-
Notifications
You must be signed in to change notification settings - Fork 1
/
ft_results.py
400 lines (353 loc) · 22 KB
/
ft_results.py
1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
50
51
52
53
54
55
56
57
58
59
60
61
62
63
64
65
66
67
68
69
70
71
72
73
74
75
76
77
78
79
80
81
82
83
84
85
86
87
88
89
90
91
92
93
94
95
96
97
98
99
100
101
102
103
104
105
106
107
108
109
110
111
112
113
114
115
116
117
118
119
120
121
122
123
124
125
126
127
128
129
130
131
132
133
134
135
136
137
138
139
140
141
142
143
144
145
146
147
148
149
150
151
152
153
154
155
156
157
158
159
160
161
162
163
164
165
166
167
168
169
170
171
172
173
174
175
176
177
178
179
180
181
182
183
184
185
186
187
188
189
190
191
192
193
194
195
196
197
198
199
200
201
202
203
204
205
206
207
208
209
210
211
212
213
214
215
216
217
218
219
220
221
222
223
224
225
226
227
228
229
230
231
232
233
234
235
236
237
238
239
240
241
242
243
244
245
246
247
248
249
250
251
252
253
254
255
256
257
258
259
260
261
262
263
264
265
266
267
268
269
270
271
272
273
274
275
276
277
278
279
280
281
282
283
284
285
286
287
288
289
290
291
292
293
294
295
296
297
298
299
300
301
302
303
304
305
306
307
308
309
310
311
312
313
314
315
316
317
318
319
320
321
322
323
324
325
326
327
328
329
330
331
332
333
334
335
336
337
338
339
340
341
342
343
344
345
346
347
348
349
350
351
352
353
354
355
356
357
358
359
360
361
362
363
364
365
366
367
368
369
370
371
372
373
374
375
376
377
378
379
380
381
382
383
384
385
386
387
388
389
390
391
392
393
394
395
396
397
398
399
400
import json
import os
import re
import numpy as np
import argparse
from tqdm import tqdm
from glob import glob
from sklearn.metrics import classification_report, confusion_matrix
from pal_code_exec import compute
from util import *
os.environ["HF_ALLOW_CODE_EVAL"] = "1"
parser = argparse.ArgumentParser(formatter_class=argparse.ArgumentDefaultsHelpFormatter)
parser.add_argument("--task", type=str, choices=TASK_LIST)
parser.add_argument("--stage", type=str, choices=['choose_best_rft_checkpoint', 'eval_greedy_refine', 'eval_sampling_refine'])
parser.add_argument("--input", type=str, choices=['prompted', 'cot_ft'], help='source of intial generations')
parser.add_argument("--input_model", type=str, choices=MODEL_LIST, default='meta-llama/Llama-2-13b-chat-hf')
parser.add_argument("--verifier_type", type=str, choices=['prompted', 'ours_ft_on_prompted_solutions', 'ours_ft_on_cot_ft_solutions', 'ours_ft_cot_init_on_prompted_solutions', 'ours_ft_base_init_on_cot_ft_solutions', 'ours_ft_on_base_rft_solutions', 'ours_ft_base_init_on_base_rft_solutions'])
parser.add_argument("--verifier_model", type=str, choices=MODEL_LIST+['gpt-35-turbo', 'gpt-35-turbo-sc', 'gpt-4'])
parser.add_argument("--refiner_type", type=str, choices=['prompted', 'filtered_feedback4_refinement_diff', 'filtered_feedback4_refinement_cot_ft_diff', 'filtered_feedback4_refinement_base_rft_diff'])
parser.add_argument("--refiner_model", type=str, choices=MODEL_LIST, default='meta-llama/Llama-2-13b-chat-hf')
parser.add_argument("--refiner_ft_data_size", type=str, choices=["3625", "7250", "10875", "more"], default='more')
parser.add_argument("--refiner_cot_init", action='store_true')
