856 |
+
non_rag_csv_result_files = [
|
857 |
+
"./data/results/gemma-1.1-2b-it_wd_non_rag.csv", # gemma-1.1-2b-it
|
858 |
+
"./data/results/Phi-3-mini-128k-instruct_wd_non_rag_batch_16.csv", # Phi-3-mini-128k-instruct(batch size:16)
|
859 |
+
"./data/results/gemma-1.1-7b-it_wd_non_rag.csv", # gemma-1.1-7b-it
|
860 |
+
"./data/results/Tune_2024-04-09_09-19-22.csv", # Llama-2-7b-chat-hf
|
861 |
+
"./data/results/Tune_2024-04-16_12-24-27.csv.csv", # Mistral-7B-Instruct-v0.2
|
862 |
+
"./data/results/Meta-Llama-3-8B-Instruct_wd_non_rag.csv", # Meta-Llama-3-8B-Instruct
|
863 |
+
"./data/results/Meta-Llama-3-8B-Instruct_wd_1_non_rag.csv", # Meta-Llama-3-8B-Instruct
|
864 |
+
"./data/results/Tune_2024-04-10_16-53-38.csv", # Llama-2-13b-chat-hf
|
865 |
+
"./data/results/Llama-2-70b-chat-hf_wd_non_rag.csv", # Llama-2-70b-chat-hf
|
866 |
+
"./data/results/Meta-Llama-3-70B-Instruct_wd_non_rag.csv", # Meta-Llama-3-70B-Instruct
|
867 |
+
]
|
868 |
+
|
869 |
+
rag_csv_result_files = [
|
870 |
+
"./data/results/gemma-1.1-2b-it_wd.csv", # gemma-1.1-2b-it
|
871 |
+
"./data/results/gemma-1.1-2b-it_wd_true.csv", # gemma-1.1-2b-it(true)
|
872 |
+
"./data/results/Phi-3-mini-128k-instruct_wd_rag_batch_4.csv", # Phi-3-mini-128k-instruct(batch size:16)
|
873 |
+
"./data/results/Phi-3-mini-128k-instruct_wd_true.csv", # Phi-3-mini-128k-instruct(batch size:16)
|
874 |
+
"./data/results/gemma-1.1-7b-it_wd.csv", # gemma-1.1-7b-it
|
875 |
+
"./data/results/gemma-1.1-7b-it_wd_true.csv", # gemma-1.1-7b-it(true)
|
876 |
+
"./data/results/Tune_2024-03-20_15-35-37.csv", # Llama-2-7b-chat-hf
|
877 |
+
"./data/results/Llama-2-7b-chat-hf_wd_true.csv", # Llama-2-7b-chat-hf(true)
|
878 |
+
"./data/results/Tune_2024-03-29_11-28-20.csv", # Mistral-7B-Instruct-v0.2
|
879 |
+
"./data/results/Mistral-7B-Instruct-v0.2_wd_true.csv", # Mistral-7B-Instruct-v0.2(true)
|
880 |
+
"./data/results/Meta-Llama-3-8B-Instruct_wd.csv", # Meta-Llama-3-8b-instruct
|
881 |
+
"./data/results/Meta-Llama-3-8B-Instruct_wd_true.csv", # Meta-Llama-3-8b-instruct(true)
|
882 |
+
"./data/results/Tune_2024-03-25_23-32-57.csv", # Llama-2-13b-chat-hf
|
883 |
+
"./data/results/Llama-2-13b-chat-hf_wd_true.csv", # Llama-2-13b-chat-hf(true)
|
884 |
+
"./data/results/Llama-2-70b-chat-hf_wd.csv", # Llama-2-70b-chat-hf
|
885 |
+
"./data/results/Llama-2-70b-chat-hf_wd_true.csv", # Llama-2-70b-chat-hf
|
886 |
+
"./data/results/Meta-Llama-3-70B-Instruct_wd.csv", # Meta-Llama-3-70B-Instruct
|
887 |
+
"./data/results/Meta-Llama-3-70B-Instruct_wd_true.csv", # Meta-Llama-3-70B-Instruct(true)
|
888 |
+
]
|
889 |
+
|
890 |
+
df_ms_macro = pd.read_json("./data/datasets/ms_macro.json")
|
891 |
+
|
892 |
+
|
893 |
+
def load_for_repetition_penalty_ms_macro(
|
894 |
+
csv_result_file, repetition_penalty, force_recalculate=False
|
895 |
+
):
|
896 |
+
result_file = replace_last(
|
897 |
+
csv_result_file, ".csv", f"_RP_{repetition_penalty:.3f}.csv"
|
898 |
+
)
|
899 |
+
df = load_with_newline_and_repetition_scores(
|
900 |
+
result_file, force_recalculate=force_recalculate
|
901 |
+
)
|
902 |
+
|
903 |
+
if len(df) != len(df_ms_macro):
|
904 |
+
print(f"error: len(df) != {len(df_ms_macro)}")
|
905 |
+
missing_ids = [
|
906 |
+
id for id in df_ms_macro["id"].unique() if id not in df["id"].unique()
|
907 |
+
]
|
908 |
+
print(f"missing_ids: {missing_ids}")
|
909 |
+
|
910 |
+
if df["ground_truth"][0] != str(df_ms_macro["wellFormedAnswers"][0]):
|
911 |
+
df["ground_truth"] = df_ms_macro["wellFormedAnswers"]
|
912 |
+
print("ground_truth updated for:", result_file)
|
913 |
+
df.to_csv(result_file, index=False)
|
914 |
+
return df
|
915 |
+
|
916 |
+
|
917 |
+
# MS MACRO
|
918 |
+
def plot_performance_scores_ms_macro(
|
919 |
+
result,
|
920 |
+
models=None,
|
921 |
+
title="Performance",
|
922 |
+
):
|
923 |
+
|
924 |
+
if models is None:
|
925 |
+
models = result.keys()
|
926 |
+
for model in models:
|
927 |
+
print(f"model: {model}")
|
928 |
+
df = result[model]["df_overall"]
|
929 |
+
# print(result[model]["df_list_repetition_penalty"][0].describe())
|
930 |
+
|
931 |
+
# Calculate the statistics
|
932 |
+
bleu1 = list(df["bleu1"])
|
933 |
+
rougeL = list(df["rougeL"])
|
934 |
+
f1 = [2 * (p * r) / (p + r) for p, r in zip(bleu1, rougeL)]
|
935 |
+
best_f1 = max(f1)
|
936 |
+
best_f1_index = f1.index(best_f1)
|
937 |
+
|
938 |
+
bleu1, rougeL = adjust_perf_scores_with_repetition_penalty(
|
939 |
+
result[model], bleu1, rougeL
|
940 |
+
)
|
941 |
+
afrp = [2 * (p * r) / (p + r) for p, r in zip(bleu1, rougeL)]
|
942 |
+
|
943 |
+
# f1 = [df["f1"].mean() for df in result[model]["df_list_repetition_penalty"]]
|
944 |
+
best_afrp = max(afrp)
|
945 |
+
best_afrp_index = afrp.index(best_afrp)
|
946 |
+
|
947 |
+
repetition_penalties = list(df["repetition_penalty"])
|
948 |
+
|
949 |
+
# line plot for precision, recall, f1
|
950 |
+
plt.figure(figsize=(10, 6))
|
951 |
+
|
952 |
+
plt.axvspan(
|
953 |
+
repetition_penalties[best_f1_index] - 0.01,
|
954 |
+
repetition_penalties[best_f1_index] + 0.01,
|
955 |
+
alpha=0.5,
|
956 |
+
edgecolor="none",
|
957 |
+
facecolor="blue",
|
958 |
+
)
|
959 |
+
|
960 |
+
plt.axvspan(
|
961 |
+
repetition_penalties[best_afrp_index] - 0.01,
|
962 |
+
repetition_penalties[best_afrp_index] + 0.01,
|
963 |
+
alpha=0.5,
|
964 |
+
edgecolor="none",
|
965 |
+
facecolor="orange",
|
966 |
+
)
|
967 |
+
|
968 |
+
plt.plot(
|
969 |
+
repetition_penalties,
|
970 |
+
f1,
|
971 |
+
label="Overall Perf Score",
|
972 |
+
marker="D",
|
973 |
+
color="blue",
|
974 |
+
)
|
975 |
+
plt.plot(
|
976 |
+
repetition_penalties,
|
977 |
+
afrp,
|
978 |
+
label="RF Adjusted Perf Score",
|
979 |
+
marker="o",
|
980 |
+
color="orange",
|
981 |
+
)
|
982 |
+
|
983 |
+
plt.xlabel("Repetition Penalties")
|
984 |
+
plt.ylabel("Score")
|
985 |
+
plt.xlim(0.99, 1.31)
|
986 |
+
# y in percentage
|
987 |
+
plt.gca().yaxis.set_major_formatter(mtick.PercentFormatter(1.0))
|
988 |
+
plt.title(f"{model} {title}")
|
989 |
+
plt.legend(bbox_to_anchor=(1.0, 0.5), loc="center left")
|
990 |
+
|
991 |
+
plt.show()
|
992 |
+
|
993 |
+
|
994 |
+
def plot_repetition_factors(result, groups):
|
995 |
+
for group in groups:
|
996 |
+
# Plot the statistics
|
997 |
+
plt.figure(figsize=(10, 6))
|
998 |
+
|
999 |
+
max_value = 0
|
1000 |
+
for model in result.keys():
|
1001 |
+
if not group in model.lower():
|
1002 |
+
continue
|
1003 |
+
print(f"model: {model}")
|
1004 |
+
df = result[model]["df_overall"]
|
1005 |
+
repetition_panelties = [
|
1006 |
+
repetition_penalty for repetition_penalty in df["repetition_penalty"]
|
1007 |
+
]
|
1008 |
+
|
1009 |
+
mean_score = [
|
1010 |
+
math.log10(10 + df["total_repetitions"].mean())
|
1011 |
+
for df in result[model]["df_list_repetition_penalty"]
|
1012 |
+
]
|
1013 |
+
|
1014 |
+
sns.lineplot(x=repetition_panelties, y=mean_score, label=model)
|
1015 |
+
|
1016 |
+
new_max = max(mean_score)
|
1017 |
+
if new_max > max_value:
|
1018 |
+
max_value = new_max
|
1019 |
+
|
1020 |
+
max_value = max_value * 1.05
|
1021 |
+
if max_value < 1.5:
|
1022 |
+
max_value = 1.5
|
1023 |
+
# set ylimit
|
1024 |
+
plt.ylim(1, max_value)
|
1025 |
+
|
1026 |
+
# show grid
|
1027 |
+
plt.grid(True)
|
1028 |
+
plt.xlabel("Repetition Penalties")
|
1029 |
+
plt.ylabel("Repetition Factors")
|
1030 |
+
plt.title("Repetition Factors vs Repetition Penalties")
|
1031 |
+
plt.legend()
|
1032 |
+
|
1033 |
+
plt.show()
|
1034 |
+
|
1035 |
+
|
1036 |
+
def plot_repetition_factors_by_group(result, group_filter=None):
|
1037 |
+
markers = ["D", "o", "s", "x"]
|
1038 |
+
colors = ["blue", "orange", "green", "red"]
|
1039 |
+
|
1040 |
+
# Plot the statistics
|
1041 |
+
plt.figure(figsize=(10, 6))
|
1042 |
+
index = 0
|
1043 |
+
max_value = 0
|
1044 |
+
|
1045 |
+
for model in result.keys():
|
1046 |
+
if group_filter is not None and group_filter not in model:
|
1047 |
+
continue
|
1048 |
+
|
1049 |
+
print(f"model: {model}")
|
1050 |
+
|
1051 |
+
df = result[model]["df_overall"]
|
1052 |
+
repetition_panelties = [
|
1053 |
+
repetition_penalty for repetition_penalty in df["repetition_penalty"]
|
1054 |
+
]
|
1055 |
+
|
1056 |
+
# Calculate the statistics
|
1057 |
+
mean_score = [
|
1058 |
+
math.log10(10 + df["total_repetitions"].mean())
|
1059 |
+
for df in result[model]["df_list_repetition_penalty"]
|
1060 |
+
]
|
1061 |
+
if len(mean_score) != len(repetition_panelties):
|
1062 |
+
print(
|
1063 |
+
f"model: {model} has different length of repetition penalties and mean score"
|
1064 |
+
)
|
1065 |
+
print("repetition_panelties:", len(repetition_panelties))
|
1066 |
+
print("mean_score:", len(mean_score))
|
1067 |
+
continue
|
1068 |
+
|
1069 |
+
new_max = max(mean_score)
|
1070 |
+
if new_max > max_value:
|
1071 |
+
max_value = new_max
|
1072 |
+
|
1073 |
+
sns.lineplot(
|
1074 |
+
x=repetition_panelties,
|
1075 |
+
y=mean_score,
|
1076 |
+
label=model,
|
1077 |
+
marker=markers[index],
|
1078 |
+
color=colors[index],
|
1079 |
+
)
|
1080 |
+
|
1081 |
+
index += 1
|
1082 |
+
|
1083 |
+
max_value = max_value * 1.05
|
1084 |
+
if max_value < 1.5:
|
1085 |
+
max_value = 1.5
|
1086 |
+
# set ylimit
|
1087 |
+
plt.ylim(1, max_value)
|
1088 |
+
max_value = 0
|
1089 |
+
|
1090 |
+
plt.xlabel("Repetition Penalties")
|
1091 |
+
plt.ylabel("Repetition Factors")
|
1092 |
+
plt.title("Repetition Factors vs Repetition Penalties")
|
1093 |
+
plt.legend(bbox_to_anchor=(1.0, 0.5), loc="center left")
|
1094 |
+
|
1095 |
+
plt.show()
|
eval_modules/calc_repetitions_v4.py
ADDED
@@ -0,0 +1,1296 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
1 |
+
import os
|
2 |
+
import re
|
3 |
+
import math
|
4 |
+
import pandas as pd
|
5 |
+
import numpy as np
|
6 |
+
import matplotlib.pyplot as plt
|
7 |
+
import matplotlib.ticker as mtick
|
8 |
+
import seaborn as sns
|
9 |
+
import nltk
|
10 |
+
|
11 |
+
print(f"loading: {__file__}")
|
12 |
+
|
13 |
+
# final version
|
14 |
+
pattern_excessive_whitespaces = re.compile(r"\s{5,}")
|
15 |
+
pattern_text_repetitions = re.compile(r"(.{5}.*)\s*((\1)\s*)+", re.M | re.DOTALL)
|
16 |
+
|
17 |
+
|
18 |
+
# final version for repetition detection
|
19 |
+
def detect_repetitions(text, debug=False):
|
20 |
+
subtotals = [0, 0]
|
21 |
+
|
22 |
+
if isinstance(text, str):
|
23 |
+
patterns = [pattern_excessive_whitespaces, pattern_text_repetitions]
|
24 |
+
for i, pattern in enumerate(patterns):
|
25 |
+
if debug:
|
26 |
+
print(
|
27 |
+
f"----detect {'excessive whitespaces' if i == 0 else 'text repetitions'}----"
|
28 |
+
)
|
29 |
+
matches = pattern.finditer(text)
|
30 |
+
for match in matches:
|
31 |
+
if debug:
|
32 |
+
print(match)
|
33 |
+
for groupNum in range(0, len(match.groups())):
|
34 |
+
groupNum = groupNum + 1
|
35 |
+
print(
|
36 |
+
"Group {groupNum} found at {start}-{end}: `{group}`".format(
|
37 |
+
groupNum=groupNum,
|
38 |
+
start=match.start(groupNum),
|
39 |
+
end=match.end(groupNum),
|
40 |
+
group=match.group(groupNum),
|
41 |
+
)
|
42 |
+
)
|
43 |
+
|
44 |
+
start, end = match.span()
|
45 |
+
subtotals[i] += end - start
|
46 |
+
|
47 |
+
if i == 0 and subtotals[i] > 0:
|
48 |
+
text = pattern.sub("", text)
|
49 |
+
if debug:
|
50 |
+
print(f"removed excessive whitespaces: {subtotals[i]}")
|
51 |
+
|
52 |
+
result = (subtotals[0], subtotals[1], subtotals[0] + subtotals[1])
|
53 |
+
|
54 |
+
if debug:
|
55 |
+
print(result)
|
56 |
+
return result
|
57 |
+
|
58 |
+
|
59 |
+
def detect_excessive_whitespaces(text, debug=False):
|
60 |
+
return detect_repetitions(text, debug=debug)[0]
|
61 |
+
|
62 |
+
|
63 |
+
def detect_text_repetitions(text, debug=False):
|
64 |
+
return detect_repetitions(text, debug=debug)[1]
|
65 |
+
|
66 |
+
|
67 |
+
def detect_scores(text, debug=False):
|
68 |
+
newline_score, repetition_score, total_repetitions = detect_repetitions(
|
69 |
+
text, debug=debug
|
70 |
+
)
|
71 |
+
return pd.Series([newline_score, repetition_score, total_repetitions])
|
72 |
+
|
73 |
+
|
74 |
+
def load_with_newline_and_repetition_scores(result_file, force_recalculate=False):
|
75 |
+
print(f"loading result file: {result_file}")
|
76 |
+
df = pd.read_csv(result_file, comment="#", on_bad_lines="warn")
|
77 |
+
|
78 |
+
if (
|
79 |
+
force_recalculate
|
80 |
+
or "newline_score" not in df.columns
|
81 |
+
or "repetition_score" not in df.columns
|
82 |
+
or "total_repetitions" not in df.columns
|
83 |
+
):
|
84 |
+
df[["newline_score", "repetition_score", "total_repetitions"]] = df[
|
85 |
+
"answer"
|
86 |
+
].apply(detect_scores)
|
87 |
+
df.to_csv(result_file, index=False)
|
88 |
+
|
89 |
+
return df
|
90 |
+
|
91 |
+
|
92 |
+
def replace_last(source_string, old_string, new_string):
|
93 |
+
head, _sep, tail = source_string.rpartition(old_string)
|
94 |
+
return head + new_string + tail
|
95 |
+
|
96 |
+
|
97 |
+
def load_for_repetition_penalty(
|
98 |
+
csv_result_file, repetition_penalty, force_recalculate=False
|
99 |
+
):
|
100 |
+
result_file = replace_last(
|
101 |
+
csv_result_file, ".csv", f"_RP_{repetition_penalty:.3f}.csv"
|
102 |
+
)
|
103 |
+
return load_with_newline_and_repetition_scores(
|
104 |
+
result_file, force_recalculate=force_recalculate
|
105 |
+
)
|
106 |
+
|
107 |
+
|
108 |
+
def calc_adjusted_performance(f, r):
|
109 |
+
return f / math.log10(10 + r)
|
110 |
+
|
111 |
+
|
112 |
+
def calculate_adjusted_performance(row):
|
113 |
+
r = row["total_repetitions"]
|
114 |
+
