Create scrolls.py
Browse files- metrics/scrolls.py +261 -0
metrics/scrolls.py
ADDED
@@ -0,0 +1,261 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
1 |
+
# coding=utf-8
|
2 |
+
# Copyright 2020 The HuggingFace Datasets Authors.
|
3 |
+
#
|
4 |
+
# Licensed under the Apache License, Version 2.0 (the "License");
|
5 |
+
# you may not use this file except in compliance with the License.
|
6 |
+
# You may obtain a copy of the License at
|
7 |
+
#
|
8 |
+
# http://www.apache.org/licenses/LICENSE-2.0
|
9 |
+
#
|
10 |
+
# Unless required by applicable law or agreed to in writing, software
|
11 |
+
# distributed under the License is distributed on an "AS IS" BASIS,
|
12 |
+
# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
|
13 |
+
# See the License for the specific language governing permissions and
|
14 |
+
# limitations under the License.
|
15 |
+
""" Scrolls benchmark metric. """
|
16 |
+
|
17 |
+
from collections import defaultdict
|
18 |
+
from copy import deepcopy
|
19 |
+
import datasets
|
20 |
+
|
21 |
+
# fmt: off
|
22 |
+
from .rouge import compute_rouge, postprocess_text as rouge_postprocess_text # From: https://huggingface.co/datasets/tau/scrolls/raw/main/metrics/rouge.py
|
23 |
+
from .exact_match import compute_exact_match # From: https://huggingface.co/datasets/tau/scrolls/raw/main/metrics/exact_match.py
|
24 |
+
from .f1 import compute_f1 # From: https://huggingface.co/datasets/tau/scrolls/raw/main/metrics/f1.py
|
25 |
+
# fmt: on
|
26 |
+
|
27 |
+
_CITATION = """\
|
28 |
+
# TODO: Add citation
|
29 |
+
"""
|
30 |
+
|
31 |
+
_DESCRIPTION = """\
|
32 |
+
Scrolls: Standardized CompaRison Over Long Language Sequences
|
33 |
+
Recent progress in NLP has created models that can process long inputs consisting of thousands of words.
|
34 |
+
But how well do these models understand the information in the input text?
|
35 |
+
The Scrolls benchmark aims to measure the ability of models to semantically understand long texts.
|
36 |
+
"""
|
37 |
+
|
38 |
+
_KWARGS_DESCRIPTION = """
|
39 |
+
Compute Scrolls evaluation metric associated to each Scrolls dataset.
|
40 |
+
Args:
|
41 |
+
predictions: list of predictions to score.
|
42 |
+
Each prediction should be a string.
|
43 |
+
references: list of lists of references for each example.
|
44 |
+
Each reference should be a string.
|
45 |
+
Returns: depending on the Scrolls subset, one or several of:
|
46 |
+
"exact_match": Exact Match score
|
47 |
+
"f1": F1 score
|
48 |
+
"rouge": ROUGE score
|
49 |
+
Examples:
|
50 |
+
predictions = ["exact match example", "hello there", "general kenobi"] # List[str]
|
51 |
+
references = [["exact match example"], ["hello", "hi there"], ["commander kenobi"]] # List[List[str]]
|
52 |
+
|
53 |
+
>>> scrolls_metric = datasets.load_metric('src/metrics/scrolls.py', 'gov_report') # 'gov_report' or any of ["qmsum", "summ_screen_fd"]
|
54 |
+
>>> results = scrolls_metric.compute(predictions=predictions, references=references)
|
55 |
+
>>> print(results)
|
56 |
+
{'rouge/rouge1': 72.2222, 'rouge/rouge2': 33.3333, 'rouge/rougeL': 72.2222, 'rouge/rougeLsum': 72.2222, 'rouge/geometric_mean': 55.8136, 'num_predicted': 3, 'mean_prediction_length_characters': 14.6667}
|
57 |
+
|
58 |
+
>>> scrolls_metric = datasets.load_metric('src/metrics/scrolls.py', 'contract_nli') # 'contract_nli' or any of ["quality", "quality_hard"]