parser.add_argument("--verbose", action='store_true')
parser.add_argument("--split", type=str, choices=["dev", "test"])
parser.add_argument("--best_dev_confidence", type=float, default=None)
args = parser.parse_args()
TASK = args.task
split = args.split
stage = args.stage
INPUT = args.input
INPUT_MODEL = args.input_model.replace('/', '-')
VERIFIER_TYPE = args.verifier_type
VERIFIER_MODEL = args.verifier_model.replace('/', '-')
REFINER_TYPE = args.refiner_type
REFINER_MODEL = args.refiner_model.replace('/', '-')
VERBOSE = args.verbose
best_dev_confidence = args.best_dev_confidence
refiner_cot_init = args.refiner_cot_init
refiner_ft_data_size = args.refiner_ft_data_size
already_correct_message = TASK2ALREADY_CORRECT_MSG[TASK]
parsing_func = TASK2PARSING_FUNC[TASK]
dev_size = TASK2DEV_SIZE[TASK]
def clean_code(c):
c_findall = re.findall("\`\`\`\s*(.*?)\`\`\`", c, re.DOTALL)
if len(c_findall) == 0:
code_part = c
else:
code_part = c_findall[-1] # correction or already-correct generation
return code_part.replace("```", "").replace("### END ###", "").strip() + "\nprint(solution())"
if stage == "eval_greedy_refine":
if split == "dev":
initial_prompting_solution_infile = f"logs/{TASK}/generations_{INPUT_MODEL}_score.jsonl_greedy"
elif split == "test":
initial_prompting_solution_infile = f"logs/{TASK}/generations_{INPUT_MODEL}_score_test.jsonl_greedy"
with open(initial_prompting_solution_infile, 'r') as reader:
items = [json.loads(l) for l in reader]
if split == "dev" and len(items) > dev_size:
items = items[dev_size:]
initial_prompting_solution_scores = [item['score'][0] for item in items]
if stage == "eval_sampling_refine":
if split == "dev":
initial_prompting_sampling_solution_infile = f"logs/{TASK}/generations_{INPUT_MODEL}_score.jsonl"
elif split == "test":
initial_prompting_sampling_solution_infile = f"logs/{TASK}/generations_{INPUT_MODEL}_score_test.jsonl"
with open(initial_prompting_sampling_solution_infile, 'r') as reader:
items = [json.loads(l) for l in reader]
if split == "dev" and len(items) > dev_size:
items = items[dev_size:]
initial_prompting_sampling_solution_scores = [majority_voting(item['extracted_answers'], item['reference'])[-1] for item in items]
if split == "dev":
refs_infile = f"data/{TASK}/train_shuf.jsonl"
elif split == "test":
refs_infile = f"data/{TASK}/test_shuf.jsonl"
with open(refs_infile, 'r') as reader:
if TASK in ["gsm8k", "gsm8k_nl", "math"]:
refs = [json.loads(l)['answer'].split('### ')[-1] for l in reader]
elif TASK in ["csqa", "riddlesense", "qasc"]:
refs = [json.loads(l)['answer'] for l in reader]
else:
raise ValueError(f"invalid TASK name {TASK}")
if split == "dev":
refs = refs[dev_size:]
test_example_cnt = len(refs)
# pred_dir = "../LLaMA-Factory/predictions/gsm8k_filtered_rebalanced_feedback_refinement_diff_meta-llama-Llama-2-13b-chat-hf"
# pred_dir = "../LLaMA-Factory/predictions/gsm8k_test_rebalanced_refinement_meta-llama-Llama-2-13b-chat-hf"
# pred_dir = "../LLaMA-Factory/predictions/gsm8k_ft_rationale_meta-llama-Llama-2-7b-chat-hf/"
def get_scores(pred_dir, return_preds=False):
if "jsonl" not in pred_dir:
pred_file = os.path.join(pred_dir, "generated_predictions.jsonl.00")
if not os.path.exists(pred_file):