adjusted_precision = calc_adjusted_performance(row["precision"], r)
|
115 |
+
adjusted_recall = calc_adjusted_performance(row["recall"], r)
|
116 |
+
return pd.Series([adjusted_precision, adjusted_recall])
|
117 |
+
|
118 |
+
|
119 |
+
def load_performance_df(csv_result_file, repetition_penalty):
|
120 |
+
result_file = replace_last(
|
121 |
+
csv_result_file, ".csv", f"_RP_{repetition_penalty:.3f}-t2_evaluated.json"
|
122 |
+
)
|
123 |
+
result_file = result_file.replace("/results/", "/eval/")
|
124 |
+
print(f"loading json file: {result_file}")
|
125 |
+
df = pd.read_json(result_file)
|
126 |
+
|
127 |
+
return df
|
128 |
+
|
129 |
+
|
130 |
+
def calculate_performance_score_v1(
|
131 |
+
csv_result_file, repetition_penalty, force_recalculate=False
|
132 |
+
):
|
133 |
+
result_file = replace_last(
|
134 |
+
csv_result_file, ".csv", f"_RP_{repetition_penalty:.3f}.csv"
|
135 |
+
)
|
136 |
+
print(f"loading result file: {result_file}")
|
137 |
+
df = pd.read_csv(result_file, comment="#", on_bad_lines="warn")
|
138 |
+
|
139 |
+
if force_recalculate or "f2" in df.columns or "f1" not in df.columns:
|
140 |
+
df.drop(
|
141 |
+
columns=[
|
142 |
+
"precision",
|
143 |
+
"recall",
|
144 |
+
"f1",
|
145 |
+
"f2",
|
146 |
+
"entities_in_answer",
|
147 |
+
"entities_in_question",
|
148 |
+
],
|
149 |
+
errors="ignore",
|
150 |
+
inplace=True,
|
151 |
+
)
|
152 |
+
perf_df = load_performance_df(csv_result_file, repetition_penalty)
|
153 |
+
filtered_df = perf_df[perf_df["id"].isin(df["id"])]
|
154 |
+
perf_df = filtered_df.reset_index(drop=True)
|
155 |
+
print(f"perf_df len: {len(perf_df)}")
|
156 |
+
# print(perf_df.head())
|
157 |
+
|
158 |
+
df["eval_gemini_1.0_pro"] = perf_df["eval_gemini_1.0_pro"]
|
159 |
+
|
160 |
+
df["precision"] = perf_df["score"].apply(lambda x: x[0])
|
161 |
+
df["recall"] = perf_df["score"].apply(lambda x: x[1])
|
162 |
+
df["f1"] = perf_df["score"].apply(lambda x: x[2])
|
163 |
+
|
164 |
+
df[["adjusted_precision", "adjusted_recall"]] = df.apply(
|
165 |
+
calculate_adjusted_performance, axis=1
|
166 |
+
)
|
167 |
+
|
168 |
+
df.to_csv(result_file, index=False)
|
169 |
+
print(f"performance scores saved to result file: {result_file}")
|
170 |
+
|
171 |
+
print(f"df len: {len(df)}")
|
172 |
+
|
173 |
+
return df
|
174 |
+
|
175 |
+
|
176 |
+
ref_df = pd.read_csv(
|
177 |
+
"./data/results/gpt-3.5-turbo_non_rag.csv", comment="#", on_bad_lines="warn"
|
178 |
+
)
|
179 |
+
|
180 |
+
|
181 |
+
def calculate_performance_score(
|
182 |
+
csv_result_file, repetition_penalty, force_recalculate=False
|
183 |
+
):
|
184 |
+
result_file = replace_last(
|
185 |
+
csv_result_file, ".csv", f"_RP_{repetition_penalty:.3f}.csv"
|
186 |
+
)
|
187 |
+
|
188 |
+
re_creating = False
|
189 |
+
if os.path.exists(result_file):
|
190 |
+
print(f"loading result file: {result_file}")
|
191 |
+
df = pd.read_csv(result_file, comment="#", on_bad_lines="warn")
|
192 |
+
else:
|
193 |
+
print(f"re-creating result file: {result_file}")
|
194 |
+
df = pd.DataFrame()
|
195 |
+
force_recalculate = True
|
196 |
+
|
197 |
+
if force_recalculate or "f2" in df.columns or "f1" not in df.columns:
|
198 |
+
df.drop(
|
199 |
+
columns=[
|
200 |
+
"precision",
|
201 |
+
"recall",
|
202 |
+
"f1",
|
203 |
+
"f2",
|
204 |
+
"entities_in_answer",
|
205 |
+
"entities_in_question",
|
206 |
+
"word_count",
|
207 |
+
],
|
208 |
+
errors="ignore",
|
209 |
+
inplace=True,
|
210 |
+
)
|
211 |
+
perf_df = load_performance_df(csv_result_file, repetition_penalty)
|
212 |
+
filtered_df = perf_df[perf_df["id"].isin(ref_df["id"])]
|
213 |
+
perf_df = filtered_df.reset_index(drop=True)
|
214 |
+
print(f"perf_df len: {len(perf_df)}")
|
215 |
+
|
216 |
+
if len(perf_df) != len(ref_df):
|
217 |
+
print(f"error: len(perf_df) != {len(ref_df)}")
|
218 |
+
missing_ids = [
|
219 |
+
id for id in ref_df["id"].unique() if id not in perf_df["id"].unique()
|
220 |
+
]
|
221 |
+
print(f"missing_ids: {missing_ids}")
|
222 |
+
|
223 |
+
# print(perf_df.head())
|
224 |
+
|
225 |
+
df["id"] = perf_df["id"]
|
226 |
+
df["question"] = perf_df["question"]
|
227 |
+
df["answer"] = perf_df["pred_answer"]
|
228 |
+
df["word_count"] = df["answer"].apply(
|
229 |
+
lambda x: len(nltk.word_tokenize(x)) if isinstance(x, str) else 0
|
230 |
+
)
|
231 |
+
df["ground_truth"] = perf_df["ground_truth"]
|
232 |
+
df[["newline_score", "repetition_score", "total_repetitions"]] = df[
|
233 |
+
"answer"
|
234 |
+
].apply(detect_scores)
|
235 |
+
|
236 |
+
df["eval_gemini_1.0_pro"] = perf_df["eval_gemini_1.0_pro"]
|
237 |
+
df["precision"] = perf_df["score"].apply(lambda x: x[0])
|
238 |
+
df["recall"] = perf_df["score"].apply(lambda x: x[1])
|
239 |
+
df["f1"] = perf_df["score"].apply(lambda x: x[2])
|
240 |
+
|
241 |
+
df[["adjusted_precision", "adjusted_recall"]] = df.apply(
|
242 |
+
calculate_adjusted_performance, axis=1
|
243 |
+
)
|
244 |
+
|
245 |
+
df.to_csv(result_file, index=False)
|
246 |
+
print(f"performance scores saved to result file: {result_file}")
|
247 |
+
|
248 |
+
print(f"df len: {len(df)}")
|
249 |
+
|
250 |
+
return df
|
251 |
+
|
252 |
+
|
253 |
+
def adjust_perf_scores_with_repetition_penalty(result, precision, recall):
|
254 |
+
newline_score = [
|
255 |
+
df["newline_score"].mean() for df in result["df_list_repetition_penalty"]
|
256 |
+
]
|
257 |
+
|
258 |
+
repetition_score = [
|
259 |
+
df["repetition_score"].mean() for df in result["df_list_repetition_penalty"]
|
260 |
+
]
|
261 |
+
|
262 |
+
precision = [
|
263 |
+
f / math.log10(10 + n + r)
|
264 |
+
for f, n, r in zip(precision, newline_score, repetition_score)
|
265 |
+
]
|
266 |
+
recall = [
|
267 |
+
f / math.log10(10 + n + r)
|
268 |
+
for f, n, r in zip(recall, newline_score, repetition_score)
|
269 |
+
]
|
270 |
+
|
271 |
+
return precision, recall
|
272 |
+
|
273 |
+
|
274 |
+
def plot_performance_scores(
|
275 |
+
result,
|
276 |
+
models=None,
|
277 |
+
title="Performance",
|
278 |
+
):
|
279 |
+
|
280 |
+
if models is None:
|
281 |
+
models = result.keys()
|
282 |
+
for model in models:
|
283 |
+
print(f"model: {model}")
|
284 |
+
df = result[model]["df_overall"]
|
285 |
+
|
286 |
+
# Calculate the statistics
|
287 |
+
precision = [
|
288 |
+
df["precision"].mean() for df in result[model]["df_list_repetition_penalty"]
|
289 |
+
]
|
290 |
+
recall = [
|
291 |
+
df["recall"].mean() for df in result[model]["df_list_repetition_penalty"]
|
292 |
+
]
|
293 |
+
f1 = [2 * (p * r) / (p + r) for p, r in zip(precision, recall)]
|
294 |
+
best_f1 = max(f1)
|
295 |
+
best_f1_index = f1.index(best_f1)
|
296 |
+
|
297 |
+
precision, recall = adjust_perf_scores_with_repetition_penalty(
|
298 |
+
result[model], precision, recall
|
299 |
+
)
|
300 |
+
afrp = [2 * (p * r) / (p + r) for p, r in zip(precision, recall)]
|
301 |
+
|
302 |
+
# f1 = [df["f1"].mean() for df in result[model]["df_list_repetition_penalty"]]
|
303 |
+
best_afrp = max(afrp)
|
304 |
+
best_afrp_index = afrp.index(best_afrp)
|
305 |
+
|
306 |
+
adjusted_precision = [
|
307 |
+
df["adjusted_precision"].mean()
|
308 |
+
for df in result[model]["df_list_repetition_penalty"]
|
309 |
+
]
|
310 |
+
adjusted_recall = [
|
311 |
+
df["adjusted_recall"].mean()
|
312 |
+
for df in result[model]["df_list_repetition_penalty"]
|
313 |
+
]
|
314 |
+
afrp2 = [
|
315 |
+
2 * (p * r) / (p + r) for p, r in zip(adjusted_precision, adjusted_recall)
|
316 |
+
]
|
317 |
+
best_afrp2 = max(afrp2)
|
318 |
+
best_afrp2_index = afrp2.index(best_afrp2)
|
319 |
+
|
320 |
+
repetition_penalties = list(df["repetition_penalty"])
|
321 |
+
|
322 |
+
# line plot for precision, recall, f1
|
323 |
+
plt.figure(figsize=(10, 6))
|
324 |
+
|
325 |
+
plt.axvspan(
|
326 |
+
repetition_penalties[best_f1_index] - 0.01,
|
327 |
+
repetition_penalties[best_f1_index] + 0.01,
|
328 |
+
alpha=0.5,
|
329 |
+
edgecolor="none",
|
330 |
+
facecolor="blue",
|
331 |
+
)
|
332 |
+
|
333 |
+
# plt.axvspan(
|
334 |
+
# repetition_penalties[best_afrp2_index] - 0.01,
|
335 |
+
# repetition_penalties[best_afrp2_index] + 0.01,
|
336 |
+
# alpha=0.5,
|
337 |
+
# edgecolor="none",
|
338 |
+
# facecolor="green",
|
339 |
+
# )
|
340 |
+
|
341 |
+
plt.axvspan(
|
342 |
+
repetition_penalties[best_afrp_index] - 0.01,
|
343 |
+
repetition_penalties[best_afrp_index] + 0.01,
|
344 |
+
alpha=0.5,
|
345 |
+
edgecolor="none",
|
346 |
+
facecolor="orange",
|
347 |
+
)
|
348 |
+
|
349 |
+
plt.plot(repetition_penalties, f1, label="F1", marker="D", color="blue")
|
350 |
+
# plt.plot(
|
351 |
+
# repetition_penalties,
|
352 |
+
# afrp2,
|
353 |
+
# label="Per-question RF Adjusted F1",
|
354 |
+
# marker="s",
|
355 |
+
# color="green",
|
356 |
+
# )
|
357 |
+
plt.plot(
|
358 |
+
repetition_penalties,
|
359 |
+
afrp,
|
360 |
+
label="RF Adjusted F1",
|
361 |
+
marker="o",
|
362 |
+
color="orange",
|
363 |
+
)
|
364 |
+
plt.xlabel("Repetition Penalties")
|
365 |
+
plt.ylabel("Score")
|
366 |
+
plt.xlim(0.99, 1.31)
|
367 |
+
# y in percentage
|
368 |
+
plt.gca().yaxis.set_major_formatter(mtick.PercentFormatter(1.0))
|
369 |
+
plt.title(f"{model} {title}")
|
370 |
+
plt.legend(bbox_to_anchor=(1.0, 0.5), loc="center left")
|
371 |
+
|
372 |
+
plt.show()
|
373 |
+
|
374 |
+
|
375 |
+
def plot_best_afrp(
|
376 |
+
result,
|
377 |
+
models=None,
|
378 |
+
title="Models with Best Repetition Factor Adjusted F1",
|
379 |
+
ref_result=None,
|
380 |
+
):
|
381 |
+
# Initialize lists to store the statistics
|
382 |
+
model_names = []
|
383 |
+
best_f1 = []
|
384 |
+
best_afrp = []
|
385 |
+
best_repetition_penalty = []
|
386 |
+
best_mtp = []
|
387 |
+
|
388 |
+
if models is None:
|
389 |
+
models = result.keys()
|
390 |
+
for model in models:
|
391 |
+
print(f"model: {model}")
|
392 |
+
df = result[model]["df_overall"]
|
393 |
+
|
394 |
+
# Calculate the statistics
|
395 |
+
precision = [
|
396 |
+
df["precision"].mean() for df in result[model]["df_list_repetition_penalty"]
|
397 |
+
]
|
398 |
+
recall = [
|
399 |
+
df["recall"].mean() for df in result[model]["df_list_repetition_penalty"]
|
400 |
+
]
|
401 |
+
# f1 = [df["f1"].mean() for df in result[model]["df_list_repetition_penalty"]]
|
402 |
+
f1 = [2 * (p * r) / (p + r) for p, r in zip(precision, recall)]
|
403 |
+
|
404 |
+
newline_score = [
|
405 |
+
df["newline_score"].mean()
|
406 |
+
for df in result[model]["df_list_repetition_penalty"]
|
407 |
+
]
|
408 |
+
# print(f"newline_score: {newline_score}")
|
409 |
+
|
410 |
+
repetition_score = [
|
411 |
+
df["repetition_score"].mean()
|
412 |
+
for df in result[model]["df_list_repetition_penalty"]
|
413 |
+
]
|
414 |
+
# print(f"repetition_score: {repetition_score}")
|
415 |
+
|
416 |
+
afrp = [
|
417 |
+
f / math.log10(10 + n + r)
|
418 |
+
for f, n, r in zip(f1, newline_score, repetition_score)
|
419 |
+
]
|
420 |
+
|
421 |
+
best_afrp.append(max(afrp))
|
422 |
+
best_afrp_index = afrp.index(best_afrp[-1])
|
423 |
+
best_repetition_penalty.append(df["repetition_penalty"][best_afrp_index])
|
424 |
+
|
425 |
+
best_f1.append(f1[best_afrp_index])
|
426 |
+
best_mtp.append(
|
427 |
+
newline_score[best_afrp_index] + repetition_score[best_afrp_index]
|
428 |
+
)
|
429 |
+
|
430 |
+
# print(
|
431 |
+
# f"best repetition penalty: {best_repetition_penalty[-1]}, best afrp: {best_afrp[-1]}, f1: {best_f1[-1]}"
|
432 |
+
# )
|
433 |
+
|
434 |
+
df = result[model]["df_list_repetition_penalty"][best_afrp_index]
|
435 |
+
|
436 |
+
model_names.append(
|
437 |
+
f"{model} (RP={best_repetition_penalty[-1]})"
|
438 |
+
) # Add the model name to the list
|
439 |
+
|
440 |
+
if ref_result is not None:
|
441 |
+
print("ref_result:", ref_result)
|
442 |
+
for model in ref_result.keys():
|
443 |
+
model_names.append(model)
|
444 |
+
df = pd.read_csv(ref_result[model])
|
445 |
+
# df = df[df["id"].isin(wikidata_df["id"])]
|
446 |
+
|
447 |
+
p = df["precision"].mean()
|
448 |
+
r = df["recall"].mean()
|
449 |
+
|
450 |
+
f1 = 2 * p * r / (p + r) if p + r > 0 else 0
|
451 |
+
best_f1.append(f1)
|
452 |
+
best_afrp.append(f1)
|
453 |
+
best_mtp.append(0)
|
454 |
+
|
455 |
+
print("model_names:", model_names)
|
456 |
+
# print("best_f1:", best_f1)
|
457 |
+
# print("best_afrp:", best_afrp)
|
458 |
+
|
459 |
+
# Create a DataFrame with the statistics
|
460 |
+
data = pd.DataFrame(
|
461 |
+
{
|
462 |
+
"Model": model_names,
|
463 |
+
"Repetition Factor Adjusted F1": best_afrp,
|
464 |
+
"F1": best_f1,
|
465 |
+
}
|
466 |
+
)
|
467 |
+
|
468 |
+
# Melt the DataFrame to a long format
|
469 |
+
data_melted = data.melt(id_vars="Model", var_name="Metric", value_name="Score")
|
470 |
+
|
471 |
+
# Pivot the DataFrame to a wide format
|
472 |
+
data_pivoted = data_melted.pivot(index="Metric", columns="Model", values="Score")
|
473 |
+
|
474 |
+
# make sure the columns are following the order of the models
|
475 |
+
data_pivoted = data_pivoted[model_names]
|
476 |
+
|
477 |
+
# make sure three groups in the order of precision, recall, f1
|
478 |
+
data_pivoted = data_pivoted.reindex(["Repetition Factor Adjusted F1", "F1"])
|
479 |
+
|
480 |
+
# Plot the statistics
|
481 |
+
plt.figure(figsize=(15, 6))
|
482 |
+
ax = data_pivoted.plot(kind="bar", ax=plt.gca(), width=0.9)
|
483 |
+
plt.title(title)
|
484 |
+
plt.legend(bbox_to_anchor=(1.0, 0.5), loc="center left")
|
485 |
+
|
486 |
+
# Set the rotation of the x-axis labels to 0 degrees
|
487 |
+
plt.xticks(rotation=0)
|
488 |
+
|
489 |
+
# Format the y-axis to display as percentage
|