|
59 |
+
>>> results = scrolls_metric.compute(predictions=predictions, references=references)
|
60 |
+
>>> print(results)
|
61 |
+
{'exact_match': 33.3333, 'num_predicted': 3, 'mean_prediction_length_characters': 14.6667}
|
62 |
+
|
63 |
+
>>> scrolls_metric = datasets.load_metric('src/metrics/scrolls.py', 'narrative_qa') # 'narrative_qa' or any of ["qasper"]
|
64 |
+
>>> results = scrolls_metric.compute(predictions=predictions, references=references)
|
65 |
+
>>> print(results)
|
66 |
+
{'f1': 72.2222, 'num_predicted': 3, 'mean_prediction_length_characters': 14.6667}
|
67 |
+
"""
|
68 |
+
|
69 |
+
DATASET_TO_METRICS = {
|
70 |
+
"contract_nli": {"metrics_to_compute": ["exact_match"], "score": "exact_match"},
|
71 |
+
"gov_report": {"metrics_to_compute": ["rouge"], "score": "rouge/geometric_mean"},
|
72 |
+
"narrative_qa": {"metrics_to_compute": ["f1"], "score": "f1"},
|
73 |
+
"qasper": {"metrics_to_compute": ["f1"], "score": "f1"},
|
74 |
+
"qmsum": {"metrics_to_compute": ["rouge"], "score": "rouge/geometric_mean"},
|
75 |
+
"summ_screen_fd": {"metrics_to_compute": ["rouge"], "score": "rouge/geometric_mean"},
|
76 |
+
"quality": {"metrics_to_compute": ["exact_match"], "score": "exact_match"},
|
77 |
+
"quality_hard": {"metrics_to_compute": ["exact_match"], "score": "exact_match"},
|
78 |
+
}
|
79 |
+
|
80 |
+
|
81 |
+
@datasets.utils.file_utils.add_start_docstrings(_DESCRIPTION, _KWARGS_DESCRIPTION)
|
82 |
+
class Scrolls(datasets.Metric):
|
83 |
+
def __init__(self, *args, **kwargs):
|
84 |
+
super().__init__(*args, **kwargs)
|
85 |
+
|
86 |
+
self._compute_helper_kwargs_fn = {
|
87 |
+
"rouge": lambda: {
|
88 |
+
"metric_fn": compute_rouge,
|
89 |
+
"agg_fn": max,
|
90 |
+
"metric_fn_kwargs": {"use_stemmer": False},
|
91 |
+
"metric_returns_per_example": True,
|
92 |
+
"transform_single_input_fn": lambda text: rouge_postprocess_text(text),
|
93 |
+
"transform_result_fn": lambda output: {
|
94 |
+
key: (value[0] if isinstance(value, list) else value).fmeasure * 100
|
95 |
+
for key, value in output.items()
|
96 |
+
},
|
97 |
+
"transform_aggregated_result_fn": lambda output: output.update(
|
98 |
+
{"geometric_mean": (output["rouge1"] * output["rouge2"] * output["rougeL"]) ** (1.0 / 3.0)}
|
99 |
+
)
|
100 |
+
or output,
|
101 |
+
},
|
102 |
+
"exact_match": lambda: {
|
103 |
+
"metric_fn": compute_exact_match,
|
104 |
+
"agg_fn": None, # compute_exact_match already takes max
|
105 |
+
"transform_result_fn": lambda output: {None: output},
|
106 |
+
},
|
107 |
+
"f1": lambda: {
|
108 |
+
"metric_fn": compute_f1,
|
109 |
+
"agg_fn": None, # compute_f1 already takes max
|
110 |
+
"transform_result_fn": lambda output: {None: output},
|
111 |
+
},
|
112 |
+
}
|
113 |
+
|
114 |
+
custom_metrics = (
|
115 |
+
[metric for metric in self.config_name.split(",") if len(metric) > 0]
|
116 |
+
if self.config_name.startswith(",")
|
117 |
+
else None
|
118 |
+
)
|
119 |
+
if custom_metrics is not None:
|
120 |
+
for metric in custom_metrics:
|
121 |
+
if metric not in self._compute_helper_kwargs_fn:
|
122 |
+
raise KeyError(
|
123 |
+
f"You should supply a metric name selected in {list(self._compute_helper_kwargs_fn.keys())}"
|
124 |
+
)
|
125 |
+
self._metrics_to_compute = custom_metrics
|
126 |
+
else:
|
127 |
+
