wehub-resource-sync
142 lines
6.4 KiB
Python
142 lines
6.4 KiB
Python
import argparse
|
|
import json
|
|
import os.path
|
|
import sys
|
|
from glob import glob
|
|
from tqdm import tqdm
|
|
import collections
|
|
|
|
sys.set_int_max_str_digits(0)
|
|
|
|
"""
|
|
Soft version of constructing preference pair. This would be useful for teacher-generated pseudo test cases.
|
|
"""
|
|
|
|
|
|
def main():
|
|
parser = argparse.ArgumentParser()
|
|
parser.add_argument("--input_file", type=str, required=True)
|
|
parser.add_argument("--output_file", type=str, required=True)
|
|
parser.add_argument("--response_field", type=str, default="response")
|
|
parser.add_argument("--test_case_field", type=str, default="input_output")
|
|
parser.add_argument("--pass_case_margin", type=float, default=1)
|
|
parser.add_argument("--pass_case_lower_bound", type=float, default=0.5)
|
|
args = parser.parse_args()
|
|
|
|
cnt = 0
|
|
|
|
if os.path.exists(args.input_file):
|
|
if args.input_file.endswith(".json"):
|
|
data = json.load(open(args.input_file))
|
|
else:
|
|
data = [json.loads(line) for line in open(args.input_file).readlines()]
|
|
else:
|
|
data = []
|
|
for file in glob(args.input_file):
|
|
print(file)
|
|
if file.endswith(".json"):
|
|
data += json.load(open(file))
|
|
else:
|
|
data += [json.loads(line) for line in open(file).readlines()]
|
|
|
|
if len(data) == 0:
|
|
raise ValueError(f"No data found in {args.input_file}")
|
|
|
|
print(len(data))
|
|
pass_cnt = collections.Counter()
|
|
for item in tqdm(data):
|
|
pos = []
|
|
neg = []
|
|
pos_code = []
|
|
neg_code = []
|
|
if isinstance(item[args.response_field], str): # We cannot make pairs if there is only one response.
|
|
item["pos"] = []
|
|
item["pos_code"] = []
|
|
item["neg"] = []
|
|
item["neg_code"] = []
|
|
continue
|
|
|
|
if len(item[args.test_case_field]["inputs"]) == 0:
|
|
item["pos"] = []
|
|
item["pos_code"] = []
|
|
item["neg"] = []
|
|
item["neg_code"] = []
|
|
continue
|
|
|
|
if "res" in item and "full_res" in item and item[args.test_case_field]: # If there is no test-cases, we cannot determine the correctness
|
|
assert len(item["res"]) == len(item["full_res"]) == len(item[args.response_field]), (len(item["res"]),
|
|
len(item["full_res"]),
|
|
len(item[args.response_field]),
|
|
item[args.response_field])
|
|
|
|
pred_pass_cnt = []
|
|
for pg_i, pg_res in enumerate(item["full_res"]):
|
|
pred_pass_cnt.append(sum([1 for r in pg_res if r == 1]))
|
|
pass_cnt[pred_pass_cnt[-1]] += 1
|
|
num_test_cases = len(item[args.test_case_field]["inputs"])
|
|
|
|
for i in range(len(pred_pass_cnt)):
|
|
resp_i = item["response"][i]
|
|
prog_i = item["pred"][i]
|
|
pass_cnt_i = pred_pass_cnt[i]
|
|
if pass_cnt_i / num_test_cases < args.pass_case_lower_bound:
|
|
continue
|
|
for j in range(len(pred_pass_cnt)):
|
|
if i == j:
|
|
continue
|
|
resp_j = item["response"][j]
|
|
prog_j = item["pred"][j]
|
|
pass_cnt_j = pred_pass_cnt[j]
|
|
if pass_cnt_i - pass_cnt_j >= args.pass_case_margin:
|
|
pos.append(resp_i)
|
|
pos_code.append(prog_i)
|
|
neg.append(resp_j)
|
|
neg_code.append(prog_j)
|
|
|
|
item["pos"] = pos
|
|
item["neg"] = neg
|
|
item["pos_code"] = pos_code
|
|
item["neg_code"] = neg_code
|
|
cnt += len(pos)
|
|
|
|
if args.response_field != "response":
|
|
item["response"] = item.pop(args.response_field)
|
|
if args.test_case_field != "input_output":
|
|
item["input_output"] = item.pop(args.test_case_field)
|
|
|
|
json.dump(data, open(args.output_file, "w"), ensure_ascii=False)
|
|
print(len(data), cnt, cnt / len(data))
|
|
|
|
|
|
if __name__ == '__main__':
|
|
main()
|
|
|
|
"""
|
|
>>> python scripts/apps/construct_prefer_pair.py --input_file ../msranlpintern/share/models/deepseek-coder-7b-instruct-v1.5/apps/train.0shot.tem1.0.n10.v1.0.pseudo_test_case.exec.sc.json \
|
|
--output_file ../msranlpintern/share/models/deepseek-coder-7b-instruct-v1.5/apps/train.0shot.tem1.0.n10.v1.0.pseudo_test_case.exec.sc.dpo_v1.0.json --test_case_field test_cases
|
|
|
|
4223
|
|
100%|████████████████████████████████████████████████████████████████████████████████████| 4223/4223 [00:00<00:00, 148189.90it/s]
|
|
4223 18383 4.353066540374142
|
|
|
|
>>> python scripts/apps/construct_prefer_pair.py --input_file ../msranlpintern/reward_modeling/experiments/deepseek-coder-v1.5-ins.7b.apps.r2c.gpt4o.distil.A100.w8.v3.0.s42/apps/checkpoint-400/train.0shot.tem1.0.n10.self_s43_pseudo_cases.exec.json \
|
|
--output_file ../msranlpintern/reward_modeling/experiments/deepseek-coder-v1.5-ins.7b.apps.r2c.gpt4o.distil.A100.w8.v3.0.s42/apps/checkpoint-400/train.0shot.tem1.0.n10.self_s43_pseudo_cases.exec_dpo.json --test_case_field test_cases
|
|
|
|
4715
|
|
100%|██████████████████████████████████████████████████████████| 4715/4715 [00:00<00:00, 107036.93it/s]
|
|
4715 15921 3.3766702014846235
|
|
|
|
>>> python ~/gpt-chat-examples/scripts/apps/construct_prefer_pair.py --input_file "train.0shot.tem1.0.n10.?-of-8.v2.0.json" \
|
|
--output_file "train.0shot.tem1.0.n10.v2.0.dpo_v1.0.json" --test_case_field test_cases
|
|
|
|
4500
|
|
100%|██████████████████████| 4500/4500 [00:00<00:00, 142921.59it/s]
|
|
4500 42550 9.455555555555556
|
|
|
|
>>> python scripts/apps/construct_prefer_pair.py \
|
|
--input_file "../msranlpintern/reward_modeling/experiments/deepseek-coder-v1.5-ins.7b.apps.r2c.gpt4o.distil.V100.w8.v3.1.dp4.tp4.s42/apps/checkpoint-200/train.0shot.tem1.0.n10.?-of-4.v2.0.json" \
|
|
--output_file ../msranlpintern/reward_modeling/experiments/deepseek-coder-v1.5-ins.7b.apps.r2c.gpt4o.distil.V100.w8.v3.1.dp4.tp4.s42/apps/checkpoint-200/train.0shot.tem1.0.n10.dpo_v1.0.json \
|
|
--test_case_field test_cases
|
|
|
|
"""
|