pred_file = pred_file.replace(".00", "")
if not os.path.exists(pred_file):
pred_file = pred_file + "_greedy"
else:
pred_file = pred_dir
# pred_file = "../LLaMA-Factory/data/gsm8k_test_feedback_meta-llama-Llama-2-13b-chat-hf.json"
with open(pred_file, "r") as reader:
refined_predictions = [clean_cot(json.loads(l)['predict']) for l in reader]
# refined_predictions = [l['instruction'].replace("```", "") + "\nprint(solution())" for l in json.load(reader)]
# print(refined_predictions[:3])
already_correct_file = os.path.join(pred_dir, "already_correct.jsonl")
if os.path.exists(already_correct_file):
with open(already_correct_file, 'r') as reader:
items = [json.loads(l) for l in reader]
already_correct_predictions = [clean_code(item['predict']) for item in items]
already_correct_idx = [item['idx'] for item in items]
# print(f"{len(refined_predictions)} refined_predictions, {len(already_correct_predictions)} already_correct_predictions")
# print(already_correct_idx)
predictions = []
for idx in range(test_example_cnt):
if len(already_correct_idx) and idx == already_correct_idx[0]:
already_correct_idx.pop(0)
predictions.append(already_correct_predictions.pop(0))
else:
predictions.append(refined_predictions.pop(0))
else:
predictions = refined_predictions
# with open("data/gsm8k/test_shuf.jsonl", 'r') as reader:
# refs = [json.loads(l)['answer'].split('#### ')[-1] for l in reader]
# predictions = predictions[:test_example_cnt]
# predictions = [[pred] for pred in predictions]
# refs = refs[:test_example_cnt]
# scores = compute(predictions, refs)
"""
print(predictions)
print(refs)
print(scores)
"""
# scores = [s[0] for s in scores]
if return_preds:
return [parsing_func(pred) for pred in predictions]
else:
assert len(predictions) == len(refs)
scores = [int(is_same(parsing_func(pred), ref)) for pred, ref in zip(predictions, refs)]
correct = len([s for s in scores if s == 1])
accu = correct * 100 / len(scores)
print(pred_file)
print(f"accu", accu)
"""
# 0 for refine (initial solution incorrect), 1 for copy intital solution
pred_when_to_sc = [int(already_correct_message in pred) for pred in predictions]
if "cot_ft_as_input" not in pred_file:
# only support greedy output now
gold_when_to_sc = initial_prompting_solution_scores
else:
raise ValueError("get_scores() and initial_cot_scores reply on each other, should decompose get_scores into two functions")
assert len(pred_when_to_sc) == len(gold_when_to_sc)
sc_contribution = sum([1 for p, c in zip(pred_when_to_sc, scores) if p == 0 and c == 1]) / sum([1 for p in pred_when_to_sc if p == 0])
sc_freq = 1 - sum(pred_when_to_sc) / len(pred_when_to_sc)
print("When to self-correct")
print("initial correctness -> critic (determine when to self-correct) correctness")
print(f"Self-correct Frequency: {100*sc_freq:.2f}%")
print(f"Self-correct Contribution: {100*sc_contribution:.2f}%")
print(classification_report(gold_when_to_sc, pred_when_to_sc))
print(confusion_matrix(gold_when_to_sc, pred_when_to_sc))
print("initial correctness -> refined correctness")
print(confusion_matrix(initial_prompting_solution_scores, scores))
"""
"""
if "feeback_refinement" not in pred_file: # separate critic refiner acrh
