microsoft--unilm
wehub-resource-sync
120 行
5.2 KiB
Python
120 行
5.2 KiB
Python
import json
|
|
from collections import defaultdict, Counter
|
|
import argparse
|
|
import os
|
|
import sys
|
|
from glob import glob
|
|
import copy
|
|
from tqdm import tqdm
|
|
from datasets import load_dataset
|
|
import random
|
|
|
|
sys.set_int_max_str_digits(0)
|
|
sys.path.append(os.path.dirname(os.path.dirname(os.path.dirname(os.path.abspath(__file__)))))
|
|
|
|
from data.apps import APPsWithFunctionName
|
|
|
|
|
|
def main():
|
|
"""
|
|
This script takes the completions from GPT4, the corresponding inputs for worsening code, and the original dataset, to construct a new combined dataset,
|
|
which contains the solutions from the policy model, as well as the synthesized negative codes for each plausible solution.
|
|
:return:
|
|
"""
|
|
parser = argparse.ArgumentParser()
|
|
parser.add_argument("--worsen_file", type=str,
|
|
help="The file contains completion from the teacher model for worsening code"
|
|
"The inputs for this file are generated by `pp_worsen_inputs.py` script.")
|
|
parser.add_argument("--completion_file", type=str,
|
|
help="The file contains the completion for each query.")
|
|
# parser.add_argument("--completion_response_field", type=str, default="completion")
|
|
parser.add_argument("--completion_problem_id_field", type=str, default="problem_id")
|
|
parser.add_argument("--output_file", type=str)
|
|
args = parser.parse_args()
|
|
|
|
worsen_codes = [json.loads(line) for line in open(args.worsen_file).readlines()]
|
|
print(f"Total number of worsen codes: {len(worsen_codes)}")
|
|
|
|
if os.path.exists(args.completion_file):
|
|
data = json.load(open(args.completion_file))
|
|
else:
|
|
data = []
|
|
for file in glob(args.completion_file):
|
|
data += json.load(open(file))
|
|
|
|
p_id2response = {item[args.completion_problem_id_field]: item for item in data}
|
|
|
|
p_id2pairs = defaultdict(list)
|
|
for item in worsen_codes:
|
|
if item["completion"]:
|
|
try:
|
|
completion = json.loads(item["completion"])
|
|
except:
|
|
print(f"Json parsing error: {item['completion']}")
|
|
continue
|
|
neg_code = completion["incorrect_program"]
|
|
# if isinstance(neg_code, dict):
|
|
# print(json.dumps(completion, indent=2))
|
|
|
|
if not isinstance(neg_code, str):
|
|
print(f"Bad format: {neg_code}.")
|
|
continue
|
|
|
|
if not neg_code.strip():
|
|
continue
|
|
|
|
p_id, pred_id = item["id"].split("_")
|
|
pred_id = int(pred_id[3:])
|
|
p_id = int(p_id)
|
|
|
|
response = p_id2response[p_id]
|
|
pred = response["pred"][pred_id]
|
|
assert pred in item["prompt"]
|
|
|
|
p_id2pairs[p_id].append((pred, neg_code))
|
|
|
|
outputs = []
|
|
num_pairs = 0
|
|
for p_id, pairs in p_id2pairs.items():
|
|
pos = []
|
|
neg = []
|
|
for pred, neg_code in pairs:
|
|
pos.append(pred)
|
|
neg.append(neg_code)
|
|
assert len(pos) == len(neg)
|
|
num_pairs += len(pos)
|
|
outputs.append({
|
|
"problem_id": p_id,
|
|
"pos": pos,
|
|
"neg": neg,
|
|
})
|
|
|
|
print(f"Total number of outputs: {len(outputs)}")
|
|
print(f"Total number of pairs: {num_pairs}")
|
|
json.dump(outputs, open(args.output_file, "w"), indent=2, ensure_ascii=False)
|
|
|
|
|
|
if __name__ == '__main__':
|
|
main()
|
|
|
|
"""
|
|
>>> python scripts/apps/worsen_gpt4_combine.py --worsen_file outputs/apps/critique/r2c.sft.train.0shot.tem1.0.n10.v1.1.worsen_4o_critic.s42.f2000.gpt4o.tem1.0.s42.n1.json_obj.jsonl ]
|
|
--completion_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.v1.1.s43.json \
|
|
--completion_problem_id_field id --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.v1.1.s43.gpt4o.worsen.f2000.json
|
|
|
|
>>> python scripts/apps/worsen_gpt4_combine.py \
|
|
--worsen_file outputs/apps/critique/r2c.sft.train.0shot.tem1.0.n10.v1.1.worsen_4o_critic.s42.f100k.gpt4o.tem1.0.s42.n1.json_obj.jsonl \
|
|
--completion_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.?-of-8.v1.1.json" \
|
|
--completion_problem_id_field id \
|
|
--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.v1.1.s43.gpt4o.worsen.f100k.json
|
|
>>> The above code is incorrect.
|
|
|
|
FIXME:
|
|
>>> python scripts/apps/worsen_gpt4_combine.py \
|
|
--worsen_file outputs/apps/critique/r2c.sft.train.0shot.tem1.0.n10.v1.1.worsen_4o_critic.s42.f100k.gpt4o.tem1.0.s42.n1.json_obj.jsonl \
|
|
--completion_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.v1.1.s43.json" \
|
|
--completion_problem_id_field id \
|
|
--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.v1.1.s43.gpt4o.worsen.f100k.fix0708.json
|
|
|
|
"""
|