490 |
+
ax.yaxis.set_major_formatter(mtick.PercentFormatter(1.0))
|
491 |
+
|
492 |
+
# get the max value of the y-axis
|
493 |
+
a1 = max(best_afrp)
|
494 |
+
a2 = max(best_f1)
|
495 |
+
|
496 |
+
max_value = max([a1, a2]) * 1.12
|
497 |
+
print("max_value:", max_value)
|
498 |
+
|
499 |
+
# Set the y-axis limit up to 70%
|
500 |
+
ax.set_ylim(0, max_value)
|
501 |
+
|
502 |
+
# Add the values above each bar
|
503 |
+
for p in ax.patches:
|
504 |
+
ax.annotate(
|
505 |
+
f"{p.get_height() * 100:.1f}",
|
506 |
+
(p.get_x() + p.get_width() / 2.0, p.get_height()),
|
507 |
+
ha="center",
|
508 |
+
va="bottom",
|
509 |
+
xytext=(0, 10),
|
510 |
+
textcoords="offset points",
|
511 |
+
rotation=90,
|
512 |
+
)
|
513 |
+
|
514 |
+
plt.show()
|
515 |
+
return data_pivoted, best_mtp
|
516 |
+
|
517 |
+
|
518 |
+
def plot_best_performance(
|
519 |
+
result,
|
520 |
+
models=None,
|
521 |
+
title="Models with Best F1 Score",
|
522 |
+
adjusted_f1=False,
|
523 |
+
ref_result=None,
|
524 |
+
):
|
525 |
+
# Initialize lists to store the statistics
|
526 |
+
model_names = []
|
527 |
+
best_precision = []
|
528 |
+
best_recall = []
|
529 |
+
best_f1 = []
|
530 |
+
best_repetition_penalty = []
|
531 |
+
best_mtp = []
|
532 |
+
|
533 |
+
if models is None:
|
534 |
+
models = result.keys()
|
535 |
+
for model in models:
|
536 |
+
print(f"model: {model}")
|
537 |
+
df = result[model]["df_overall"]
|
538 |
+
|
539 |
+
# Calculate the statistics
|
540 |
+
precision = [
|
541 |
+
df["precision"].mean() for df in result[model]["df_list_repetition_penalty"]
|
542 |
+
]
|
543 |
+
recall = [
|
544 |
+
df["recall"].mean() for df in result[model]["df_list_repetition_penalty"]
|
545 |
+
]
|
546 |
+
newline_score = [
|
547 |
+
df["newline_score"].mean()
|
548 |
+
for df in result[model]["df_list_repetition_penalty"]
|
549 |
+
]
|
550 |
+
|
551 |
+
repetition_score = [
|
552 |
+
df["repetition_score"].mean()
|
553 |
+
for df in result[model]["df_list_repetition_penalty"]
|
554 |
+
]
|
555 |
+
|
556 |
+
if adjusted_f1:
|
557 |
+
precision, recall = adjust_perf_scores_with_repetition_penalty(
|
558 |
+
result[model], precision, recall
|
559 |
+
)
|
560 |
+
|
561 |
+
# f1 = [df["f1"].mean() for df in result[model]["df_list_repetition_penalty"]]
|
562 |
+
f1 = [2 * (p * r) / (p + r) for p, r in zip(precision, recall)]
|
563 |
+
|
564 |
+
best_f1.append(max(f1))
|
565 |
+
best_f1_index = f1.index(best_f1[-1])
|
566 |
+
best_repetition_penalty.append(df["repetition_penalty"][best_f1_index])
|
567 |
+
|
568 |
+
best_precision.append(precision[best_f1_index])
|
569 |
+
best_recall.append(recall[best_f1_index])
|
570 |
+
best_mtp.append(newline_score[best_f1_index] + repetition_score[best_f1_index])
|
571 |
+
|
572 |
+
print(
|
573 |
+
f"best repetition penalty: {best_repetition_penalty[-1]}, best f1: {best_f1[-1]}, precision: {best_precision[-1]}, recall: {best_recall[-1]}"
|
574 |
+
)
|
575 |
+
|
576 |
+
df = result[model]["df_list_repetition_penalty"][best_f1_index]
|
577 |
+
|
578 |
+
model_names.append(
|
579 |
+
f"{model} (RP={best_repetition_penalty[-1]})"
|
580 |
+
) # Add the model name to the list
|
581 |
+
|
582 |
+
# print sum for columns: newline_score, repetition_score
|
583 |
+
print(
|
584 |
+
f"newline_score: {df['newline_score'].sum()}, repetition_score: {df['repetition_score'].sum()}"
|
585 |
+
)
|
586 |
+
|
587 |
+
if ref_result is not None:
|
588 |
+
print("ref_result:", ref_result)
|
589 |
+
for model in ref_result.keys():
|
590 |
+
model_names.append(model)
|
591 |
+
df = pd.read_csv(ref_result[model])
|
592 |
+
# df = df[df["id"].isin(wikidata_df["id"])]
|
593 |
+
|
594 |
+
best_precision.append(df["precision"].mean())
|
595 |
+
best_recall.append(df["recall"].mean())
|
596 |
+
f1 = (
|
597 |
+
2
|
598 |
+
* (best_precision[-1] * best_recall[-1])
|
599 |
+
/ (best_precision[-1] + best_recall[-1])
|
600 |
+
)
|
601 |
+
# best_f1.append(df["f1"].mean())
|
602 |
+
best_f1.append(f1)
|
603 |
+
best_mtp.append(0)
|
604 |
+
|
605 |
+
# Create a DataFrame with the statistics
|
606 |
+
data = (
|
607 |
+
pd.DataFrame(
|
608 |
+
{
|
609 |
+
"Model": model_names,
|
610 |
+
"Adjusted Precision with RP": best_precision,
|
611 |
+
"Adjusted Recall with RP": best_recall,
|
612 |
+
"Adjusted F1 with RP": best_f1,
|
613 |
+
}
|
614 |
+
)
|
615 |
+
if adjusted_f1
|
616 |
+
else pd.DataFrame(
|
617 |
+
{
|
618 |
+
"Model": model_names,
|
619 |
+
"Precision": best_precision,
|
620 |
+
"Recall": best_recall,
|
621 |
+
"F1": best_f1,
|
622 |
+
}
|
623 |
+
)
|
624 |
+
)
|
625 |
+
columns = list(data.columns)
|
626 |
+
|
627 |
+
# Melt the DataFrame to a long format
|
628 |
+
data_melted = data.melt(id_vars="Model", var_name="Metric", value_name="Score")
|
629 |
+
|
630 |
+
# Pivot the DataFrame to a wide format
|
631 |
+
data_pivoted = data_melted.pivot(index="Metric", columns="Model", values="Score")
|
632 |
+
|
633 |
+
# make sure the columns are following the order of the models
|
634 |
+
data_pivoted = data_pivoted[model_names]
|
635 |
+
|
636 |
+
# make sure three groups in the order of precision, recall, f1
|
637 |
+
data_pivoted = data_pivoted.reindex(columns[1:])
|
638 |
+
|
639 |
+
# Plot the statistics
|
640 |
+
plt.figure(figsize=(10, 6))
|
641 |
+
ax = data_pivoted.plot(kind="bar", ax=plt.gca(), width=0.9)
|
642 |
+
plt.title(title)
|
643 |
+
plt.legend(bbox_to_anchor=(1.0, 0.5), loc="center left")
|
644 |
+
|
645 |
+
# Set the rotation of the x-axis labels to 0 degrees
|
646 |
+
plt.xticks(rotation=0)
|
647 |
+
|
648 |
+
# Format the y-axis to display as percentage
|
649 |
+
ax.yaxis.set_major_formatter(mtick.PercentFormatter(1.0))
|
650 |
+
|
651 |
+
# get the max value of the y-axis
|
652 |
+
a1 = max(best_precision)
|
653 |
+
a2 = max(best_recall)
|
654 |
+
a3 = max(best_f1)
|
655 |
+
|
656 |
+
max_value = max([a1, a2, a3]) * 1.12
|
657 |
+
print("max_value:", max_value)
|
658 |
+
|
659 |
+
# Set the y-axis limit up to 70%
|
660 |
+
ax.set_ylim(0, max_value)
|
661 |
+
|
662 |
+
# Add the values above each bar
|
663 |
+
for p in ax.patches:
|
664 |
+
ax.annotate(
|
665 |
+
f"{p.get_height() * 100:.1f}",
|
666 |
+
(p.get_x() + p.get_width() / 2.0, p.get_height()),
|
667 |
+
ha="center",
|
668 |
+
va="bottom",
|
669 |
+
xytext=(0, 10),
|
670 |
+
textcoords="offset points",
|
671 |
+
rotation=90,
|
672 |
+
)
|
673 |
+
|
674 |
+
plt.show()
|
675 |
+
return data_pivoted, best_mtp
|
676 |
+
|
677 |
+
|
678 |
+
def plot_best_performance_ms_macro(
|
679 |
+
result,
|
680 |
+
models=None,
|
681 |
+
title="Models with Best Repetition Factor Adjusted Performance",
|
682 |
+
ref_result=None,
|
683 |
+
skip_generic_prompt=False,
|
684 |
+
include_adjusted_performance=True,
|
685 |
+
):
|
686 |
+
# Initialize lists to store the statistics
|
687 |
+
model_names = []
|
688 |
+
best_f1 = []
|
689 |
+
best_afrp = []
|
690 |
+
best_repetition_penalty = []
|
691 |
+
best_bleu1 = []
|
692 |
+
best_rougeL = []
|
693 |
+
best_mtp = []
|
694 |
+
|
695 |
+
if models is None:
|
696 |
+
models = result.keys()
|
697 |
+
for model in models:
|
698 |
+
if skip_generic_prompt and "generic prompt" in model:
|
699 |
+
continue
|
700 |
+
print(f"model: {model}")
|
701 |
+
df = result[model]["df_overall"]
|
702 |
+
|
703 |
+
# Calculate the statistics
|
704 |
+
bleu1 = [x for x in df["bleu1"]]
|
705 |
+
rougeL = [x for x in df["rougeL"]]
|
706 |
+
f1 = [2 * (p * r) / (p + r) for p, r in zip(bleu1, rougeL)]
|
707 |
+
|
708 |
+
newline_score = [
|
709 |
+
df["newline_score"].mean()
|
710 |
+
for df in result[model]["df_list_repetition_penalty"]
|
711 |
+
]
|
712 |
+
# print(f"newline_score: {newline_score}")
|
713 |
+
|
714 |
+
repetition_score = [
|
715 |
+
df["repetition_score"].mean()
|
716 |
+
for df in result[model]["df_list_repetition_penalty"]
|
717 |
+
]
|
718 |
+
# print(f"repetition_score: {repetition_score}")
|
719 |
+
|
720 |
+
afrp = [
|
721 |
+
f / math.log10(10 + n + r)
|
722 |
+
for f, n, r in zip(f1, newline_score, repetition_score)
|
723 |
+
]
|
724 |
+
|
725 |
+
best_afrp.append(max(afrp if include_adjusted_performance else f1))
|
726 |
+
best_afrp_index = (
|
727 |
+
afrp.index(best_afrp[-1])
|
728 |
+
if include_adjusted_performance
|
729 |
+
else f1.index(best_afrp[-1])
|
730 |
+
)
|
731 |
+
best_repetition_penalty.append(df["repetition_penalty"][best_afrp_index])
|
732 |
+
|
733 |
+
best_f1.append(f1[best_afrp_index])
|
734 |
+
best_bleu1.append(bleu1[best_afrp_index])
|
735 |
+
best_rougeL.append(rougeL[best_afrp_index])
|
736 |
+
best_mtp.append(
|
737 |
+
newline_score[best_afrp_index] + repetition_score[best_afrp_index]
|
738 |
+
)
|
739 |
+
|
740 |
+
# print(
|
741 |
+
# f"best repetition penalty: {best_repetition_penalty[-1]}, best afrp: {best_afrp[-1]}, f1: {best_f1[-1]}"
|
742 |
+
# )
|
743 |
+
|
744 |
+
df = result[model]["df_list_repetition_penalty"][best_afrp_index]
|
745 |
+
|
746 |
+
model_names.append(
|
747 |
+
f"{model} (RP={best_repetition_penalty[-1]})"
|
748 |
+
) # Add the model name to the list
|
749 |
+
|
750 |
+
if ref_result is not None:
|
751 |
+
print("ref_result:", ref_result)
|
752 |
+
for model in ref_result.keys():
|
753 |
+
model_names.append(model)
|
754 |
+
df = pd.read_csv(ref_result[model], comment="#", on_bad_lines="warn")
|
755 |
+
# df = df[df["id"].isin(wikidata_df["id"])]
|
756 |
+
|
757 |
+
p = df["bleu1"][0]
|
758 |
+
best_bleu1.append(p)
|
759 |
+
|
760 |
+
r = df["rougeL"][0]
|
761 |
+
best_rougeL.append(r)
|
762 |
+
|
763 |
+
f1 = 2 * p * r / (p + r) if p + r > 0 else 0
|
764 |
+
best_f1.append(f1)
|
765 |
+
best_afrp.append(f1)
|
766 |
+
best_mtp.append(0)
|
767 |
+
|
768 |
+
# print("model_names:", model_names)
|
769 |
+
# print("best_f1:", best_f1)
|
770 |
+
# print("best_afrp:", best_afrp)
|
771 |
+
|
772 |
+
# Create a DataFrame with the statistics
|
773 |
+
data = (
|
774 |
+
pd.DataFrame(
|
775 |
+
{
|
776 |
+
"Model": model_names,
|
777 |
+
"Repetition Factor Adjusted Perf Score": best_afrp,
|
778 |
+
"Overall Perf Score": best_f1,
|
779 |
+
}
|
780 |
+
)
|
781 |
+
if include_adjusted_performance
|
782 |
+
else pd.DataFrame(
|
783 |
+
{
|
784 |
+
"Model": model_names,
|
785 |
+
"Bleu-1": best_bleu1,
|
786 |
+
"Rouge-L": best_rougeL,
|
787 |
+
"Overall Perf Score": best_f1,
|
788 |
+
}
|
789 |
+
)
|
790 |
+
)
|
791 |
+
|
792 |
+
# Melt the DataFrame to a long format
|
793 |
+
data_melted = data.melt(id_vars="Model", var_name="Metric", value_name="Score")
|
794 |
+
|
795 |
+
# Pivot the DataFrame to a wide format
|
796 |
+
data_pivoted = data_melted.pivot(index="Metric", columns="Model", values="Score")
|
797 |
+
|
798 |
+
# make sure the columns are following the order of the models
|
799 |
+
data_pivoted = data_pivoted[model_names]
|
800 |
+
|
801 |
+
columns = list(data.columns)
|
802 |
+
data_pivoted = data_pivoted.reindex(columns[1:])
|
803 |
+
|
804 |
+
# Plot the statistics
|
805 |
+
plt.figure(figsize=(10, 6))
|
806 |
+
ax = data_pivoted.plot(kind="bar", ax=plt.gca(), width=0.9)
|
807 |
+
plt.title(title)
|
808 |
+
plt.legend(bbox_to_anchor=(1.0, 0.5), loc="center left")
|
809 |
+
|
810 |
+
# Set the rotation of the x-axis labels to 0 degrees
|
811 |
+
plt.xticks(rotation=0)
|
812 |
+
|
813 |
+
# Format the y-axis to display as percentage
|
814 |
+
ax.yaxis.set_major_formatter(mtick.PercentFormatter(1.0))
|
815 |
+
|
816 |
+
# get the max value of the y-axis
|
817 |
+
a1 = max(best_afrp)
|
818 |
+
a2 = max(best_f1)
|
819 |
+
a3 = max(best_bleu1)
|
820 |
+
a4 = max(best_rougeL)
|
821 |
+
|
822 |
+
max_value = (
|
823 |
+
max([a1, a2] if include_adjusted_performance else [a1, a2, a3, a4]) * 1.12
|
824 |
+
)
|
825 |
+
print("max_value:", max_value)
|
826 |
+
|
827 |
+
# Set the y-axis limit up to 70%
|
828 |
+
ax.set_ylim(0, max_value)
|
829 |
+
|
830 |
+
# Add the values above each bar
|
831 |
+
for p in ax.patches:
|
832 |
+
ax.annotate(
|
833 |
+
f"{p.get_height() * 100:.1f}",
|
834 |
+
(p.get_x() + p.get_width() / 2.0, p.get_height()),
|
835 |
+
ha="center",
|
836 |
+
va="bottom",
|
837 |
+
xytext=(0, 10),
|
838 |
+
textcoords="offset points",
|
839 |
+
rotation=90,
|
840 |
+
)
|
841 |
+
|
842 |
+
plt.show()
|
843 |
+
return data_pivoted, best_mtp
|
844 |
+
|
845 |
+
|
846 |
+
all_open_source_models = [
|
847 |
+
"gemma-1.1-2b-it",
|
848 |
+
"Phi-3-mini-128k-instruct",
|
849 |
+
"gemma-1.1-7b-it",
|
850 |
+
"Llama-2-7b-chat-hf",
|
851 |
+
"Mistral-7B-Instruct-v0.2",
|
852 |
+
"Meta-Llama-3-8B-Instruct",
|
853 |
+
"Llama-2-13b-chat-hf",
|
854 |
+
"Llama-2-70b-chat-hf",
|
855 |
+
"Meta-Llama-3-70B-Instruct",
|
856 |
+
]
|
857 |
+
|
858 |
+
|
859 |
+
non_rag_csv_result_files = [
|
860 |
+
"./data/results/gemma-1.1-2b-it_wd_non_rag.csv", # gemma-1.1-2b-it
|
861 |
+
"./data/results/Phi-3-mini-128k-instruct_wd_non_rag_batch_16.csv", # Phi-3-mini-128k-instruct(batch size:16)
|
862 |
+
"./data/results/gemma-1.1-7b-it_wd_non_rag.csv", # gemma-1.1-7b-it
|
863 |
+
"./data/results/Tune_2024-04-09_09-19-22.csv", # Llama-2-7b-chat-hf
|
864 |
+
"./data/results/Tune_2024-04-16_12-24-27.csv.csv", # Mistral-7B-Instruct-v0.2
|
865 |
+
"./data/results/Meta-Llama-3-8B-Instruct_wd_non_rag.csv", # Meta-Llama-3-8B-Instruct
|
866 |
+
"./data/results/Meta-Llama-3-8B-Instruct_wd_1_non_rag.csv", # Meta-Llama-3-8B-Instruct
|
867 |
+
"./data/results/Tune_2024-04-10_16-53-38.csv", # Llama-2-13b-chat-hf
|
868 |
+
"./data/results/Llama-2-70b-chat-hf_wd_non_rag.csv", # Llama-2-70b-chat-hf