gen_feedback_file = pred_file.replace("_refinement", "")
else:
gen_feedback_file = pred_file
with open(gen_feedback_file, 'r') as reader:
gen_feedbacks = [json.loads(l)['predict']) for l in reader]
"""
return scores
def get_sampling_evals(dataset_model, SAMPLE_SIZE=10):
# SAMPLE_SIZE
# return (prediction, confidence, correctness)
if dataset_model in ["dev_rationale_train_rationale", "dev_feedback4_refinement_filtered_feedback4_refinement_diff"]:
sampling_preds = [s for sample_idx in range(SAMPLE_SIZE) for part_idx in range(16) for s in get_scores(f"../LLaMA-Factory/predictions/more_{TASK}_part{part_idx:02}_{dataset_model}_meta-llama-Llama-2-13b-chat-hf/generated_predictions.jsonl.{sample_idx:02}", return_preds=True)]
else:
sampling_preds = [s for i in range(16) for s in get_scores(f"../LLaMA-Factory/predictions/{TASK}_part{i:02}_{dataset_model}_meta-llama-Llama-2-13b-chat-hf/generated_predictions.jsonl.00", return_preds=True)]
assert len(sampling_preds) == len(refs) * SAMPLE_SIZE, f"len(sampling_preds)={len(sampling_preds)}, len(refs)={len(refs)}, SAMPLE_SIZE={SAMPLE_SIZE}"
sampling_preds = [sampling_preds[i:i+len(refs)] for i in range(0, len(sampling_preds), len(refs))]
assert len(sampling_preds) == SAMPLE_SIZE
sampling_preds = [[one_version[data_idx] for one_version in sampling_preds] for data_idx in range(len(refs))]
assert len(sampling_preds) == len(refs)
sampling_evals = [majority_voting(preds, ref)[-1] for preds, ref in zip(sampling_preds, refs)]
return sampling_evals
def get_oracle_scores(initial_scores, refined_scores):
assert len(initial_scores) == len(refined_scores)
oracle_scores = [int(ini or s) for ini, s in zip(initial_scores, refined_scores)]
print(f"oracle refinement accu: {100*sum(oracle_scores)/len(oracle_scores):.2f}")
def get_verifier_scores(original_verifier_scores, initial_scores, refined_scores, confidence=None, detailed_report=False):
if confidence is None:
all_confidence = [round(i, 2) for i in np.linspace(0.5,1,51).tolist()]
else:
all_confidence = [confidence]
assert len(original_verifier_scores) == len(initial_scores)
assert len(initial_scores) == len(refined_scores)
best_final_accu = 0
best_conf = None
for conf in all_confidence:
if type(original_verifier_scores[0]) is int:
verifier_scores = original_verifier_scores
else:
verifier_scores = [0 if v['predict'] == 0 and v['confidence'] >= conf else 1 for v in original_verifier_scores]
system_scores = [ini if v else r for v, ini, r in zip(verifier_scores, initial_scores, refined_scores)]
final_accu_at_conf = 100*sum(system_scores)/len(system_scores)
if final_accu_at_conf > best_final_accu:
best_final_accu = final_accu_at_conf
best_conf = conf
# if detailed_report:
# print(f"ours verifier (confidence @ {conf:.2f}) accu: {final_accu_at_conf:.2f}")
print(f"ours verifier (confidence @ {best_conf:.2f}) accu: {best_final_accu:.2f}")
if not detailed_report:
return
print(f"For Best confidence ({best_conf:.2f})")
if type(original_verifier_scores[0]) is int:
verifier_scores = original_verifier_scores
else:
verifier_scores = [0 if v['predict'] == 0 and v['confidence'] > best_conf else 1 for v in original_verifier_scores]
system_scores = [ini if v else r for v, ini, r in zip(verifier_scores, initial_scores, refined_scores)]