if self.config_name not in DATASET_TO_METRICS:
|
128 |
+
raise KeyError(f"You should supply a configuration name selected in {list(DATASET_TO_METRICS.keys())}")
|
129 |
+
self._metrics_to_compute = DATASET_TO_METRICS[self.config_name]["metrics_to_compute"]
|
130 |
+
|
131 |
+
def _info(self):
|
132 |
+
return datasets.MetricInfo(
|
133 |
+
description=_DESCRIPTION,
|
134 |
+
citation=_CITATION,
|
135 |
+
inputs_description=_KWARGS_DESCRIPTION,
|
136 |
+
features=datasets.Features(
|
137 |
+
{
|
138 |
+
"predictions": datasets.Value("string"),
|
139 |
+
"references": datasets.Sequence(datasets.Value("string")),
|
140 |
+
}
|
141 |
+
),
|
142 |
+
codebase_urls=[],
|
143 |
+
reference_urls=[],
|
144 |
+
)
|
145 |
+
|
146 |
+
def convert_from_map_format(self, id_to_pred, id_to_labels):
|
147 |
+
index_to_id = list(id_to_pred.keys())
|
148 |
+
predictions = [id_to_pred[id_] for id_ in index_to_id]
|
149 |
+
references = [id_to_labels[id_] for id_ in index_to_id]
|
150 |
+
return {"predictions": predictions, "references": references}
|
151 |
+
|
152 |
+
def _compute(self, predictions, references):
|
153 |
+
metrics = {}
|
154 |
+
for metric in self._metrics_to_compute:
|
155 |
+
result = _compute_helper(
|
156 |
+
deepcopy(predictions),
|
157 |
+
deepcopy(references),
|
158 |
+
**self._compute_helper_kwargs_fn[metric](),
|
159 |
+
)
|
160 |
+
metrics.update(
|
161 |
+
{(f"{metric}/{key}" if key is not None else metric): value for key, value in result.items()}
|
162 |
+
)
|
163 |
+
metrics["num_predicted"] = len(predictions)
|
164 |
+
prediction_lengths = [len(prediction) for prediction in predictions]
|
165 |
+
metrics["mean_prediction_length_characters"] = sum(prediction_lengths) / len(prediction_lengths)
|
166 |
+
|
167 |
+
metrics = {key: round(value, 4) for key, value in metrics.items()}
|
168 |
+
|
169 |
+
if self.config_name in DATASET_TO_METRICS:
|
170 |
+
metrics["score"] = metrics[DATASET_TO_METRICS[self.config_name]["score"]]
|
171 |
+
|
172 |
+
return metrics
|
173 |
+
|
174 |
+
|
175 |
+
def _compute_helper(
|
176 |
+
predictions,
|
177 |
+
references,
|
178 |
+
metric_fn,
|
179 |
+
agg_fn,
|
180 |
+
metric_fn_kwargs=None,
|
181 |
+
transform_single_input_fn=None,
|
182 |
+
transform_result_fn=None,
|
183 |
+
transform_aggregated_result_fn=None,
|
184 |
+
metric_returns_per_example=False,
|
185 |
+
):
|
186 |
+
if metric_fn_kwargs is None:
|
187 |
+
metric_fn_kwargs = {}
|
188 |
+
|
189 |
+
if agg_fn is None:
|
190 |
+
assert metric_returns_per_example is False
|
191 |
+
|
192 |
+
if transform_single_input_fn is not None:
|
193 |
+
predictions = [transform_single_input_fn(prediction) for prediction in predictions]
|
194 |
+
references = [
|
195 |
+
[transform_single_input_fn(reference) for reference in reference_list] for reference_list in references
|
196 |
+
]
|
197 |
+
|
198 |
+
if transform_result_fn is None:
|
199 |
+
transform_result_fn = lambda x: x
|
200 |
+
do_transform_result = False
|
201 |
+
else:
|
202 |
+
do_transform_result = True
|
203 |
+
|
204 |
+
if transform_aggregated_result_fn is None:
|
205 |
+
transform_aggregated_result_fn = lambda x: x
|
206 |
+
|
207 |
+
if agg_fn is not None:
|
208 |
+
# Required when the metric doesn't do the aggregation we need
|
209 |
+
scores = defaultdict(list)
|
210 |
+
if metric_returns_per_example is False:
|
211 |
+
# If when given a list of prediction and references the metric returns an aggregated score,
|
212 |
+
# we need to compute the metric for each prediction and reference and then aggregate the results.
|
213 |
+
# This is only an issue when we want to get the best aggregated score (e.g. max) for prediction
|
214 |
+
# with multiple references.
|
215 |
+
for prediction, reference_list in zip(predictions, references):
|
216 |
+
prediction_scores = defaultdict(list)
|
217 |
+
for reference in reference_list:
|
218 |
+
result = transform_result_fn(metric_fn([prediction], [reference], **metric_fn_kwargs))
|
219 |
+
for key in result:
|
220 |
+
prediction_scores[key].append(result[key])
|
221 |
+
for key in prediction_scores:
|
222 |
+
scores[key].append(agg_fn(prediction_scores[key]))
|
223 |
+
else:
|
224 |
+
# Flatten the references and then aggregate per prediction with agg_fn
|
225 |
+
mapping = [[] for _ in range(len(predictions))]
|
226 |
+
flattened_predictions = []
|
227 |
+
flattened_references = []
|
228 |
+
for i, prediction in enumerate(predictions):
|
229 |
+
for reference in references[i]:
|
230 |
+
flattened_predictions.append(prediction)
|
231 |
+
flattened_references.append(reference)
|
232 |
+
mapping[i].append(len(flattened_references) - 1)
|
233 |
+
|
234 |
+
results = metric_fn(flattened_predictions, flattened_references, **metric_fn_kwargs)
|
235 |
+
if isinstance(results, dict):
|
236 |
+
# Convert a dictionary with lists per key to a list with dictionary with the same keys per element
|
237 |
+
results_list = [{k: None for k in results} for _ in range(len(flattened_predictions))]
|
238 |
+
for k, v in results.items():
|
239 |
+
for i in range(len(v)):
|
240 |
+
results_list[i][k] = v[i]
|
241 |
+
else:
|
242 |
+
results_list = results
|
243 |
+
|
244 |
+
if do_transform_result:
|
245 |
+
for i in range(len(results_list)):
|
246 |
+
results_list[i] = transform_result_fn(results_list[i])
|
247 |
+
|
248 |
+
for reference_indexes in mapping:
|
249 |
+
prediction_scores = defaultdict(list)
|
250 |
+
for reference_index in reference_indexes:
|
251 |
+
result = results_list[reference_index]
|
252 |
+
for key in result:
|
253 |
+
prediction_scores[key].append(result[key])
|
254 |
+
for key in prediction_scores:
|
255 |
+
scores[key].append(agg_fn(prediction_scores[key]))
|
256 |
+
|
257 |
+
return transform_aggregated_result_fn({key: sum(value) / len(value) for key, value in scores.items()})
|
258 |
+
else:
|
259 |
+
return transform_aggregated_result_fn(
|
260 |
+
transform_result_fn(metric_fn(predictions, references, **metric_fn_kwargs))
|
261 |
+
)
|