|
869 |
+
"./data/results/Meta-Llama-3-70B-Instruct_wd_non_rag.csv", # Meta-Llama-3-70B-Instruct
|
870 |
+
]
|
871 |
+
|
872 |
+
rag_csv_result_files = [
|
873 |
+
"./data/results/gemma-1.1-2b-it_wd.csv", # gemma-1.1-2b-it
|
874 |
+
"./data/results/gemma-1.1-2b-it_wd_true.csv", # gemma-1.1-2b-it(true)
|
875 |
+
"./data/results/Phi-3-mini-128k-instruct_wd_rag_batch_4.csv", # Phi-3-mini-128k-instruct(batch size:16)
|
876 |
+
"./data/results/Phi-3-mini-128k-instruct_wd_true.csv", # Phi-3-mini-128k-instruct(batch size:16)
|
877 |
+
"./data/results/gemma-1.1-7b-it_wd.csv", # gemma-1.1-7b-it
|
878 |
+
"./data/results/gemma-1.1-7b-it_wd_true.csv", # gemma-1.1-7b-it(true)
|
879 |
+
"./data/results/Tune_2024-03-20_15-35-37.csv", # Llama-2-7b-chat-hf
|
880 |
+
"./data/results/Llama-2-7b-chat-hf_wd_true.csv", # Llama-2-7b-chat-hf(true)
|
881 |
+
"./data/results/Tune_2024-03-29_11-28-20.csv", # Mistral-7B-Instruct-v0.2
|
882 |
+
"./data/results/Mistral-7B-Instruct-v0.2_wd_true.csv", # Mistral-7B-Instruct-v0.2(true)
|
883 |
+
"./data/results/Meta-Llama-3-8B-Instruct_wd.csv", # Meta-Llama-3-8b-instruct
|
884 |
+
"./data/results/Meta-Llama-3-8B-Instruct_wd_true.csv", # Meta-Llama-3-8b-instruct(true)
|
885 |
+
"./data/results/Tune_2024-03-25_23-32-57.csv", # Llama-2-13b-chat-hf
|
886 |
+
"./data/results/Llama-2-13b-chat-hf_wd_true.csv", # Llama-2-13b-chat-hf(true)
|
887 |
+
"./data/results/Llama-2-70b-chat-hf_wd.csv", # Llama-2-70b-chat-hf
|
888 |
+
"./data/results/Llama-2-70b-chat-hf_wd_true.csv", # Llama-2-70b-chat-hf
|
889 |
+
"./data/results/Meta-Llama-3-70B-Instruct_wd.csv", # Meta-Llama-3-70B-Instruct
|
890 |
+
"./data/results/Meta-Llama-3-70B-Instruct_wd_true.csv", # Meta-Llama-3-70B-Instruct(true)
|
891 |
+
]
|
892 |
+
|
893 |
+
df_ms_macro = pd.read_json("./data/datasets/ms_macro.json")
|
894 |
+
|
895 |
+
|
896 |
+
def load_for_repetition_penalty_ms_macro(
|
897 |
+
csv_result_file, repetition_penalty, force_recalculate=False
|
898 |
+
):
|
899 |
+
result_file = replace_last(
|
900 |
+
csv_result_file, ".csv", f"_RP_{repetition_penalty:.3f}.csv"
|
901 |
+
)
|
902 |
+
df = load_with_newline_and_repetition_scores(
|
903 |
+
result_file, force_recalculate=force_recalculate
|
904 |
+
)
|
905 |
+
|
906 |
+
if len(df) != len(df_ms_macro):
|
907 |
+
print(f"error: len(df) != {len(df_ms_macro)}")
|
908 |
+
missing_ids = [
|
909 |
+
id for id in df_ms_macro["id"].unique() if id not in df["id"].unique()
|
910 |
+
]
|
911 |
+
print(f"missing_ids: {missing_ids}")
|
912 |
+
|
913 |
+
if df["ground_truth"][0] != str(df_ms_macro["wellFormedAnswers"][0]):
|
914 |
+
df["ground_truth"] = df_ms_macro["wellFormedAnswers"]
|
915 |
+
print("ground_truth updated for:", result_file)
|
916 |
+
df.to_csv(result_file, index=False)
|
917 |
+
return df
|
918 |
+
|
919 |
+
|
920 |
+
# MS MACRO
|
921 |
+
def plot_performance_scores_ms_macro(
|
922 |
+
result,
|
923 |
+
models=None,
|
924 |
+
title="Performance",
|
925 |
+
):
|
926 |
+
|
927 |
+
if models is None:
|
928 |
+
models = result.keys()
|
929 |
+
for model in models:
|
930 |
+
print(f"model: {model}")
|
931 |
+
df = result[model]["df_overall"]
|
932 |
+
# print(result[model]["df_list_repetition_penalty"][0].describe())
|
933 |
+
|
934 |
+
# Calculate the statistics
|
935 |
+
bleu1 = list(df["bleu1"])
|
936 |
+
rougeL = list(df["rougeL"])
|
937 |
+
f1 = [2 * (p * r) / (p + r) for p, r in zip(bleu1, rougeL)]
|
938 |
+
best_f1 = max(f1)
|
939 |
+
best_f1_index = f1.index(best_f1)
|
940 |
+
|
941 |
+
bleu1, rougeL = adjust_perf_scores_with_repetition_penalty(
|
942 |
+
result[model], bleu1, rougeL
|
943 |
+
)
|
944 |
+
afrp = [2 * (p * r) / (p + r) for p, r in zip(bleu1, rougeL)]
|
945 |
+
|
946 |
+
# f1 = [df["f1"].mean() for df in result[model]["df_list_repetition_penalty"]]
|
947 |
+
best_afrp = max(afrp)
|
948 |
+
best_afrp_index = afrp.index(best_afrp)
|
949 |
+
|
950 |
+
repetition_penalties = list(df["repetition_penalty"])
|
951 |
+
|
952 |
+
# line plot for precision, recall, f1
|
953 |
+
plt.figure(figsize=(10, 6))
|
954 |
+
|
955 |
+
plt.axvspan(
|
956 |
+
repetition_penalties[best_f1_index] - 0.01,
|
957 |
+
repetition_penalties[best_f1_index] + 0.01,
|
958 |
+
alpha=0.5,
|
959 |
+
edgecolor="none",
|
960 |
+
facecolor="blue",
|
961 |
+
)
|
962 |
+
|
963 |
+
plt.axvspan(
|
964 |
+
repetition_penalties[best_afrp_index] - 0.01,
|
965 |
+
repetition_penalties[best_afrp_index] + 0.01,
|
966 |
+
alpha=0.5,
|
967 |
+
edgecolor="none",
|
968 |
+
facecolor="orange",
|
969 |
+
)
|
970 |
+
|
971 |
+
plt.plot(
|
972 |
+
repetition_penalties,
|
973 |
+
f1,
|
974 |
+
label="Overall Perf Score",
|
975 |
+
marker="D",
|
976 |
+
color="blue",
|
977 |
+
)
|
978 |
+
plt.plot(
|
979 |
+
repetition_penalties,
|
980 |
+
afrp,
|
981 |
+
label="RF Adjusted Perf Score",
|
982 |
+
marker="o",
|
983 |
+
color="orange",
|
984 |
+
)
|
985 |
+
|
986 |
+
plt.xlabel("Repetition Penalties")
|
987 |
+
plt.ylabel("Score")
|
988 |
+
plt.xlim(0.99, 1.31)
|
989 |
+
# y in percentage
|
990 |
+
plt.gca().yaxis.set_major_formatter(mtick.PercentFormatter(1.0))
|
991 |
+
plt.title(f"{model} {title}")
|
992 |
+
plt.legend(bbox_to_anchor=(1.0, 0.5), loc="center left")
|
993 |
+
|
994 |
+
plt.show()
|
995 |
+
|
996 |
+
|
997 |
+
def plot_repetition_factors(result, groups):
|
998 |
+
for group in groups:
|
999 |
+
# Plot the statistics
|
1000 |
+
plt.figure(figsize=(10, 6))
|
1001 |
+
|
1002 |
+
max_value = 0
|
1003 |
+
for model in result.keys():
|
1004 |
+
if not group in model.lower():
|
1005 |
+
continue
|
1006 |
+
print(f"model: {model}")
|
1007 |
+
df = result[model]["df_overall"]
|
1008 |
+
repetition_panelties = [
|
1009 |
+
repetition_penalty for repetition_penalty in df["repetition_penalty"]
|
1010 |
+
]
|
1011 |
+
|
1012 |
+
mean_score = [
|
1013 |
+
math.log10(10 + df["total_repetitions"].mean())
|
1014 |
+
for df in result[model]["df_list_repetition_penalty"]
|
1015 |
+
]
|
1016 |
+
|
1017 |
+
sns.lineplot(x=repetition_panelties, y=mean_score, label=model)
|
1018 |
+
|
1019 |
+
new_max = max(mean_score)
|
1020 |
+
if new_max > max_value:
|
1021 |
+
max_value = new_max
|
1022 |
+
|
1023 |
+
max_value = max_value * 1.05
|
1024 |
+
if max_value < 1.5:
|
1025 |
+
max_value = 1.5
|
1026 |
+
# set ylimit
|
1027 |
+
plt.ylim(1, max_value)
|
1028 |
+
|
1029 |
+
# show grid
|
1030 |
+
plt.grid(True)
|
1031 |
+
plt.xlabel("Repetition Penalties")
|
1032 |
+
plt.ylabel("Repetition Factors")
|
1033 |
+
plt.title("Repetition Factors vs Repetition Penalties")
|
1034 |
+
plt.legend()
|
1035 |
+
|
1036 |
+
plt.show()
|
1037 |
+
|
1038 |
+
|
1039 |
+
def plot_repetition_factors_by_group(result, group_filter=None):
|
1040 |
+
markers = ["D", "o", "s", "x"]
|
1041 |
+
colors = ["blue", "orange", "green", "red"]
|
1042 |
+
|
1043 |
+
# Plot the statistics
|
1044 |
+
plt.figure(figsize=(10, 6))
|
1045 |
+
index = 0
|
1046 |
+
max_value = 0
|
1047 |
+
|
1048 |
+
for model in result.keys():
|
1049 |
+
if group_filter is not None and group_filter not in model:
|
1050 |
+
continue
|
1051 |
+
|
1052 |
+
print(f"model: {model}")
|
1053 |
+
|
1054 |
+
df = result[model]["df_overall"]
|
1055 |
+
repetition_panelties = [
|
1056 |
+
repetition_penalty for repetition_penalty in df["repetition_penalty"]
|
1057 |
+
]
|
1058 |
+
|
1059 |
+
# Calculate the statistics
|
1060 |
+
mean_score = [
|
1061 |
+
math.log10(10 + df["total_repetitions"].mean())
|
1062 |
+
for df in result[model]["df_list_repetition_penalty"]
|
1063 |
+
]
|
1064 |
+
if len(mean_score) != len(repetition_panelties):
|
1065 |
+
print(
|
1066 |
+
f"model: {model} has different length of repetition penalties and mean score"
|
1067 |
+
)
|
1068 |
+
print("repetition_panelties:", len(repetition_panelties))
|
1069 |
+
print("mean_score:", len(mean_score))
|
1070 |
+
continue
|
1071 |
+
|
1072 |
+
new_max = max(mean_score)
|
1073 |
+
if new_max > max_value:
|
1074 |
+
max_value = new_max
|
1075 |
+
|
1076 |
+
sns.lineplot(
|
1077 |
+
x=repetition_panelties,
|
1078 |
+
y=mean_score,
|
1079 |
+
label=model,
|
1080 |
+
marker=markers[index],
|
1081 |
+
color=colors[index],
|
1082 |
+
)
|
1083 |
+
|
1084 |
+
index += 1
|
1085 |
+
|
1086 |
+
max_value = max_value * 1.05
|
1087 |
+
if max_value < 1.5:
|
1088 |
+
max_value = 1.5
|
1089 |
+
# set ylimit
|
1090 |
+
plt.ylim(1, max_value)
|
1091 |
+
max_value = 0
|
1092 |
+
|
1093 |
+
plt.xlabel("Repetition Penalties")
|
1094 |
+
plt.ylabel("Repetition Factors")
|
1095 |
+
plt.title("Repetition Factors vs Repetition Penalties")
|
1096 |
+
plt.legend(bbox_to_anchor=(1.0, 0.5), loc="center left")
|
1097 |
+
|
1098 |
+
plt.show()
|
1099 |
+
|
1100 |
+
|
1101 |
+
ms_marco_csv_result_files = [
|
1102 |
+
"data/results/gemma-1.1-2b-it_mm_true_false.csv",
|
1103 |
+
"data/results/gemma-1.1-2b-it_mm_true.csv",
|
1104 |
+
"data/results/gemma-1.1-2b-it_mm_true_false_non_rag.csv",
|
1105 |
+
"data/results/Phi-3-mini-128k-instruct_mm_false.csv",
|
1106 |
+
"data/results/Phi-3-mini-128k-instruct_mm_true.csv",
|
1107 |
+
"data/results/Phi-3-mini-128k-instruct_mm_non_rag.csv",
|
1108 |
+
"data/results/gemma-1.1-7b-it_mm_false.csv",
|
1109 |
+
"data/results/gemma-1.1-7b-it_mm_true.csv",
|
1110 |
+
"data/results/gemma-1.1-7b-it_mm_non_rag.csv",
|
1111 |
+
"data/results/Llama-2-7b-chat-hf_mm_true_false.csv",
|
1112 |
+
"data/results/Llama-2-7b-chat-hf_mm_true.csv",
|
1113 |
+
"data/results/Llama-2-7b-chat-hf_mm_true_false_non_rag.csv",
|
1114 |
+
"data/results/Mistral-7B-Instruct-v0.2_mm_false.csv",
|
1115 |
+
"data/results/Mistral-7B-Instruct-v0.2_mm_true.csv",
|
1116 |
+
"data/results/Mistral-7B-Instruct-v0.2_mm_non_rag.csv",
|
1117 |
+
"data/results/Meta-Llama-3-8B-Instruct_mm_true_false.csv",
|
1118 |
+
"data/results/Meta-Llama-3-8B-Instruct_mm_true.csv",
|
1119 |
+
"data/results/Meta-Llama-3-8B-Instruct_mm_true_false_non_rag.csv",
|
1120 |
+
"data/results/Llama-2-13b-chat-hf_mm_false.csv",
|
1121 |
+
"data/results/Llama-2-13b-chat-hf_mm_true.csv",
|
1122 |
+
"data/results/Llama-2-13b-chat-hf_mm_non_rag.csv",
|
1123 |
+
"data/results/Llama-2-70b-chat-hf_mm_false.csv",
|
1124 |
+
"data/results/Llama-2-70b-chat-hf_mm_true.csv",
|
1125 |
+
"data/results/Llama-2-70b-chat-hf_mm_non_rag.csv",
|
1126 |
+
"data/results/Meta-Llama-3-70B-Instruct_mm_false.csv",
|
1127 |
+
"data/results/Meta-Llama-3-70B-Instruct_mm_true.csv",
|
1128 |
+
"data/results/Meta-Llama-3-70B-Instruct_mm_non_rag.csv",
|
1129 |
+
]
|
1130 |
+
|
1131 |
+
webqsp_csv_result_files = []
|
1132 |
+
webqsp_model_result_counts = {}
|
1133 |
+
|
1134 |
+
|
1135 |
+
def find_model_name(file_path):
|
1136 |
+
df = pd.read_csv(file_path, comment="#", on_bad_lines="warn")
|
1137 |
+
return df["model"][0]
|
1138 |
+
|
1139 |
+
|
1140 |
+
def add_file(file):
|
1141 |
+
model_name = find_model_name(file)
|
1142 |
+
if "(generic prompt)" not in model_name:
|
1143 |
+
webqsp_csv_result_files.append(file)
|
1144 |
+
if model_name not in webqsp_model_result_counts:
|
1145 |
+
webqsp_model_result_counts[model_name] = 1
|
1146 |
+
else:
|
1147 |
+
webqsp_model_result_counts[model_name] += 1
|
1148 |
+
|
1149 |
+
|
1150 |
+
last_model_name = None
|
1151 |
+
non_rag_index = 0
|
1152 |
+
|
1153 |
+
for csv_result_file in rag_csv_result_files:
|
1154 |
+
try:
|
1155 |
+
model_name = find_model_name(csv_result_file)
|
1156 |
+
# print(f"processing model: {model_name} - {csv_result_file}")
|
1157 |
+
|
1158 |
+
if last_model_name != model_name and last_model_name is not None:
|
1159 |
+
while non_rag_index < len(non_rag_csv_result_files):
|
1160 |
+
# print(f"processing non-rag file - {file}")
|
1161 |
+
file = non_rag_csv_result_files[non_rag_index]
|
1162 |
+
non_model_name = find_model_name(file)
|
1163 |
+
if non_model_name.startswith(last_model_name):
|
1164 |
+
add_file(file)
|
1165 |
+
non_rag_index += 1
|
1166 |
+
else:
|
1167 |
+
break
|
1168 |
+
|
1169 |
+
add_file(csv_result_file)
|
1170 |
+
last_model_name = model_name
|
1171 |
+
except FileNotFoundError as e:
|
1172 |
+
print("\terror processing file: ", csv_result_file, e)
|
1173 |
+
continue
|
1174 |
+
|
1175 |
+
for file in non_rag_csv_result_files[non_rag_index:]:
|
1176 |
+
add_file(file)
|
1177 |
+
|
1178 |
+
|
1179 |
+
def calc_rap_scores(result, precision="precision", recall="recall"):
|
1180 |
+
newline_score = [
|
1181 |
+
df["newline_score"].mean() for df in result["df_list_repetition_penalty"]
|
1182 |
+
]
|
1183 |
+
|
1184 |
+
repetition_score = [
|
1185 |
+
df["repetition_score"].mean() for df in result["df_list_repetition_penalty"]
|
1186 |
+
]
|
1187 |
+
|
1188 |
+
if precision in result["df_list_repetition_penalty"][0].columns:
|
1189 |
+
precision = [
|
1190 |
+
df[precision].mean() for df in result["df_list_repetition_penalty"]
|
1191 |
+
]
|
1192 |
+
recall = [df[recall].mean() for df in result["df_list_repetition_penalty"]]
|
1193 |
+
else:
|
1194 |
+
precision = result["df_overall"][precision]
|
1195 |
+
recall = result["df_overall"][recall]
|
1196 |
+
|
1197 |
+
f1 = [2 * (p * r) / (p + r) for p, r in zip(precision, recall)]
|
1198 |
+
|
1199 |
+
rap = [
|
1200 |
+
f / math.log10(10 + n + r)
|
1201 |
+
for f, n, r in zip(f1, newline_score, repetition_score)
|
1202 |
+
]
|
1203 |
+
|
1204 |
+
return newline_score, repetition_score, f1, rap
|
1205 |
+
|
1206 |
+
|
1207 |
+
def load_webqsp_result(csv_result_files, force_recalculate=False):
|
1208 |
