Upload 4 files
Browse files- LLaMIPa3/adapter_config.json +26 -0
- LLaMIPa3/adapter_model.safetensors +3 -0
- data/parser_test_moves_15.jsonl +0 -0
- parser_generate.py +127 -0
LLaMIPa3/adapter_config.json
ADDED
@@ -0,0 +1,26 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
1 |
+
{
|
2 |
+
"alpha_pattern": {},
|
3 |
+
"auto_mapping": null,
|
4 |
+
"base_model_name_or_path": "/tmpdir/thompson/Meta-Llama-3-8B/",
|
5 |
+
"bias": "none",
|
6 |
+
"fan_in_fan_out": false,
|
7 |
+
"inference_mode": true,
|
8 |
+
"init_lora_weights": true,
|
9 |
+
"layers_pattern": null,
|
10 |
+
"layers_to_transform": null,
|
11 |
+
"loftq_config": {},
|
12 |
+
"lora_alpha": 16,
|
13 |
+
"lora_dropout": 0.1,
|
14 |
+
"megatron_config": null,
|
15 |
+
"megatron_core": "megatron.core",
|
16 |
+
"modules_to_save": null,
|
17 |
+
"peft_type": "LORA",
|
18 |
+
"r": 64,
|
19 |
+
"rank_pattern": {},
|
20 |
+
"revision": null,
|
21 |
+
"target_modules": [
|
22 |
+
"q_proj",
|
23 |
+
"v_proj"
|
24 |
+
],
|
25 |
+
"task_type": "CAUSAL_LM"
|
26 |
+
}
|
LLaMIPa3/adapter_model.safetensors
ADDED
@@ -0,0 +1,3 @@
|
|
|
|
|
|
|
|
|
1 |
+
version https://git-lfs.github.com/spec/v1
|
2 |
+
oid sha256:956bbad58e3c7d8101baa809d11f3aa025ddf5d36925d0c78eea43d656ad2b37
|
3 |
+
size 109069176
|
data/parser_test_moves_15.jsonl
ADDED
The diff for this file is too large to render.
See raw diff
|
|
parser_generate.py
ADDED
@@ -0,0 +1,127 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
1 |
+
import torch
|
2 |
+
import json
|
3 |
+
from transformers import AutoModelForCausalLM, AutoTokenizer, pipeline
|
4 |
+
from datasets import load_dataset
|
5 |
+
from tqdm import tqdm
|
6 |
+
|
7 |
+
device_map = "auto"
|
8 |
+
model = AutoModelForCausalLM.from_pretrained(
|
9 |
+
"/path/to/meta-llama3-8b/",
|
10 |
+
return_dict=True,
|
11 |
+
torch_dtype=torch.float16,
|
12 |
+
device_map=device_map)
|
13 |
+
|
14 |
+
|
15 |
+
tokenizer = AutoTokenizer.from_pretrained("/path/to/meta-llama3-8b/",add_eos_token=True)
|
16 |
+
|
17 |
+
tokenizer.pad_token_id = tokenizer.eos_token_id + 1
|
18 |
+
tokenizer.padding_side = "right"
|
19 |
+
|
20 |
+
pipe = pipeline(task="text-generation", model=model, tokenizer=tokenizer, pad_token_id=tokenizer.pad_token_id, max_new_tokens=100)
|
21 |
+
|
22 |
+
test_dataset = load_dataset("json", data_files={'test':'/path/to/parser_test_moves_15.jsonl'})["test"]
|
23 |
+
|
24 |
+
def is_first_moves(sample):
|
25 |
+
answer = 0
|
26 |
+
slist = sample.split('\n')
|
27 |
+
if slist[0].startswith('Context: 0 <Buil> Mission has started.'):
|
28 |
+
struct = [i for i in slist if i.startswith('Structure:')]
|
29 |
+
rels = struct[0].split(':')[1].strip()
|
30 |
+
if len(rels) == 0:
|
31 |
+
answer = 1
|
32 |
+
return answer
|
33 |
+
|
34 |
+
|
35 |
+
def check_endpoints(struct, head):
|
36 |
+
"""
|
37 |
+
takes a struct string and a head int and returns only
|
38 |
+
the struct rels with sources that are >= head
|
39 |
+
"""
|
40 |
+
new_rels_list = []
|
41 |
+
new_rels = None
|
42 |
+
if struct:
|
43 |
+
rels = struct.split(' ')
|
44 |
+
for rel in rels:
|
45 |
+
if len(rel) > 0:
|
46 |
+
source = int(rel.split('(')[1].split(',')[0].strip())
|
47 |
+
if source >= head:
|
48 |
+
new_rels_list.append(rel)
|
49 |
+
if len(new_rels_list) > 0:
|
50 |
+
new_rels = ' '.join(new_rels_list)
|
51 |
+
return new_rels
|
52 |
+
|
53 |
+
def add_previous(sample, previous, predictions):
|
54 |
+
new_output = []
|
55 |
+
keep_str = None
|
56 |
+
#get head
|
57 |
+
slist = sample.split('\n')
|
58 |
+
head = int(slist[0].split('Context:')[1].split('<')[0].strip())
|
59 |
+
# check current structure
|
60 |
+
for s in slist:
|
61 |
+
if s.startswith('Structure:'):
|
62 |
+
new_structure = check_endpoints(previous, head)
|
63 |
+
if new_structure:
|
64 |
+
s = 'Structure: ' + new_structure + ' ' + predictions
|
65 |
+
keep_str = new_structure + ' ' + predictions
|
66 |
+
else:
|
67 |
+
s = 'Structure: ' + predictions
|
68 |
+
keep_str = predictions
|
69 |
+
new_output.append(s)
|
70 |
+
new_output_string = '\n'.join(new_output)
|
71 |
+
return keep_str, new_output_string
|
72 |
+
|
73 |
+
def format_gen(preds):
|
74 |
+
labels = ['COM','CONTR','CORR','QAP','ACK','ELAB','CLARIFQ','COND','CONTIN',
|
75 |
+
'RES','EXPL','QELAB','ALT','NARR','CONFQ','SEQ']
|
76 |
+
split_list = [st.strip() for st in preds.split(' ')]
|
77 |
+
clean_list = []
|
78 |
+
for a in split_list:
|
79 |
+
s_tuple = None
|
80 |
+
rel = None
|
81 |
+
try:
|
82 |
+
s = a.split('(')[1].split(')')[0].split(',')
|
83 |
+
r = a.split('(')[0].strip()
|
84 |
+
except IndexError:
|
85 |
+
print('split error one')
|
86 |
+
else:
|
87 |
+
try:
|
88 |
+
s_tuple = (int(s[0]), int(s[1]))
|
89 |
+
except IndexError:
|
90 |
+
print('split error two')
|
91 |
+
except ValueError:
|
92 |
+
print('value error three')
|
93 |
+
if r in labels:
|
94 |
+
#make sure the label is well-formed
|
95 |
+
rel = r
|
96 |
+
if rel != None and s_tuple != None:
|
97 |
+
clean_list.append(rel + '(' + str(s_tuple[0]) + ',' + str(s_tuple[1]) + ')')
|
98 |
+
clean_preds = ' '.join(clean_list)
|
99 |
+
return clean_preds
|
100 |
+
|
101 |
+
|
102 |
+
def formatting_prompts_func(example):
|
103 |
+
output_text = '<|begin_of_text|>Identify the discourse structure (DS) for the new turn in the following excerpt :\n' + example + '\n ### DS:'
|
104 |
+
return output_text
|
105 |
+
|
106 |
+
|
107 |
+
f = open("/path/to/val-output-file.txt","w")
|
108 |
+
|
109 |
+
new_generations = None
|
110 |
+
previous_generations = None
|
111 |
+
for datum in tqdm(test_dataset['sample']):
|
112 |
+
|
113 |
+
#figure out if it's a first example
|
114 |
+
if is_first_moves(datum):
|
115 |
+
text = formatting_prompts_func(datum)
|
116 |
+
previous_generations = None
|
117 |
+
else:
|
118 |
+
#need to make sure head edu and relations match up
|
119 |
+
update_prev, amended_text = add_previous(datum, previous_generations, new_generations)
|
120 |
+
previous_generations = update_prev
|
121 |
+
text = formatting_prompts_func(amended_text)
|
122 |
+
generated = pipe(text)[0]['generated_text']
|
123 |
+
print(generated, file=f)
|
124 |
+
new_generations = format_gen(generated.split('### DS:')[1])
|
125 |
+
|
126 |
+
f.close()
|
127 |
+
|