try:
sc_contribution = len([r for v, r in zip(verifier_scores, refined_scores) if v == 0 and r == 1]) / len([v for v in verifier_scores if v == 0])
except:
sc_contribution = 0
sc_freq = len([v for v in verifier_scores if v == 0]) / len(verifier_scores)
print("When to self-correct")
print("initial correctness -> critic (determine when to self-correct) correctness")
print(f"Self-correct Frequency: {100*sc_freq:.2f}%")
print(f"Self-correct Contribution: {100*sc_contribution:.2f}%")
print(classification_report(initial_scores, verifier_scores, digits=3))
print(confusion_matrix(initial_scores, verifier_scores))
print("initial correctness -> refined correctness")
print(confusion_matrix(initial_scores, system_scores))
if stage == "choose_best_rft_checkpoint":
for pred_dir in [f"../LLaMA-Factory/predictions/{TASK}_dev_rationale_train_rationale_meta-llama-Llama-2-13b-chat-hf_step_{ckpt}" for ckpt in ["3309", "6618", "9927"]]:
get_scores(pred_dir)
exit(0)
# prompted_as_input_scores = get_scores(f"../LLaMA-Factory/predictions/{TASK}_dev_feedback_refinement_filtered_rebalanced_feedback_refinement_diff_meta-llama-Llama-2-13b-chat-hf")
# get_oracle_scores(initial_prompting_solution_scores, prompted_as_input_scores)
"""
assert len(initial_prompting_solution_scores) == len(prompted_as_input_scores)
oracle_scores = [int(ini or s) for ini, s in zip(initial_prompting_solution_scores, prompted_as_input_scores)]
print(f"oracle refinement accu: {100*sum(oracle_scores)/len(oracle_scores):.2f}")
"""
if VERIFIER_TYPE == "ours_ft_on_prompted_solutions":
if split == "dev":
verifier_pred_file = f"checkpoint/verifier/{TASK}/{VERIFIER_MODEL}/dev.json"
elif split == "test":
verifier_pred_file = f"checkpoint/verifier/{TASK}/{VERIFIER_MODEL}/test.json"
elif VERIFIER_TYPE == "prompted":
verifier_pred_file = f"checkpoint/verifier/{TASK}/cot_ft_train_rationale_{INPUT_MODEL}_verified_by_{VERIFIER_MODEL}/generated_predictions_test.jsonl_greedy"
if INPUT == "prompted":
verifier_pred_file = verifier_pred_file.replace("cot_ft_train_rationale_", "")
if split == "dev":
verifier_pred_file = verifier_pred_file.replace("_test", "")
if "-sc" in VERIFIER_MODEL:
verifier_pred_file = verifier_pred_file.replace("_greedy", "").replace("-sc", "")
elif VERIFIER_TYPE == "ours_ft_on_cot_ft_solutions":
if split == "dev":
verifier_pred_file = f"checkpoint/verifier/{TASK}/cot_ft_train_rationale_{VERIFIER_MODEL}/generated_predictions.jsonl"
elif split == "test":
verifier_pred_file = f"predictions/verifier/{TASK}/cot_ft_train_rationale_{VERIFIER_MODEL}/test_cot_ft.json"
elif VERIFIER_TYPE == "ours_ft_cot_init_on_prompted_solutions":
if split == "dev":
verifier_pred_file = f"checkpoint/verifier/{TASK}/train_rationale_{VERIFIER_MODEL}/generated_predictions.jsonl"
elif split == "test":
verifier_pred_file = f"predictions/verifier/{TASK}/{VERIFIER_MODEL}/test.json"
elif VERIFIER_TYPE == "ours_ft_base_init_on_cot_ft_solutions":
if split == "dev":
verifier_pred_file = f"checkpoint/verifier/{TASK}/cot_ft_{VERIFIER_MODEL}/generated_predictions.jsonl"
elif split == "test":
assert 0, "only for dev ablation"
elif VERIFIER_TYPE in ["ours_ft_on_base_rft_solutions", "ours_ft_base_init_on_base_rft_solutions"]:
if INPUT == "prompted":
verifier_pred_file = f"predictions/verifier/{TASK}/base_rft_train_rationale_{VERIFIER_MODEL}/{split}.json"
elif INPUT == "cot_ft":
verifier_pred_file = f"predictions/verifier/{TASK}/base_rft_train_rationale_{VERIFIER_MODEL}/{split}_cot_ft.json"
if "base_init" in VERIFIER_TYPE:
verifier_pred_file = verifier_pred_file.replace("_train_rationale", "")
else:
raise ValueError(f"invalid VERIFIER_TYPE: {VERIFIER_TYPE}")
# verifier_pred_file = f"checkpoint/verifier/{TASK}/cot_ft_train_rationale_meta-llama-Llama-2-13b-chat-hf_verified_by_gpt-4/generated_predictions.jsonl_greedy"
with open(verifier_pred_file, 'r') as reader:
initial_prompting_verifier_scores = [json.loads(l) for l in reader]
assert len(initial_prompting_verifier_scores) == len(initial_prompting_solution_scores)
if INPUT == "prompted":
initial_scores = initial_prompting_solution_scores
print(f"{split} initial prompting solution accu:", sum([s for s in initial_prompting_solution_scores if s == 1]) / len(initial_prompting_solution_scores))
elif INPUT == "cot_ft":
initial_cot_indir = f"../LLaMA-Factory/predictions/{TASK}_{split}_rationale_train_rationale_{INPUT_MODEL}/"
initial_cot_scores = get_scores(initial_cot_indir)
initial_scores = initial_cot_scores
print(f"{split} initial cot-finetuned solution accu:", sum([s for s in initial_cot_scores if s == 1]) / len(initial_cot_scores))
else:
raise ValueError(f"invalid INPUT: {INPUT}")
"""
print('='*30, "feedback4_separate_refine")
feedback4_separate_refine_scores = get_scores(f"../LLaMA-Factory/predictions/{TASK}_dev_feedback4_filtered_feedback4_diff_refinement_meta-llama-Llama-2-13b-chat-hf")
get_oracle_scores(initial_scores, feedback4_separate_refine_scores)
get_verifier_scores(initial_prompting_verifier_scores, initial_scores, feedback4_separate_refine_scores, detailed_report=False)
"""
if stage == "eval_greedy_refine":
print('='*30, "feedback4_combined_critic_refine")
if REFINER_TYPE == "prompted":
refine_infile = f"predictions/{TASK}_{split}_feedback4_refinement_1samples_self_refine_{REFINER_MODEL}/generated_predictions_test.jsonl_greedy"
else:
if INPUT == "prompted":
refine_infile = f"../LLaMA-Factory/predictions/more_{TASK}_{split}_feedback4_refinement_1samples_{REFINER_TYPE}_{REFINER_MODEL}"
elif INPUT == "cot_ft":
refine_infile = f"../LLaMA-Factory/predictions/more_{TASK}_{split}_cot_ft_as_input_feedback4_refinement_1samples_{REFINER_TYPE}_{REFINER_MODEL}"
if refiner_cot_init:
refine_infile = refine_infile.replace("more", "more_cot_init")
refine_infile = refine_infile.replace("more", refiner_ft_data_size)
if not os.path.exists(refine_infile):
refine_infile = refine_infile.replace("_1samples", "")
feedback4_combined_critic_refine_scores = get_scores(refine_infile)
# f"../LLaMA-Factory/predictions/more_{TASK}_dev_feedback4_refinement_1samples_filtered_feedback4_refinement_base_rft_diff_meta-llama-Llama-2-13b-chat-hf")
get_oracle_scores(initial_scores, feedback4_combined_critic_refine_scores)
get_verifier_scores(initial_prompting_verifier_scores, initial_scores, feedback4_combined_critic_refine_scores, detailed_report=VERBOSE, confidence=best_dev_confidence)
print('='*30, "oracle verifier")