+
model_name_exts = {
|
1209 |
+
"true": "(RAG - Chat Template)",
|
1210 |
+
"wd": "(RAG - Generic Prompt)",
|
1211 |
+
"rag": "(Non-RAG)",
|
1212 |
+
}
|
1213 |
+
|
1214 |
+
result = {}
|
1215 |
+
for i, csv_result_file in enumerate(csv_result_files):
|
1216 |
+
try:
|
1217 |
+
df = pd.read_csv(csv_result_file)
|
1218 |
+
parts = re.split(r"[_\.]", csv_result_file)
|
1219 |
+
if parts[-2] in model_name_exts.keys():
|
1220 |
+
key = parts[-2]
|
1221 |
+
elif csv_result_file in non_rag_csv_result_files:
|
1222 |
+
key = "rag"
|
1223 |
+
else:
|
1224 |
+
key = "wd"
|
1225 |
+
model_name = f'{df["model"][0]}{model_name_exts[key]}'
|
1226 |
+
dfs = [
|
1227 |
+
calculate_performance_score(
|
1228 |
+
csv_result_file,
|
1229 |
+
repetition_penalty,
|
1230 |
+
force_recalculate=force_recalculate,
|
1231 |
+
)
|
1232 |
+
for repetition_penalty in df["repetition_penalty"]
|
1233 |
+
]
|
1234 |
+
|
1235 |
+
result[model_name] = {
|
1236 |
+
"df_overall": df,
|
1237 |
+
"df_list_repetition_penalty": dfs,
|
1238 |
+
"file": csv_result_file,
|
1239 |
+
}
|
1240 |
+
newline_score, repetition_score, perf, rap = calc_rap_scores(
|
1241 |
+
result[model_name]
|
1242 |
+
)
|
1243 |
+
df["newline_score"] = newline_score
|
1244 |
+
df["repetition_score"] = repetition_score
|
1245 |
+
df["total_repetitions"] = df["newline_score"] + df["repetition_score"]
|
1246 |
+
df["perf"] = perf
|
1247 |
+
df["rap"] = rap
|
1248 |
+
except Exception as e:
|
1249 |
+
print(f"Error: {e}")
|
1250 |
+
|
1251 |
+
return result
|
1252 |
+
|
1253 |
+
|
1254 |
+
def load_ms_marco_result(csv_result_files, force_recalculate=False):
|
1255 |
+
model_name_exts = {
|
1256 |
+
"true": "(RAG - Chat Template)",
|
1257 |
+
"false": "(RAG - Generic Prompt)",
|
1258 |
+
"rag": "(Non-RAG)",
|
1259 |
+
}
|
1260 |
+
|
1261 |
+
result = {}
|
1262 |
+
for csv_result_file in csv_result_files:
|
1263 |
+
try:
|
1264 |
+
df = pd.read_csv(csv_result_file)
|
1265 |
+
|
1266 |
+
parts = re.split(r"[_\.]", csv_result_file)
|
1267 |
+
model_name = f'{df["model"][0]}{model_name_exts[parts[-2]]}'
|
1268 |
+
|
1269 |
+
print(f"\tmodel_name: {model_name}")
|
1270 |
+
dfs = [
|
1271 |
+
load_for_repetition_penalty_ms_macro(
|
1272 |
+
csv_result_file,
|
1273 |
+
repetition_penalty,
|
1274 |
+
force_recalculate=force_recalculate,
|
1275 |
+
)
|
1276 |
+
for repetition_penalty in df["repetition_penalty"]
|
1277 |
+
]
|
1278 |
+
result[model_name] = {
|
1279 |
+
"df_overall": df,
|
1280 |
+
"df_list_repetition_penalty": dfs,
|
1281 |
+
"file": csv_result_file,
|
1282 |
+
}
|
1283 |
+
newline_score, repetition_score, perf, rap = calc_rap_scores(
|
1284 |
+
result[model_name],
|
1285 |
+
precision="bleu1",
|
1286 |
+
recall="rougeL",
|
1287 |
+
)
|
1288 |
+
df["newline_score"] = newline_score
|
1289 |
+
df["repetition_score"] = repetition_score
|
1290 |
+
df["total_repetitions"] = df["newline_score"] + df["repetition_score"]
|
1291 |
+
df["perf"] = perf
|
1292 |
+
df["rap"] = rap
|
1293 |
+
except Exception as e:
|
1294 |
+
print(f"Error: {e}")
|
1295 |
+
|
1296 |
+
return result
|
eval_modules/calc_repetitions_v5.py
ADDED
@@ -0,0 +1,1383 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
1 |
+
import os
|
2 |
+
import re
|
3 |
+
import math
|
4 |
+
import pandas as pd
|
5 |
+
import numpy as np
|
6 |
+
import matplotlib.pyplot as plt
|
7 |
+
import matplotlib.ticker as mtick
|
8 |
+
import seaborn as sns
|
9 |
+
import nltk
|
10 |
+
import evaluate
|
11 |
+
|
12 |
+
meteor = evaluate.load("meteor")
|
13 |
+
|
14 |
+
print(f"loading: {__file__}")
|
15 |
+
|
16 |
+
# final version
|
17 |
+
pattern_excessive_whitespaces = re.compile(r"\s{5,}")
|
18 |
+
pattern_text_repetitions = re.compile(r"(.{5}.*)\s*((\1)\s*)+", re.M | re.DOTALL)
|
19 |
+
|
20 |
+
|
21 |
+
def del_excessive_whitespaces(text, debug=False):
|
22 |
+
count = 0
|
23 |
+
|
24 |
+
if isinstance(text, str):
|
25 |
+
if debug:
|
26 |
+
print("----detect excessive whitespaces----")
|
27 |
+
count = len(text)
|
28 |
+
text = pattern_excessive_whitespaces.sub("", text)
|
29 |
+
count -= len(text)
|
30 |
+
if debug and count:
|
31 |
+
print(f"removed excessive whitespaces: {count}")
|
32 |
+
return text, count
|
33 |
+
|
34 |
+
|
35 |
+
# final version for repetition detection
|
36 |
+
def detect_text_repetitions(text, debug=False):
|
37 |
+
count = 0
|
38 |
+
|
39 |
+
if isinstance(text, str):
|
40 |
+
if debug:
|
41 |
+
print("----detect text repetitions----")
|
42 |
+
matches = pattern_text_repetitions.finditer(text)
|
43 |
+
for match in matches:
|
44 |
+
if debug:
|
45 |
+
print(match)
|
46 |
+
for groupNum in range(0, len(match.groups())):
|
47 |
+
groupNum = groupNum + 1
|
48 |
+
print(
|
49 |
+
"Group {groupNum} found at {start}-{end}: `{group}`".format(
|
50 |
+
groupNum=groupNum,
|
51 |
+
start=match.start(groupNum),
|
52 |
+
end=match.end(groupNum),
|
53 |
+
group=match.group(groupNum),
|
54 |
+
)
|
55 |
+
)
|
56 |
+
|
57 |
+
start, end = match.span()
|
58 |
+
count += end - start
|
59 |
+
|
60 |
+
return count
|
61 |
+
|
62 |
+
|
63 |
+
def detect_repetitions(text, debug=False):
|
64 |
+
text, count_excessive_whitespaces = del_excessive_whitespaces(text, debug=debug)
|
65 |
+
count_text_repetitions = detect_text_repetitions(text, debug=debug)
|
66 |
+
total_repetitions = count_excessive_whitespaces + count_text_repetitions
|
67 |
+
|
68 |
+
result = (count_excessive_whitespaces, count_text_repetitions, total_repetitions)
|
69 |
+
|
70 |
+
if debug:
|
71 |
+
print(result)
|
72 |
+
return result
|
73 |
+
|
74 |
+
|
75 |
+
def detect_scores(text, debug=False):
|
76 |
+
newline_score, repetition_score, total_repetitions = detect_repetitions(
|
77 |
+
text, debug=debug
|
78 |
+
)
|
79 |
+
return pd.Series([newline_score, repetition_score, total_repetitions])
|
80 |
+
|
81 |
+
|
82 |
+
def load_with_newline_and_repetition_scores(result_file, force_recalculate=False):
|
83 |
+
print(f"loading result file: {result_file}")
|
84 |
+
df = pd.read_csv(result_file, comment="#", on_bad_lines="warn")
|
85 |
+
|
86 |
+
if (
|
87 |
+
force_recalculate
|
88 |
+
or "newline_score" not in df.columns
|
89 |
+
or "repetition_score" not in df.columns
|
90 |
+
or "total_repetitions" not in df.columns
|
91 |
+
):
|
92 |
+
df[["newline_score", "repetition_score", "total_repetitions"]] = df[
|
93 |
+
"answer"
|
94 |
+
].apply(detect_scores)
|
95 |
+
df.to_csv(result_file, index=False)
|
96 |
+
|
97 |
+
return df
|
98 |
+
|
99 |
+
|
100 |
+
def replace_last(source_string, old_string, new_string):
|
101 |
+
head, _sep, tail = source_string.rpartition(old_string)
|
102 |
+
return head + new_string + tail
|
103 |
+
|
104 |
+
|
105 |
+
def load_for_repetition_penalty(
|
106 |
+
csv_result_file, repetition_penalty, force_recalculate=False
|
107 |
+
):
|
108 |
+
result_file = replace_last(
|
109 |
+
csv_result_file, ".csv", f"_RP_{repetition_penalty:.3f}.csv"
|
110 |
+
)
|
111 |
+
return load_with_newline_and_repetition_scores(
|
112 |
+
result_file, force_recalculate=force_recalculate
|
113 |
+
)
|
114 |
+
|
115 |
+
|
116 |
+
def calc_adjusted_performance(f, r):
|
117 |
+
return f / math.log10(10 + r)
|
118 |
+
|
119 |
+
|
120 |
+
def calculate_adjusted_performance(row):
|
121 |
+
r = row["total_repetitions"]
|
122 |
+
adjusted_precision = calc_adjusted_performance(row["precision"], r)
|
123 |
+
adjusted_recall = calc_adjusted_performance(row["recall"], r)
|
124 |
+
return pd.Series([adjusted_precision, adjusted_recall])
|
125 |
+
|
126 |
+
|
127 |
+
def load_performance_df(csv_result_file, repetition_penalty):
|
128 |
+
result_file = replace_last(
|
129 |
+
csv_result_file, ".csv", f"_RP_{repetition_penalty:.3f}-t2_evaluated.json"
|
130 |
+
)
|
131 |
+
result_file = result_file.replace("/results/", "/eval/")
|
132 |
+
print(f"loading json file: {result_file}")
|
133 |
+
df = pd.read_json(result_file)
|
134 |
+
|
135 |
+
return df
|
136 |
+
|
137 |
+
|
138 |
+
def calculate_performance_score_v1(
|
139 |
+
csv_result_file, repetition_penalty, force_recalculate=False
|
140 |
+
):
|
141 |
+
result_file = replace_last(
|
142 |
+
csv_result_file, ".csv", f"_RP_{repetition_penalty:.3f}.csv"
|
143 |
+
)
|
144 |
+
print(f"loading result file: {result_file}")
|
145 |
+
df = pd.read_csv(result_file, comment="#", on_bad_lines="warn")
|
146 |
+
|
147 |
+
if force_recalculate or "f2" in df.columns or "f1" not in df.columns:
|
148 |
+
df.drop(
|
149 |
+
columns=[
|
150 |
+
"precision",
|
151 |
+
"recall",
|
152 |
+
"f1",
|
153 |
+
"f2",
|
154 |
+
"entities_in_answer",
|
155 |
+
"entities_in_question",
|
156 |
+
],
|
157 |
+
errors="ignore",
|
158 |
+
inplace=True,
|
159 |
+
)
|
160 |
+
perf_df = load_performance_df(csv_result_file, repetition_penalty)
|
161 |
+
filtered_df = perf_df[perf_df["id"].isin(df["id"])]
|
162 |
+
perf_df = filtered_df.reset_index(drop=True)
|
163 |
+
print(f"perf_df len: {len(perf_df)}")
|
164 |
+
# print(perf_df.head())
|
165 |
+
|
166 |
+
df["eval_gemini_1.0_pro"] = perf_df["eval_gemini_1.0_pro"]
|
167 |
+
|
168 |
+
df["precision"] = perf_df["score"].apply(lambda x: x[0])
|
169 |
+
df["recall"] = perf_df["score"].apply(lambda x: x[1])
|
170 |
+
df["f1"] = perf_df["score"].apply(lambda x: x[2])
|
171 |
+
|
172 |
+
df[["adjusted_precision", "adjusted_recall"]] = df.apply(
|
173 |
+
calculate_adjusted_performance, axis=1
|
174 |
+
)
|
175 |
+
|
176 |
+
df.to_csv(result_file, index=False)
|
177 |
+
print(f"performance scores saved to result file: {result_file}")
|
178 |
+
|
179 |
+
print(f"df len: {len(df)}")
|
180 |
+
|
181 |
+
return df
|
182 |
+
|
183 |
+
|
184 |
+
ref_df = pd.read_csv(
|
185 |
+
"./data/results/gpt-3.5-turbo_non_rag.csv", comment="#", on_bad_lines="warn"
|
186 |
+
)
|
187 |
+
|
188 |
+
|
189 |
+
def calculate_performance_score(
|
190 |
+
csv_result_file, repetition_penalty, force_recalculate=False
|
191 |
+
):
|
192 |
+
result_file = replace_last(
|
193 |
+
csv_result_file, ".csv", f"_RP_{repetition_penalty:.3f}.csv"
|
194 |
+
)
|
195 |
+
|
196 |
+
re_creating = False
|
197 |
+
if os.path.exists(result_file):
|
198 |
+
print(f"loading result file: {result_file}")
|
199 |
+
df = pd.read_csv(result_file, comment="#", on_bad_lines="warn")
|
200 |
+
else:
|
201 |
+
print(f"re-creating result file: {result_file}")
|
202 |
+
df = pd.DataFrame()
|
203 |
+
force_recalculate = True
|
204 |
+
|
205 |
+
if force_recalculate or "f2" in df.columns or "f1" not in df.columns:
|
206 |
+
df.drop(
|
207 |
+
columns=[
|
208 |
+
"precision",
|
209 |
+
"recall",
|
210 |
+
"f1",
|
211 |
+
"f2",
|
212 |
+
"entities_in_answer",
|
213 |
+
"entities_in_question",
|
214 |
+
"word_count",
|
215 |
+
],
|
216 |
+
errors="ignore",
|
217 |
+
inplace=True,
|
218 |
+
)
|
219 |
+
perf_df = load_performance_df(csv_result_file, repetition_penalty)
|
220 |
+
filtered_df = perf_df[perf_df["id"].isin(ref_df["id"])]
|
221 |
+
perf_df = filtered_df.reset_index(drop=True)
|
222 |
+
print(f"perf_df len: {len(perf_df)}")
|
223 |
+
|
224 |
+
if len(perf_df) != len(ref_df):
|
225 |
+
print(f"error: len(perf_df) != {len(ref_df)}")
|
226 |
+
missing_ids = [
|
227 |
+
id for id in ref_df["id"].unique() if id not in perf_df["id"].unique()
|
228 |
+
]
|
229 |
+
print(f"missing_ids: {missing_ids}")
|
230 |
+
|
231 |
+
# print(perf_df.head())
|
232 |
+
|
233 |
+
df["id"] = perf_df["id"]
|
234 |
+
df["question"] = perf_df["question"]
|
235 |
+
df["answer"] = perf_df["pred_answer"]
|
236 |
+
df["word_count"] = df["answer"].apply(
|
237 |
+
lambda x: len(nltk.word_tokenize(x)) if isinstance(x, str) else 0
|
238 |
+
)
|
239 |
+
df["ground_truth"] = perf_df["ground_truth"]
|
240 |
+
df[["newline_score", "repetition_score", "total_repetitions"]] = df[
|
241 |
+
"answer"
|
242 |
+
].apply(detect_scores)
|
243 |
+
|
244 |
+
df["eval_gemini_1.0_pro"] = perf_df["eval_gemini_1.0_pro"]
|
245 |
+
df["precision"] = perf_df["score"].apply(lambda x: x[0])
|
246 |
+
df["recall"] = perf_df["score"].apply(lambda x: x[1])
|
247 |
+
df["f1"] = perf_df["score"].apply(lambda x: x[2])
|
248 |
+
|
249 |
+
df[["adjusted_precision", "adjusted_recall"]] = df.apply(
|
250 |
+
calculate_adjusted_performance, axis=1
|
251 |
+
)
|
252 |
+
|
253 |
+
df.to_csv(result_file, index=False)
|
254 |
+
print(f"performance scores saved to result file: {result_file}")
|
255 |
+
|
256 |
+
print(f"df len: {len(df)}")
|
257 |
+
|
258 |
+
return df
|
259 |
+
|
260 |
+
|
261 |
+
def adjust_perf_scores_with_repetition_penalty(result, precision, recall):
|
262 |
+
newline_score = [
|
263 |
+
df["newline_score"].mean() for df in result["df_list_repetition_penalty"]
|
264 |
+
]
|
265 |
+
|
266 |
+
repetition_score = [
|
267 |
+
df["repetition_score"].mean() for df in result["df_list_repetition_penalty"]
|
268 |
+
]
|
269 |
+
|
270 |
+
precision = [
|
271 |
+
f / math.log10(10 + n + r)
|
272 |
+
for f, n, r in zip(precision, newline_score, repetition_score)
|
273 |
+
]
|
274 |
+
recall = [
|
275 |
+
f / math.log10(10 + n + r)
|
276 |
+
for f, n, r in zip(recall, newline_score, repetition_score)
|
277 |
+
]
|
278 |
+
|
279 |
+
return precision, recall
|
280 |
+
|
281 |
+
|
282 |
+
def plot_performance_scores(
|
283 |
+
result,
|
284 |
+
models=None,
|
285 |
+
title="Performance",
|
286 |
+
):
|
287 |
+
if models is None:
|
288 |
+
models = result.keys()
|
289 |
+
for model in models:
|
290 |
+
print(f"model: {model}")
|
291 |
+
df = result[model]["df_overall"]
|
292 |
+
|
293 |
+
# Calculate the statistics
|
294 |
+
precision = [
|
295 |
+
df["precision"].mean() for df in result[model]["df_list_repetition_penalty"]
|
296 |
+
]
|
297 |
+
recall = [
|
298 |
+
df["recall"].mean() for df in result[model]["df_list_repetition_penalty"]
|
299 |
+
]
|
300 |
+
f1 = [2 * (p * r) / (p + r) for p, r in zip(precision, recall)]
|
301 |
+
best_f1 = max(f1)
|
302 |
+
best_f1_index = f1.index(best_f1)
|
303 |
+
|
304 |
+
precision, recall = adjust_perf_scores_with_repetition_penalty(
|
305 |
+
result[model], precision, recall
|
306 |
+
)
|
307 |
+
afrp = [2 * (p * r) / (p + r) for p, r in zip(precision, recall)]
|
308 |
+
|
309 |
+
# f1 = [df["f1"].mean() for df in result[model]["df_list_repetition_penalty"]]
|
310 |
+
best_afrp = max(afrp)
|
311 |
+
best_afrp_index = afrp.index(best_afrp)
|
312 |
+
|
313 |
+
adjusted_precision = [
|
314 |
+
df["adjusted_precision"].mean()
|
315 |
+
for df in result[model]["df_list_repetition_penalty"]
|
316 |
+
]
|
317 |
+
adjusted_recall = [
|
318 |
+
df["adjusted_recall"].mean()
|
319 |
+
for df in result[model]["df_list_repetition_penalty"]
|
320 |
+
]
|
321 |
+
afrp2 = [
|
322 |
+
2 * (p * r) / (p + r) for p, r in zip(adjusted_precision, adjusted_recall)
|
323 |
+
]
|
324 |
+
best_afrp2 = max(afrp2)
|
325 |
+
best_afrp2_index = afrp2.index(best_afrp2)
|
326 |
+
|
327 |
+
repetition_penalties = list(df["repetition_penalty"])
|
328 |
+
|
329 |
+
# line plot for precision, recall, f1
|
330 |
+
plt.figure(figsize=(10, 6))
|
331 |
+
|
332 |
+
plt.axvspan(
|
333 |
+
repetition_penalties[best_f1_index] - 0.01,
|
334 |
+
repetition_penalties[best_f1_index] + 0.01,
|
335 |
+
alpha=0.5,
|
336 |
+
edgecolor="none",
|
337 |
+
facecolor="blue",
|
338 |
+
)
|
339 |
+
|
340 |
+
# plt.axvspan(
|
341 |
+
# repetition_penalties[best_afrp2_index] - 0.01,
|
342 |
+
# repetition_penalties[best_afrp2_index] + 0.01,
|
343 |
+
# alpha=0.5,
|
344 |
+
# edgecolor="none",
|
345 |
+
# facecolor="green",
|
346 |
+
# )
|
347 |
+
|
348 |
+
plt.axvspan(
|
349 |
+
repetition_penalties[best_afrp_index] - 0.01,
|
350 |
+
repetition_penalties[best_afrp_index] + 0.01,
|
351 |
+
alpha=0.5,
|
352 |
+
edgecolor="none",
|
353 |
+
facecolor="orange",
|
354 |
+
)
|
355 |
+
|
356 |
+
plt.plot(repetition_penalties, f1, label="F1", marker="D", color="blue")
|
357 |
+
# plt.plot(
|
358 |
+
# repetition_penalties,
|
359 |
+
# afrp2,
|
360 |
+
# label="Per-question RAP - F1",
|
361 |
+
# marker="s",
|
362 |
+
# color="green",
|
363 |
+
# )
|
364 |
+
plt.plot(
|
365 |
+
repetition_penalties,
|
366 |
+
afrp,
|
367 |
+
label="RAP - F1",
|
368 |
+
marker="o",
|
369 |
+
color="orange",
|
370 |
+
)
|
371 |
+
plt.xlabel("Repetition Penalties")
|
372 |
+
plt.ylabel("Score")
|
373 |
+
# plt.xlim(0.99, 1.31)
|
374 |
+
# y in percentage
|
375 |
+
plt.gca().yaxis.set_major_formatter(mtick.PercentFormatter(1.0))
|
376 |
+
plt.title(f"{model} {title}")
|
377 |
+
plt.legend(bbox_to_anchor=(1.0, 0.5), loc="center left")
|
378 |
+
|
379 |
+
plt.show()
|
380 |
+
|
381 |
+
|
382 |
+
def plot_best_afrp(
|
383 |
+
result,
|
384 |
+
models=None,
|
385 |
+
title="Models with Best RAP - F1",
|
386 |
+
ref_result=None,
|
387 |
+
):
|
388 |
+
# Initialize lists to store the statistics
|
389 |
+
model_names = []
|
390 |
+
best_f1 = []
|
391 |
+
best_afrp = []
|
392 |
+
best_repetition_penalty = []
|
393 |
+
best_mtr = []
|
394 |
+
|
395 |
+
if models is None:
|
396 |
+
models = result.keys()
|
397 |
+
for model in models:
|
398 |
+
print(f"model: {model}")
|
399 |
+
df = result[model]["df_overall"]
|
400 |
+
|
401 |
+
# Calculate the statistics
|
402 |
+
precision = [
|
403 |
+
df["precision"].mean() for df in result[model]["df_list_repetition_penalty"]
|
404 |
+
]
|
405 |
+
recall = [
|
406 |
+
df["recall"].mean() for df in result[model]["df_list_repetition_penalty"]
|
407 |
+
]
|
408 |
+
# f1 = [df["f1"].mean() for df in result[model]["df_list_repetition_penalty"]]
|
409 |
+
f1 = [2 * (p * r) / (p + r) for p, r in zip(precision, recall)]
|
410 |
+
|
411 |
+
newline_score = [
|
412 |
+
df["newline_score"].mean()
|
413 |
+
for df in result[model]["df_list_repetition_penalty"]
|
414 |
+
]
|
415 |
+
# print(f"newline_score: {newline_score}")
|
416 |
+
|
417 |
+
repetition_score = [
|
418 |
+
df["repetition_score"].mean()
|
419 |
+
for df in result[model]["df_list_repetition_penalty"]
|
420 |
+
]
|
421 |
+
# print(f"repetition_score: {repetition_score}")
|
422 |
+
|
423 |
+
afrp = [
|
424 |
+
f / math.log10(10 + n + r)
|
425 |
+
for f, n, r in zip(f1, newline_score, repetition_score)
|
426 |
+
]
|
427 |
+
|
428 |
+
best_afrp.append(max(afrp))
|
429 |
+
best_afrp_index = afrp.index(best_afrp[-1])
|
430 |
+
best_repetition_penalty.append(df["repetition_penalty"][best_afrp_index])
|
431 |
+
|
432 |
+
best_f1.append(f1[best_afrp_index])
|
433 |
+
best_mtr.append(
|
434 |
+
newline_score[best_afrp_index] + repetition_score[best_afrp_index]
|
435 |
+
)
|
436 |
+
|
437 |
+
# print(
|
438 |
+
# f"best repetition penalty: {best_repetition_penalty[-1]}, best afrp: {best_afrp[-1]}, f1: {best_f1[-1]}"
|
439 |
+
# )
|
440 |
+
|
441 |
+
df = result[model]["df_list_repetition_penalty"][best_afrp_index]
|
442 |
+
|
443 |
+
model_names.append(
|
444 |
+
f"{model} (RP={best_repetition_penalty[-1]})"
|
445 |
+
) # Add the model name to the list
|
446 |
+
|
447 |
+
if ref_result is not None:
|
448 |
+
print("ref_result:", ref_result)
|
449 |
+
for model in ref_result.keys():
|
450 |
+
model_names.append(model)
|
451 |
+
df = pd.read_csv(ref_result[model])
|
452 |
+
# df = df[df["id"].isin(wikidata_df["id"])]
|
453 |
+
|
454 |
+
p = df["precision"].mean()
|
455 |
+
r = df["recall"].mean()
|
456 |
+
|
457 |
+
f1 = 2 * p * r / (p + r) if p + r > 0 else 0
|
458 |
+
best_f1.append(f1)
|
459 |
+
best_afrp.append(f1)
|
460 |
+
best_mtr.append(0)
|
461 |
+
|
462 |
+
print("model_names:", model_names)
|
463 |
+
# print("best_f1:", best_f1)
|
464 |
+
# print("best_afrp:", best_afrp)
|
465 |
+
|
466 |
+
# Create a DataFrame with the statistics
|
467 |
+
data = pd.DataFrame(
|
468 |
+
{
|
469 |
+
"Model": model_names,
|
470 |
+
"RAP - F1": best_afrp,
|
471 |
+
"F1": best_f1,
|
472 |
+
}
|
473 |
+
)
|
474 |
+
|
475 |
+
# Melt the DataFrame to a long format
|
476 |
+
data_melted = data.melt(id_vars="Model", var_name="Metric", value_name="Score")
|
477 |
+
|
478 |
+
# Pivot the DataFrame to a wide format
|
479 |
+
data_pivoted = data_melted.pivot(index="Metric", columns="Model", values="Score")
|
480 |
+
|
481 |
+
# make sure the columns are following the order of the models
|
482 |
+
data_pivoted = data_pivoted[model_names]
|
483 |
+
|
484 |
+
# make sure three groups in the order of precision, recall, f1
|
485 |
+
data_pivoted = data_pivoted.reindex(["RAP - F1", "F1"])
|
486 |
+
|
487 |
+
# Plot the statistics
|
488 |
+
plt.figure(figsize=(15, 6))
|
489 |
+
ax = data_pivoted.plot(kind="bar", ax=plt.gca(), width=0.9)
|
490 |
+
plt.title(title)
|
491 |
+
plt.legend(bbox_to_anchor=(1.0, 0.5), loc="center left")
|
492 |
+
|
493 |
+
# Set the rotation of the x-axis labels to 0 degrees
|
494 |
+
plt.xticks(rotation=0)
|
495 |
+
|
496 |
+
# Format the y-axis to display as percentage
|
497 |
+
ax.yaxis.set_major_formatter(mtick.PercentFormatter(1.0))
|
498 |
+
|
499 |
+
# get the max value of the y-axis
|
500 |
+
a1 = max(best_afrp)
|
501 |
+
a2 = max(best_f1)
|
502 |
+
|
503 |
+
max_value = max([a1, a2]) * 1.12
|
504 |
+
print("max_value:", max_value)
|
505 |
+
|
506 |
+
# Set the y-axis limit up to 70%
|
507 |
+
ax.set_ylim(0, max_value)
|
508 |
+
|
509 |
+
# Add the values above each bar
|
510 |
+
for p in ax.patches:
|
511 |
+
ax.annotate(
|
512 |
+
f"{p.get_height() * 100:.1f}",
|
513 |
+
(p.get_x() + p.get_width() / 2.0, p.get_height()),
|
514 |
+
ha="center",
|
515 |
+
va="bottom",
|
516 |
+
xytext=(0, 10),
|
517 |
+
textcoords="offset points",
|
518 |
+
rotation=90,
|
519 |
+
)
|
520 |
+
|
521 |
+
plt.show()
|
522 |
+
return data_pivoted, best_mtr
|
523 |
+
|
524 |
+
|
525 |
+
def plot_best_performance(
|
526 |
+
result,
|
527 |
+
models=None,
|
528 |
+
title="Models with Best F1 Score",
|
529 |
+
adjusted_f1=False,
|
530 |
+
ref_result=None,
|
531 |
+
):
|
532 |
+
# Initialize lists to store the statistics
|
533 |
+
model_names = []
|
534 |
+
best_precision = []
|
535 |
+
best_recall = []
|
536 |
+
best_f1 = []
|
537 |
+
best_repetition_penalty = []
|
538 |
+
best_mtr = []
|
539 |
+
|
540 |
+
if models is None:
|
541 |
+
models = result.keys()
|
542 |
+
for model in models:
|
543 |
+
print(f"model: {model}")
|
544 |
+
df = result[model]["df_overall"]
|
545 |
+
|
546 |
+
# Calculate the statistics
|
547 |
+
precision = [
|
548 |
+
df["precision"].mean() for df in result[model]["df_list_repetition_penalty"]
|
549 |
+
]
|
550 |
+
recall = [
|
551 |
+
df["recall"].mean() for df in result[model]["df_list_repetition_penalty"]
|
552 |
+
]
|
553 |
+
newline_score = [
|
554 |
+
df["newline_score"].mean()
|
555 |
+
for df in result[model]["df_list_repetition_penalty"]
|
556 |
+
]
|
557 |
+
|
558 |
+
repetition_score = [
|
559 |
+
df["repetition_score"].mean()
|
560 |
+
for df in result[model]["df_list_repetition_penalty"]
|
561 |
+
]
|
562 |
+
|
563 |
+
if adjusted_f1:
|
564 |
+
precision, recall = adjust_perf_scores_with_repetition_penalty(
|
565 |
+
result[model], precision, recall
|
566 |
+
)
|
567 |
+
|
568 |
+
# f1 = [df["f1"].mean() for df in result[model]["df_list_repetition_penalty"]]
|
569 |
+
f1 = [2 * (p * r) / (p + r) for p, r in zip(precision, recall)]
|
570 |
+
|
571 |
+
best_f1.append(max(f1))
|
572 |
+
best_f1_index = f1.index(best_f1[-1])
|
573 |
+
best_repetition_penalty.append(df["repetition_penalty"][best_f1_index])
|
574 |
+
|
575 |
+
best_precision.append(precision[best_f1_index])
|
576 |
+
best_recall.append(recall[best_f1_index])
|
577 |
+
best_mtr.append(newline_score[best_f1_index] + repetition_score[best_f1_index])
|
578 |
+
|
579 |
+
print(
|
580 |
+
f"best repetition penalty: {best_repetition_penalty[-1]}, best f1: {best_f1[-1]}, precision: {best_precision[-1]}, recall: {best_recall[-1]}"
|
581 |
+
)
|
582 |
+
|
583 |
+
df = result[model]["df_list_repetition_penalty"][best_f1_index]
|
584 |
+
|
585 |
+
model_names.append(
|
586 |
+
f"{model} (RP={best_repetition_penalty[-1]})"
|
587 |
+
) # Add the model name to the list
|
588 |
+
|
589 |
+
# print sum for columns: newline_score, repetition_score
|
590 |
+
print(
|
591 |
+
f"newline_score: {df['newline_score'].sum()}, repetition_score: {df['repetition_score'].sum()}"
|
592 |
+
)
|
593 |
+
|
594 |
+
if ref_result is not None:
|
595 |
+
print("ref_result:", ref_result)
|
596 |
+
for model in ref_result.keys():
|
597 |
+
model_names.append(model)
|
598 |
+
df = pd.read_csv(ref_result[model])
|
599 |
+
# df = df[df["id"].isin(wikidata_df["id"])]
|
600 |
+
|
601 |
+
best_precision.append(df["precision"].mean())
|
602 |
+
best_recall.append(df["recall"].mean())
|
603 |
+
f1 = (
|
604 |
+
2
|
605 |
+
* (best_precision[-1] * best_recall[-1])
|
606 |
+
/ (best_precision[-1] + best_recall[-1])
|
607 |
+
)
|
608 |
+
# best_f1.append(df["f1"].mean())
|
609 |
+
best_f1.append(f1)
|
610 |
+
best_mtr.append(0)
|
611 |
+
|
612 |
+
# Create a DataFrame with the statistics
|
613 |
+
data = (
|
614 |
+
pd.DataFrame(
|
615 |
+
{
|
616 |
+
"Model": model_names,
|
617 |
+
"Adjusted Precision with RP": best_precision,
|
618 |
+
"Adjusted Recall with RP": best_recall,
|
619 |
+
"Adjusted F1 with RP": best_f1,
|
620 |
+
}
|
621 |
+
)
|
622 |
+
if adjusted_f1
|
623 |
+
else pd.DataFrame(
|
624 |
+
{
|
625 |
+
"Model": model_names,
|
626 |
+
"Precision": best_precision,
|
627 |
+
"Recall": best_recall,
|
628 |
+
"F1": best_f1,
|
629 |
+
}
|
630 |
+
)
|
631 |
+
)
|
632 |
+
columns = list(data.columns)
|
633 |
+
|
634 |
+
# Melt the DataFrame to a long format
|
635 |
+
data_melted = data.melt(id_vars="Model", var_name="Metric", value_name="Score")
|
636 |
+
|
637 |
+
# Pivot the DataFrame to a wide format
|
638 |
+
data_pivoted = data_melted.pivot(index="Metric", columns="Model", values="Score")
|
639 |
+
|
640 |
+
# make sure the columns are following the order of the models
|
641 |
+
data_pivoted = data_pivoted[model_names]
|
642 |
+
|
643 |
+
# make sure three groups in the order of precision, recall, f1
|
644 |
+
data_pivoted = data_pivoted.reindex(columns[1:])
|
645 |
+
|
646 |
+
# Plot the statistics
|
647 |
+
plt.figure(figsize=(10, 6))
|
648 |
+
ax = data_pivoted.plot(kind="bar", ax=plt.gca(), width=0.9)
|
649 |
+
plt.title(title)
|
650 |
+
plt.legend(bbox_to_anchor=(1.0, 0.5), loc="center left")
|
651 |
+
|
652 |
+
# Set the rotation of the x-axis labels to 0 degrees
|
653 |
+
plt.xticks(rotation=0)
|
654 |
+
|
655 |
+
# Format the y-axis to display as percentage
|
656 |
+
ax.yaxis.set_major_formatter(mtick.PercentFormatter(1.0))
|
657 |
+
|
658 |
+
# get the max value of the y-axis
|
659 |
+
a1 = max(best_precision)
|
660 |
+
a2 = max(best_recall)
|
661 |
+
a3 = max(best_f1)
|
662 |
+
|
663 |
+
max_value = max([a1, a2, a3]) * 1.12
|
664 |
+
print("max_value:", max_value)
|
665 |
+
|
666 |
+
# Set the y-axis limit up to 70%
|
667 |
+
ax.set_ylim(0, max_value)
|
668 |
+
|
669 |
+
# Add the values above each bar
|
670 |
+
for p in ax.patches:
|
671 |
+
ax.annotate(
|
672 |
+
f"{p.get_height() * 100:.1f}",
|
673 |
+