get_verifier_scores(initial_scores, initial_scores, feedback4_combined_critic_refine_scores, detailed_report=VERBOSE, confidence=1.0)
exit(0)
"""
print('='*30, "feedback4_combined_critic_refine_rebalanced_0.13")
feedback4_combined_critic_refine_rebalanced_13_scores = get_scores(f"../LLaMA-Factory/predictions/{TASK}_dev_feedback_refinement_filtered_rebalanced_feedback_refinement_diff_meta-llama-Llama-2-13b-chat-hf")
get_oracle_scores(initial_prompting_solution_scores, feedback4_combined_critic_refine_rebalanced_13_scores)
print('='*30, "feedback3_separate_rebalanced_0.13")
feedback3_separate_rebalanced_13_scores = get_scores(f"../LLaMA-Factory/predictions/{TASK}_dev_feedback3_filtered_rebalanced_feedback3_0.13_diff_refinement_meta-llama-Llama-2-13b-chat-hf")
get_oracle_scores(initial_prompting_solution_scores, feedback3_separate_rebalanced_13_scores)
"""
"""
cot_as_input_scores = get_scores("../LLaMA-Factory/predictions/{TASK}_dev_cot_ft_as_input_feedback_refinement_filtered_rebalanced_feedback_refinement_diff_meta-llama-Llama-2-13b-chat-hf")
initial_cot_scores = get_scores(f"../LLaMA-Factory/predictions/{TASK}_dev_rationale_meta-llama-Llama-2-13b-chat-hf_rank32_lr5e-5_2337")
get_oracle_scores(initial_cot_scores, cot_as_input_scores)
"""
"""
# self-consistency
sampling_cot_as_input_refined_evals = get_evals("dev_cot_ft_as_input_feedback_refinement_filtered_feedback_refinement_diff")
sampling_cot_as_input_initial_evals = get_evals("dev_rationale_train_rationale")
initial_cot_sc_accu = sum([correctness for _, _, correctness in sampling_cot_as_input_initial_evals]) / len(refs)
print(f"initial_cot_sc_accu: {100*initial_cot_sc_accu:.2f}")
for confidence_threshold in [0.0, 0.1, 0.2, 0.3, 0.4, 0.5, 0.6, 0.7, 0.8, 0.9, 1.0]:
refined_cot_sc_accu = sum([init_correctness if init_confidence >= confidence_threshold else refined_correctness for (_, init_confidence, init_correctness), (_, _, refined_correctness) in zip(sampling_cot_as_input_initial_evals, sampling_cot_as_input_refined_evals)]) / len(refs)
print(f"refined_cot_sc_accu @ {confidence_threshold}: {100*refined_cot_sc_accu:.2f}")
oracle_refined_cot_sc_accu = sum([int(init_correctness or refined_correctness) for (_, init_confidence, init_correctness), (_, _, refined_correctness) in zip(sampling_cot_as_input_initial_evals, sampling_cot_as_input_refined_evals)]) / len(refs)
print(f"oracle_refined_cot_sc_accu: {100*oracle_refined_cot_sc_accu:.2f}")
"""
if stage == "eval_sampling_refine":
print("dev initial prompting sampling@10 solution accu:", sum([s for s in initial_prompting_sampling_solution_scores if s == 1]) / len(initial_prompting_sampling_solution_scores))
print('='*30, "feedback4_combined_critic_refine, Q->1*A->10*A'")
sampling_refinement_evals = get_sampling_evals("dev_feedback4_refinement_filtered_feedback4_refinement_diff")
get_oracle_scores(initial_scores, sampling_refinement_evals)
get_verifier_scores(initial_prompting_verifier_scores, initial_scores, sampling_refinement_evals, detailed_report=False)
"""
for dtst in ["dev_feedback4"]:
for ratio in [0.11, 0.12, 0.13, 0.14]:
get_scores(f"../LLaMA-Factory/predictions/{TASK}_{dtst}_filtered_rebalanced_feedback4_{ratio}_diff_refinement_meta-llama-Llama-2-13b-chat-hf")
"""