(p.get_x() + p.get_width() / 2.0, p.get_height()),
|
674 |
+
ha="center",
|
675 |
+
va="bottom",
|
676 |
+
xytext=(0, 10),
|
677 |
+
textcoords="offset points",
|
678 |
+
rotation=90,
|
679 |
+
)
|
680 |
+
|
681 |
+
plt.show()
|
682 |
+
return data_pivoted, best_mtr
|
683 |
+
|
684 |
+
|
685 |
+
def plot_best_performance_ms_macro(
|
686 |
+
result,
|
687 |
+
models=None,
|
688 |
+
title="Models with Best RAP - Performance",
|
689 |
+
ref_result=None,
|
690 |
+
skip_generic_prompt=False,
|
691 |
+
include_adjusted_performance=True,
|
692 |
+
):
|
693 |
+
# Initialize lists to store the statistics
|
694 |
+
model_names = []
|
695 |
+
best_f1 = []
|
696 |
+
best_afrp = []
|
697 |
+
best_repetition_penalty = []
|
698 |
+
best_bleu1 = []
|
699 |
+
best_rougeL = []
|
700 |
+
best_mtr = []
|
701 |
+
|
702 |
+
if models is None:
|
703 |
+
models = result.keys()
|
704 |
+
for model in models:
|
705 |
+
if skip_generic_prompt and "generic prompt" in model:
|
706 |
+
continue
|
707 |
+
print(f"model: {model}")
|
708 |
+
df = result[model]["df_overall"]
|
709 |
+
|
710 |
+
# Calculate the statistics
|
711 |
+
bleu1 = [x for x in df["bleu1"]]
|
712 |
+
rougeL = [x for x in df["rougeL"]]
|
713 |
+
f1 = [2 * (p * r) / (p + r) for p, r in zip(bleu1, rougeL)]
|
714 |
+
|
715 |
+
newline_score = [
|
716 |
+
df["newline_score"].mean()
|
717 |
+
for df in result[model]["df_list_repetition_penalty"]
|
718 |
+
]
|
719 |
+
# print(f"newline_score: {newline_score}")
|
720 |
+
|
721 |
+
repetition_score = [
|
722 |
+
df["repetition_score"].mean()
|
723 |
+
for df in result[model]["df_list_repetition_penalty"]
|
724 |
+
]
|
725 |
+
# print(f"repetition_score: {repetition_score}")
|
726 |
+
|
727 |
+
afrp = [
|
728 |
+
f / math.log10(10 + n + r)
|
729 |
+
for f, n, r in zip(f1, newline_score, repetition_score)
|
730 |
+
]
|
731 |
+
|
732 |
+
best_afrp.append(max(afrp if include_adjusted_performance else f1))
|
733 |
+
best_afrp_index = (
|
734 |
+
afrp.index(best_afrp[-1])
|
735 |
+
if include_adjusted_performance
|
736 |
+
else f1.index(best_afrp[-1])
|
737 |
+
)
|
738 |
+
best_repetition_penalty.append(df["repetition_penalty"][best_afrp_index])
|
739 |
+
|
740 |
+
best_f1.append(f1[best_afrp_index])
|
741 |
+
best_bleu1.append(bleu1[best_afrp_index])
|
742 |
+
best_rougeL.append(rougeL[best_afrp_index])
|
743 |
+
best_mtr.append(
|
744 |
+
newline_score[best_afrp_index] + repetition_score[best_afrp_index]
|
745 |
+
)
|
746 |
+
|
747 |
+
# print(
|
748 |
+
# f"best repetition penalty: {best_repetition_penalty[-1]}, best afrp: {best_afrp[-1]}, f1: {best_f1[-1]}"
|
749 |
+
# )
|
750 |
+
|
751 |
+
df = result[model]["df_list_repetition_penalty"][best_afrp_index]
|
752 |
+
|
753 |
+
model_names.append(
|
754 |
+
f"{model} (RP={best_repetition_penalty[-1]})"
|
755 |
+
) # Add the model name to the list
|
756 |
+
|
757 |
+
if ref_result is not None:
|
758 |
+
print("ref_result:", ref_result)
|
759 |
+
for model in ref_result.keys():
|
760 |
+
model_names.append(model)
|
761 |
+
df = pd.read_csv(ref_result[model], comment="#", on_bad_lines="warn")
|
762 |
+
# df = df[df["id"].isin(wikidata_df["id"])]
|
763 |
+
|
764 |
+
p = df["bleu1"][0]
|
765 |
+
best_bleu1.append(p)
|
766 |
+
|
767 |
+
r = df["rougeL"][0]
|
768 |
+
best_rougeL.append(r)
|
769 |
+
|
770 |
+
f1 = 2 * p * r / (p + r) if p + r > 0 else 0
|
771 |
+
best_f1.append(f1)
|
772 |
+
best_afrp.append(f1)
|
773 |
+
best_mtr.append(0)
|
774 |
+
|
775 |
+
# print("model_names:", model_names)
|
776 |
+
# print("best_f1:", best_f1)
|
777 |
+
# print("best_afrp:", best_afrp)
|
778 |
+
|
779 |
+
# Create a DataFrame with the statistics
|
780 |
+
data = (
|
781 |
+
pd.DataFrame(
|
782 |
+
{
|
783 |
+
"Model": model_names,
|
784 |
+
"RAP - Perf Score": best_afrp,
|
785 |
+
"Overall Perf Score": best_f1,
|
786 |
+
}
|
787 |
+
)
|
788 |
+
if include_adjusted_performance
|
789 |
+
else pd.DataFrame(
|
790 |
+
{
|
791 |
+
"Model": model_names,
|
792 |
+
"Bleu-1": best_bleu1,
|
793 |
+
"Rouge-L": best_rougeL,
|
794 |
+
"Overall Perf Score": best_f1,
|
795 |
+
}
|
796 |
+
)
|
797 |
+
)
|
798 |
+
|
799 |
+
# Melt the DataFrame to a long format
|
800 |
+
data_melted = data.melt(id_vars="Model", var_name="Metric", value_name="Score")
|
801 |
+
|
802 |
+
# Pivot the DataFrame to a wide format
|
803 |
+
data_pivoted = data_melted.pivot(index="Metric", columns="Model", values="Score")
|
804 |
+
|
805 |
+
# make sure the columns are following the order of the models
|
806 |
+
data_pivoted = data_pivoted[model_names]
|
807 |
+
|
808 |
+
columns = list(data.columns)
|
809 |
+
data_pivoted = data_pivoted.reindex(columns[1:])
|
810 |
+
|
811 |
+
# Plot the statistics
|
812 |
+
plt.figure(figsize=(10, 6))
|
813 |
+
ax = data_pivoted.plot(kind="bar", ax=plt.gca(), width=0.9)
|
814 |
+
plt.title(title)
|
815 |
+
plt.legend(bbox_to_anchor=(1.0, 0.5), loc="center left")
|
816 |
+
|
817 |
+
# Set the rotation of the x-axis labels to 0 degrees
|
818 |
+
plt.xticks(rotation=0)
|
819 |
+
|
820 |
+
# Format the y-axis to display as percentage
|
821 |
+
ax.yaxis.set_major_formatter(mtick.PercentFormatter(1.0))
|
822 |
+
|
823 |
+
# get the max value of the y-axis
|
824 |
+
a1 = max(best_afrp)
|
825 |
+
a2 = max(best_f1)
|
826 |
+
a3 = max(best_bleu1)
|
827 |
+
a4 = max(best_rougeL)
|
828 |
+
|
829 |
+
max_value = (
|
830 |
+
max([a1, a2] if include_adjusted_performance else [a1, a2, a3, a4]) * 1.12
|
831 |
+
)
|
832 |
+
print("max_value:", max_value)
|
833 |
+
|
834 |
+
# Set the y-axis limit up to 70%
|
835 |
+
ax.set_ylim(0, max_value)
|
836 |
+
|
837 |
+
# Add the values above each bar
|
838 |
+
for p in ax.patches:
|
839 |
+
ax.annotate(
|
840 |
+
f"{p.get_height() * 100:.1f}",
|
841 |
+
(p.get_x() + p.get_width() / 2.0, p.get_height()),
|
842 |
+
ha="center",
|
843 |
+
va="bottom",
|
844 |
+
xytext=(0, 10),
|
845 |
+
textcoords="offset points",
|
846 |
+
rotation=90,
|
847 |
+
)
|
848 |
+
|
849 |
+
plt.show()
|
850 |
+
return data_pivoted, best_mtr
|
851 |
+
|
852 |
+
|
853 |
+
all_open_source_models = [
|
854 |
+
"gemma-1.1-2b-it",
|
855 |
+
"Phi-3-mini-128k-instruct",
|
856 |
+
"gemma-1.1-7b-it",
|
857 |
+
"Llama-2-7b-chat-hf",
|
858 |
+
"Mistral-7B-Instruct-v0.2",
|
859 |
+
"Meta-Llama-3-8B-Instruct",
|
860 |
+
"Llama-2-13b-chat-hf",
|
861 |
+
"Llama-2-70b-chat-hf",
|
862 |
+
"Meta-Llama-3-70B-Instruct",
|
863 |
+
]
|
864 |
+
|
865 |
+
|
866 |
+
non_rag_csv_result_files = [
|
867 |
+
"./data/results/gemma-1.1-2b-it_wd_non_rag.csv", # gemma-1.1-2b-it
|
868 |
+
"./data/results/Phi-3-mini-128k-instruct_wd_non_rag_batch_16.csv", # Phi-3-mini-128k-instruct(batch size:16)
|
869 |
+
"./data/results/gemma-1.1-7b-it_wd_non_rag.csv", # gemma-1.1-7b-it
|
870 |
+
"./data/results/Tune_2024-04-09_09-19-22.csv", # Llama-2-7b-chat-hf
|
871 |
+
"./data/results/Tune_2024-04-16_12-24-27.csv.csv", # Mistral-7B-Instruct-v0.2
|
872 |
+
"./data/results/Meta-Llama-3-8B-Instruct_wd_non_rag.csv", # Meta-Llama-3-8B-Instruct
|
873 |
+
"./data/results/Meta-Llama-3-8B-Instruct_wd_1_non_rag.csv", # Meta-Llama-3-8B-Instruct
|
874 |
+
"./data/results/Tune_2024-04-10_16-53-38.csv", # Llama-2-13b-chat-hf
|
875 |
+
"./data/results/Llama-2-70b-chat-hf_wd_non_rag.csv", # Llama-2-70b-chat-hf
|
876 |
+
"./data/results/Meta-Llama-3-70B-Instruct_wd_non_rag.csv", # Meta-Llama-3-70B-Instruct
|
877 |
+
]
|
878 |
+
|
879 |
+
rag_csv_result_files = [
|
880 |
+
"./data/results/gemma-1.1-2b-it_wd.csv", # gemma-1.1-2b-it
|
881 |
+
"./data/results/gemma-1.1-2b-it_wd_true.csv", # gemma-1.1-2b-it(true)
|
882 |
+
"./data/results/Phi-3-mini-128k-instruct_wd_rag_batch_4.csv", # Phi-3-mini-128k-instruct(batch size:16)
|
883 |
+
"./data/results/Phi-3-mini-128k-instruct_wd_true.csv", # Phi-3-mini-128k-instruct(batch size:16)
|
884 |
+
"./data/results/gemma-1.1-7b-it_wd.csv", # gemma-1.1-7b-it
|
885 |
+
"./data/results/gemma-1.1-7b-it_wd_true.csv", # gemma-1.1-7b-it(true)
|
886 |
+
"./data/results/Tune_2024-03-20_15-35-37.csv", # Llama-2-7b-chat-hf
|
887 |
+
"./data/results/Llama-2-7b-chat-hf_wd_true.csv", # Llama-2-7b-chat-hf(true)
|
888 |
+
"./data/results/Tune_2024-03-29_11-28-20.csv", # Mistral-7B-Instruct-v0.2
|
889 |
+
"./data/results/Mistral-7B-Instruct-v0.2_wd_true.csv", # Mistral-7B-Instruct-v0.2(true)
|
890 |
+
"./data/results/Meta-Llama-3-8B-Instruct_wd.csv", # Meta-Llama-3-8b-instruct
|
891 |
+
"./data/results/Meta-Llama-3-8B-Instruct_wd_true.csv", # Meta-Llama-3-8b-instruct(true)
|
892 |
+
"./data/results/Tune_2024-03-25_23-32-57.csv", # Llama-2-13b-chat-hf
|
893 |
+
"./data/results/Llama-2-13b-chat-hf_wd_true.csv", # Llama-2-13b-chat-hf(true)
|
894 |
+
"./data/results/Llama-2-70b-chat-hf_wd.csv", # Llama-2-70b-chat-hf
|
895 |
+
"./data/results/Llama-2-70b-chat-hf_wd_true.csv", # Llama-2-70b-chat-hf
|
896 |
+
"./data/results/Meta-Llama-3-70B-Instruct_wd.csv", # Meta-Llama-3-70B-Instruct
|
897 |
+
"./data/results/Meta-Llama-3-70B-Instruct_wd_true.csv", # Meta-Llama-3-70B-Instruct(true)
|
898 |
+
]
|
899 |
+
|
900 |
+
df_ms_macro = pd.read_json("./data/datasets/ms_macro.json")
|
901 |
+
|
902 |
+
|
903 |
+
def load_for_repetition_penalty_ms_macro(
|
904 |
+
csv_result_file, repetition_penalty, force_recalculate=False
|
905 |
+
):
|
906 |
+
result_file = replace_last(
|
907 |
+
csv_result_file, ".csv", f"_RP_{repetition_penalty:.3f}.csv"
|
908 |
+
)
|
909 |
+
df = load_with_newline_and_repetition_scores(
|
910 |
+
result_file, force_recalculate=force_recalculate
|
911 |
+
)
|
912 |
+
|
913 |
+
if len(df) != len(df_ms_macro):
|
914 |
+
print(f"error: len(df) != {len(df_ms_macro)}")
|
915 |
+
missing_ids = [
|
916 |
+
id for id in df_ms_macro["id"].unique() if id not in df["id"].unique()
|
917 |
+
]
|
918 |
+
print(f"missing_ids: {missing_ids}")
|
919 |
+
|
920 |
+
if df["ground_truth"][0] != str(df_ms_macro["wellFormedAnswers"][0]):
|
921 |
+
df["ground_truth"] = df_ms_macro["wellFormedAnswers"]
|
922 |
+
print("ground_truth updated for:", result_file)
|
923 |
+
df.to_csv(result_file, index=False)
|
924 |
+
return df
|
925 |
+
|
926 |
+
|
927 |
+
# MS MACRO
|
928 |
+
def plot_performance_scores_ms_macro(
|
929 |
+
result,
|
930 |
+
models=None,
|
931 |
+
title="Performance",
|
932 |
+
):
|
933 |
+
if models is None:
|
934 |
+
models = result.keys()
|
935 |
+
for model in models:
|
936 |
+
print(f"model: {model}")
|
937 |
+
df = result[model]["df_overall"]
|
938 |
+
# print(result[model]["df_list_repetition_penalty"][0].describe())
|
939 |
+
|
940 |
+
# Calculate the statistics
|
941 |
+
bleu1 = list(df["bleu1"])
|
942 |
+
rougeL = list(df["rougeL"])
|
943 |
+
f1 = [2 * (p * r) / (p + r) for p, r in zip(bleu1, rougeL)]
|
944 |
+
best_f1 = max(f1)
|
945 |
+
best_f1_index = f1.index(best_f1)
|
946 |
+
|
947 |
+
bleu1, rougeL = adjust_perf_scores_with_repetition_penalty(
|
948 |
+
result[model], bleu1, rougeL
|
949 |
+
)
|
950 |
+
afrp = [2 * (p * r) / (p + r) for p, r in zip(bleu1, rougeL)]
|
951 |
+
|
952 |
+
# f1 = [df["f1"].mean() for df in result[model]["df_list_repetition_penalty"]]
|
953 |
+
best_afrp = max(afrp)
|
954 |
+
best_afrp_index = afrp.index(best_afrp)
|
955 |
+
|
956 |
+
repetition_penalties = list(df["repetition_penalty"])
|
957 |
+
|
958 |
+
# line plot for precision, recall, f1
|
959 |
+
plt.figure(figsize=(10, 6))
|
960 |
+
|
961 |
+
plt.axvspan(
|
962 |
+
repetition_penalties[best_f1_index] - 0.01,
|
963 |
+
repetition_penalties[best_f1_index] + 0.01,
|
964 |
+
alpha=0.5,
|
965 |
+
edgecolor="none",
|
966 |
+
facecolor="blue",
|
967 |
+
)
|
968 |
+
|
969 |
+
plt.axvspan(
|
970 |
+
repetition_penalties[best_afrp_index] - 0.01,
|
971 |
+
repetition_penalties[best_afrp_index] + 0.01,
|
972 |
+
alpha=0.5,
|
973 |
+
edgecolor="none",
|
974 |
+
facecolor="orange",
|
975 |
+
)
|
976 |
+
|
977 |
+
plt.plot(
|
978 |
+
repetition_penalties,
|
979 |
+
f1,
|
980 |
+
label="Overall Perf Score",
|
981 |
+
marker="D",
|
982 |
+
color="blue",
|
983 |
+
)
|
984 |
+
plt.plot(
|
985 |
+
repetition_penalties,
|
986 |
+
afrp,
|
987 |
+
label="RAP - Perf Score",
|
988 |
+
marker="o",
|
989 |
+
color="orange",
|
990 |
+
)
|
991 |
+
|
992 |
+
plt.xlabel("Repetition Penalties")
|
993 |
+
plt.ylabel("Score")
|
994 |
+
# plt.xlim(0.99, 1.31)
|
995 |
+
# y in percentage
|
996 |
+
plt.gca().yaxis.set_major_formatter(mtick.PercentFormatter(1.0))
|
997 |
+
plt.title(f"{model} {title}")
|
998 |
+
plt.legend(bbox_to_anchor=(1.0, 0.5), loc="center left")
|
999 |
+
|
1000 |
+
plt.show()
|
1001 |
+
|
1002 |
+
|
1003 |
+
def plot_repetition_factors(result, groups):
|
1004 |
+
for group in groups:
|
1005 |
+
# Plot the statistics
|
1006 |
+
plt.figure(figsize=(10, 6))
|
1007 |
+
|
1008 |
+
max_value = 0
|
1009 |
+
for model in result.keys():
|
1010 |
+
if not group in model.lower():
|
1011 |
+
continue
|
1012 |
+
print(f"model: {model}")
|
1013 |
+
df = result[model]["df_overall"]
|
1014 |
+
repetition_panelties = [
|
1015 |
+
repetition_penalty for repetition_penalty in df["repetition_penalty"]
|
1016 |
+
]
|
1017 |
+
|
1018 |
+
mean_score = [
|
1019 |
+
# math.log10(10 + df["total_repetitions"].mean())
|
1020 |
+
df["total_repetitions"].mean()
|
1021 |
+
for df in result[model]["df_list_repetition_penalty"]
|
1022 |
+
]
|
1023 |
+
|
1024 |
+
sns.lineplot(x=repetition_panelties, y=mean_score, label=model)
|
1025 |
+
|
1026 |
+
new_max = max(mean_score)
|
1027 |
+
if new_max > max_value:
|
1028 |
+
max_value = new_max
|
1029 |
+
|
1030 |
+
max_value = max_value * 1.05
|
1031 |
+
# if max_value < 1.5:
|
1032 |
+
# max_value = 1.5
|
1033 |
+
# set ylimit
|
1034 |
+
plt.ylim(0, max_value)
|
1035 |
+
|
1036 |
+
# show grid
|
1037 |
+
plt.grid(True)
|
1038 |
+
plt.xlabel("Repetition Penalties")
|
1039 |
+
plt.ylabel("Mean Total Repetitions")
|
1040 |
+
plt.title("Mean Total Repetitions vs Repetition Penalties")
|
1041 |
+
plt.legend()
|
1042 |
+
|
1043 |
+
plt.show()
|
1044 |
+
|
1045 |
+
|
1046 |
+
def plot_repetition_factors_by_group(result, group_filter=None):
|
1047 |
+
markers = ["D", "o", "s", "x"]
|
1048 |
+
colors = ["blue", "orange", "green", "red"]
|
1049 |
+
|
1050 |
+
# Plot the statistics
|
1051 |
+
plt.figure(figsize=(10, 6))
|
1052 |
+
index = 0
|
1053 |
+
max_value = 0
|
1054 |
+
|
1055 |
+
for model in result.keys():
|
1056 |
+
if group_filter is not None and group_filter not in model:
|
1057 |
+
continue
|
1058 |
+
|
1059 |
+
print(f"model: {model}")
|
1060 |
+
|
1061 |
+
df = result[model]["df_overall"]
|
1062 |
+
repetition_panelties = [
|
1063 |
+
repetition_penalty for repetition_penalty in df["repetition_penalty"]
|
1064 |
+
]
|
1065 |
+
|
1066 |
+
# Calculate the statistics
|
1067 |
+
mean_score = [
|
1068 |
+
# math.log10(10 + df["total_repetitions"].mean())
|
1069 |
+
df["total_repetitions"].mean()
|
1070 |
+
for df in result[model]["df_list_repetition_penalty"]
|
1071 |
+
]
|
1072 |
+
if len(mean_score) != len(repetition_panelties):
|
1073 |
+
print(
|
1074 |
+
f"model: {model} has different length of repetition penalties and mean score"
|
1075 |
+
)
|
1076 |
+
print("repetition_panelties:", len(repetition_panelties))
|
1077 |
+
print("mean_score:", len(mean_score))
|
1078 |
+
continue
|
1079 |
+
|
1080 |
+
new_max = max(mean_score)
|
1081 |
+
if new_max > max_value:
|
1082 |
+
max_value = new_max
|
1083 |
+
|
1084 |
+
sns.lineplot(
|
1085 |
+
x=repetition_panelties,
|
1086 |
+
y=mean_score,
|
1087 |
+
label=model,
|
1088 |
+
marker=markers[index],
|
1089 |
+
color=colors[index],
|
1090 |
+
)
|
1091 |
+
|
1092 |
+
index += 1
|
1093 |
+
|
1094 |
+
max_value = max_value * 1.05
|
1095 |
+
# if max_value < 1.5:
|
1096 |
+
# max_value = 1.5
|
1097 |
+
# set ylimit
|
1098 |
+
plt.ylim(0, max_value)
|
1099 |
+
max_value = 0
|
1100 |
+
|
1101 |
+
plt.xlabel("Repetition Penalties")
|
1102 |
+
plt.ylabel("Mean Total Repetitions")
|
1103 |
+
plt.title("Mean Total Repetitions vs Repetition Penalties")
|
1104 |
+
plt.legend(bbox_to_anchor=(1.0, 0.5), loc="center left")
|
1105 |
+
|
1106 |
+
plt.show()
|
1107 |
+
|
1108 |
+
|
1109 |
+
ms_marco_csv_result_files = [
|
1110 |
+
"data/results/gemma-1.1-2b-it_mm_true_false.csv",
|
1111 |
+
"data/results/gemma-1.1-2b-it_mm_true.csv",
|
1112 |
+
"data/results/gemma-1.1-2b-it_mm_true_false_non_rag.csv",
|
1113 |
+
"data/results/Phi-3-mini-128k-instruct_mm_false.csv",
|
1114 |
+
"data/results/Phi-3-mini-128k-instruct_mm_true.csv",
|
1115 |
+
"data/results/Phi-3-mini-128k-instruct_mm_non_rag.csv",
|
1116 |
+
"data/results/gemma-1.1-7b-it_mm_false.csv",
|
1117 |
+
"data/results/gemma-1.1-7b-it_mm_true.csv",
|
1118 |
+
"data/results/gemma-1.1-7b-it_mm_non_rag.csv",
|
1119 |
+
"data/results/Llama-2-7b-chat-hf_mm_true_false.csv",
|
1120 |
+
"data/results/Llama-2-7b-chat-hf_mm_true.csv",
|
1121 |
+
"data/results/Llama-2-7b-chat-hf_mm_true_false_non_rag.csv",
|
1122 |
+
"data/results/Mistral-7B-Instruct-v0.2_mm_false.csv",
|
1123 |
+
"data/results/Mistral-7B-Instruct-v0.2_mm_true.csv",
|
1124 |
+
"data/results/Mistral-7B-Instruct-v0.2_mm_non_rag.csv",
|
1125 |
+
"data/results/Meta-Llama-3-8B-Instruct_mm_true_false.csv",
|
1126 |
+
"data/results/Meta-Llama-3-8B-Instruct_mm_true.csv",
|
1127 |
+
"data/results/Meta-Llama-3-8B-Instruct_mm_true_false_non_rag.csv",
|
1128 |
+
"data/results/Llama-2-13b-chat-hf_mm_false.csv",
|
1129 |
+
"data/results/Llama-2-13b-chat-hf_mm_true.csv",
|
1130 |
+
"data/results/Llama-2-13b-chat-hf_mm_non_rag.csv",
|
1131 |
+
"data/results/Llama-2-70b-chat-hf_mm_false.csv",
|
1132 |
+
"data/results/Llama-2-70b-chat-hf_mm_true.csv",
|
1133 |
+
"data/results/Llama-2-70b-chat-hf_mm_non_rag.csv",
|
1134 |
+
"data/results/Meta-Llama-3-70B-Instruct_mm_false.csv",
|
1135 |
+
"data/results/Meta-Llama-3-70B-Instruct_mm_true.csv",
|
1136 |
+
"data/results/Meta-Llama-3-70B-Instruct_mm_non_rag.csv",
|
1137 |
+
]
|
1138 |
+
|
1139 |
+
webqsp_csv_result_files = []
|
1140 |
+
webqsp_model_result_counts = {}
|
1141 |
+
|
1142 |
+
|
1143 |
+
def find_model_name(file_path):
|
1144 |
+
df = pd.read_csv(file_path, comment="#", on_bad_lines="warn")
|
1145 |
+
return df["model"][0]
|
1146 |
+
|
1147 |
+
|
1148 |
+
def add_file(file):
|
1149 |
+
model_name = find_model_name(file)
|
1150 |
+
if "(generic prompt)" not in model_name:
|
1151 |
+
webqsp_csv_result_files.append(file)
|
1152 |
+
if model_name not in webqsp_model_result_counts:
|
1153 |
+
webqsp_model_result_counts[model_name] = 1
|
1154 |
+
else:
|
1155 |
+
webqsp_model_result_counts[model_name] += 1
|
1156 |
+
|
1157 |
+
|
1158 |
+
last_model_name = None
|
1159 |
+
non_rag_index = 0
|
1160 |
+
|
1161 |
+
for csv_result_file in rag_csv_result_files:
|
1162 |
+
try:
|
1163 |
+
model_name = find_model_name(csv_result_file)
|
1164 |
+
# print(f"processing model: {model_name} - {csv_result_file}")
|
1165 |
+
|
1166 |
+
if last_model_name != model_name and last_model_name is not None:
|
1167 |
+
while non_rag_index < len(non_rag_csv_result_files):
|
1168 |
+
# print(f"processing non-rag file - {file}")
|
1169 |
+
file = non_rag_csv_result_files[non_rag_index]
|
1170 |
+
non_model_name = find_model_name(file)
|
1171 |
+
if non_model_name.startswith(last_model_name):
|
1172 |
+
add_file(file)
|
1173 |
+
non_rag_index += 1
|
1174 |
+
else:
|
1175 |
+
break
|
1176 |
+
|
1177 |
+
add_file(csv_result_file)
|
1178 |
+
last_model_name = model_name
|
1179 |
+
except FileNotFoundError as e:
|
1180 |
+
print("\terror processing file: ", csv_result_file, e)
|
1181 |
+
continue
|
1182 |
+
|
1183 |
+
for file in non_rag_csv_result_files[non_rag_index:]:
|
1184 |
+
add_file(file)
|
1185 |
+
|
1186 |
+
|
1187 |
+
def calc_rap_scores(result, precision="precision", recall="recall"):
|
1188 |
+
newline_score = [
|
1189 |
+
df["newline_score"].mean() for df in result["df_list_repetition_penalty"]
|
1190 |
+
]
|
1191 |
+
|
1192 |
+
repetition_score = [
|
1193 |
+
df["repetition_score"].mean() for df in result["df_list_repetition_penalty"]
|
1194 |
+
]
|
1195 |
+
|
1196 |
+
if precision in result["df_list_repetition_penalty"][0].columns:
|
1197 |
+
precision = [
|
1198 |
+
df[precision].mean() for df in result["df_list_repetition_penalty"]
|
1199 |
+
]
|
1200 |
+
recall = [df[recall].mean() for df in result["df_list_repetition_penalty"]]
|
1201 |
+
else:
|
1202 |
+
precision = result["df_overall"][precision]
|
1203 |
+
recall = result["df_overall"][recall]
|
1204 |
+
|
1205 |
+
f1 = [2 * (p * r) / (p + r) for p, r in zip(precision, recall)]
|
1206 |
+
|
1207 |
+
rap = [
|
1208 |
+
f / math.log10(10 + n + r)
|
1209 |
+
for f, n, r in zip(f1, newline_score, repetition_score)
|
1210 |
+
]
|
1211 |
+
|
1212 |
+
return newline_score, repetition_score, f1, rap
|
1213 |
+
|
1214 |
+
|
1215 |
+
def load_webqsp_result(csv_result_files, force_recalculate=False):
|
1216 |
+
model_name_exts = {
|
1217 |
+
"true": "(RAG - Chat Template)",
|
1218 |
+
"wd": "(RAG - Generic Prompt)",
|
1219 |
+
"rag": "(Non-RAG)",
|
1220 |
+
}
|
1221 |
+
|
1222 |
+
result = {}
|
1223 |
+
for i, csv_result_file in enumerate(csv_result_files):
|
1224 |
+
try:
|
1225 |
+
df = pd.read_csv(csv_result_file)
|
1226 |
+
parts = re.split(r"[_\.]", csv_result_file)
|
1227 |
+
if parts[-2] in model_name_exts.keys():
|
1228 |
+
key = parts[-2]
|
1229 |
+
elif csv_result_file in non_rag_csv_result_files:
|
1230 |
+
key = "rag"
|
1231 |
+
else:
|
1232 |
+
key = "wd"
|
1233 |
+
model_name = f'{df["model"][0]}{model_name_exts[key]}'
|
1234 |
+
dfs = [
|
1235 |
+
calculate_performance_score(
|
1236 |
+
csv_result_file,
|
1237 |
+
repetition_penalty,
|
1238 |
+
force_recalculate=force_recalculate,
|
1239 |
+
)
|
1240 |
+
for repetition_penalty in df["repetition_penalty"]
|
1241 |
+
]
|
1242 |
+
|
1243 |
+
answer_lens = []
|
1244 |
+
for df_rpp in dfs:
|
1245 |
+
df_rpp["answer_len"] = df_rpp["answer"].apply(
|
1246 |
+
lambda x: len(x) if isinstance(x, str) else 0
|
1247 |
+
)
|
1248 |
+
answer_lens.append(df_rpp["answer_len"].mean())
|
1249 |
+
|
1250 |
+
result[model_name] = {
|
1251 |
+
"df_overall": df,
|
1252 |
+
"df_list_repetition_penalty": dfs,
|
1253 |
+
"file": csv_result_file,
|
1254 |
+
}
|
1255 |
+
newline_score, repetition_score, perf, rap = calc_rap_scores(
|
1256 |
+
result[model_name]
|
1257 |
+
)
|
1258 |
+
df["newline_score"] = newline_score
|
1259 |
+
df["repetition_score"] = repetition_score
|
1260 |
+
df["total_repetitions"] = df["newline_score"] + df["repetition_score"]
|
1261 |
+
df["answer_len"] = answer_lens
|
1262 |
+
df["perf"] = perf
|
1263 |
+
df["rap"] = rap
|
1264 |
+
except Exception as e:
|
1265 |
+
print(f"Error: {e}")
|
1266 |
+
|
1267 |
+
return result
|
1268 |
+
|
1269 |
+
|
1270 |
+
def load_ms_marco_result(csv_result_files, force_recalculate=False):
|
1271 |
+
model_name_exts = {
|
1272 |
+
"true": "(RAG - Chat Template)",
|
1273 |
+
"false": "(RAG - Generic Prompt)",
|
1274 |
+
"rag": "(Non-RAG)",
|
1275 |
+
}
|
1276 |
+
|
1277 |
+
result = {}
|
1278 |
+
for csv_result_file in csv_result_files:
|
1279 |
+
try:
|
1280 |
+
df = pd.read_csv(csv_result_file)
|
1281 |
+
|
1282 |
+
parts = re.split(r"[_\.]", csv_result_file)
|
1283 |
+
model_name = f'{df["model"][0]}{model_name_exts[parts[-2]]}'
|
1284 |
+
|
1285 |
+
print(f"\tmodel_name: {model_name}")
|
1286 |
+
dfs = [
|
1287 |
+
load_for_repetition_penalty_ms_macro(
|
1288 |
+
csv_result_file,
|
1289 |
+
repetition_penalty,
|
1290 |
+
force_recalculate=force_recalculate,
|
1291 |
+
)
|
1292 |
+
for repetition_penalty in df["repetition_penalty"]
|
1293 |
+
]
|
1294 |
+
|
1295 |
+
answer_lens = []
|
1296 |
+
for df_rpp in dfs:
|
1297 |
+
df_rpp["answer_len"] = df_rpp["answer"].apply(
|
1298 |
+
lambda x: len(x) if isinstance(x, str) else 0
|
1299 |
+
)
|
1300 |
+
answer_lens.append(df_rpp["answer_len"].mean())
|
1301 |
+
|
1302 |
+
result[model_name] = {
|
1303 |
+
"df_overall": df,
|
1304 |
+
"df_list_repetition_penalty": dfs,
|
1305 |
+
"file": csv_result_file,
|
1306 |
+
}
|
1307 |
+
newline_score, repetition_score, perf, rap = calc_rap_scores(
|
1308 |
+
result[model_name],
|
1309 |
+
precision="bleu1",
|
1310 |
+
recall="rougeL",
|
1311 |
+
)
|
1312 |
+
df["newline_score"] = newline_score
|
1313 |
+
df["repetition_score"] = repetition_score
|
1314 |
+
df["total_repetitions"] = df["newline_score"] + df["repetition_score"]
|
1315 |
+
df["answer_len"] = answer_lens
|
1316 |
+
df["perf"] = perf
|
1317 |
+
df["rap"] = rap
|
1318 |
+
except Exception as e:
|
1319 |
+
print(f"Error: {e}")
|
1320 |
+
|
1321 |
+
return result
|
1322 |
+
|
1323 |
+
|
1324 |
+
def load_ms_marco_result_v2(csv_result_files, force_recalculate=False):
|
1325 |
+
model_name_exts = {
|
1326 |
+
"true": "(RAG - Chat Template)",
|
1327 |
+
"false": "(RAG - Generic Prompt)",
|
1328 |
+
"rag": "(Non-RAG)",
|
1329 |
+
}
|
1330 |
+
|
1331 |
+
result = {}
|
1332 |
+
for csv_result_file in csv_result_files:
|
1333 |
+
try:
|
1334 |
+
df = pd.read_csv(csv_result_file)
|
1335 |
+
|
1336 |
+
parts = re.split(r"[_\.]", csv_result_file)
|
1337 |
+
model_name = f'{df["model"][0]}{model_name_exts[parts[-2]]}'
|
1338 |
+
|
1339 |
+
print(f"\tmodel_name: {model_name}")
|
1340 |
+
dfs = [
|
1341 |
+
load_for_repetition_penalty_ms_macro(
|
1342 |
+
csv_result_file,
|
1343 |
+
repetition_penalty,
|
1344 |
+
force_recalculate=force_recalculate,
|
1345 |
+
)
|
1346 |
+
for repetition_penalty in df["repetition_penalty"]
|
1347 |
+
]
|
1348 |
+
|
1349 |
+
answer_lens = []
|
1350 |
+
for df_rpp in dfs:
|
1351 |
+
df_rpp["answer_len"] = df_rpp["answer"].apply(
|
1352 |
+
lambda x: len(x) if isinstance(x, str) else 0
|
1353 |
+
)
|
1354 |
+
answer_lens.append(df_rpp["answer_len"].mean())
|
1355 |
+
df["answer_len"] = answer_lens
|
1356 |
+
|
1357 |
+
meteor_scores = []
|
1358 |
+
for df_rpp in dfs:
|
1359 |
+
meteor_score = meteor.compute(
|
1360 |
+
predictions=df_rpp["answer"], references=df_rpp["ground_truth"]
|
1361 |
+
)["meteor"]
|
1362 |
+
meteor_scores.append(meteor_score)
|
1363 |
+
df["meteor_scores"] = meteor_scores
|
1364 |
+
|
1365 |
+
result[model_name] = {
|
1366 |
+
"df_overall": df,
|
1367 |
+
"df_list_repetition_penalty": dfs,
|
1368 |
+
"file": csv_result_file,
|
1369 |
+
}
|
1370 |
+
newline_score, repetition_score, perf, rap = calc_rap_scores(
|
1371 |
+
result[model_name],
|
1372 |
+
precision="meteor_scores",
|
1373 |
+
recall="meteor_scores",
|
1374 |
+
)
|
1375 |
+
df["newline_score"] = newline_score
|
1376 |
+
df["repetition_score"] = repetition_score
|
1377 |
+
df["total_repetitions"] = df["newline_score"] + df["repetition_score"]
|
1378 |
+
df["perf"] = perf
|
1379 |
+
df["rap"] = rap
|
1380 |
+
except Exception as e:
|
1381 |
+
print(f"Error: {e}")
|
1382 |
+
|
1383 |
+
return result
|
notebooks/00_Repetition_Algorithms_Comparison.ipynb
ADDED
The diff for this file is too large to render.
See raw diff
|
|
notebooks/04_RAPGeT_v2.ipynb
ADDED
The diff for this file is too large to render.
See raw diff
|
|