acecode87k / preprocess_acecode_data.py
guanning's picture
AceCode-87K train / held-out 1k test split used for the Qwen3-1.7B-Base code RL runs
de32d94 verified
Raw History Blame Contribute Delete
5.27 kB
# Copyright 2024 Bytedance Ltd. and/or its affiliates
#
# Licensed under the Apache License, Version 2.0 (the "License");
# you may not use this file except in compliance with the License.
# You may obtain a copy of the License at
#
# http://www.apache.org/licenses/LICENSE-2.0
#
# Unless required by applicable law or agreed to in writing, software
# distributed under the License is distributed on an "AS IS" BASIS,
# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
# See the License for the specific language governing permissions and
# limitations under the License.
"""Preprocess TIGER-Lab/AceCode-87K into verl parquet for the binary code-RL recipe.
Writes the same schema as scripts/preprocess_knights_and_knaves_data.py:
data_source | prompt (chat, single user turn) | ability | reward_model | extra_info
* data_source : "acecode" for train, "acecode_eval" for the held-out split
(both route to verl/utils/reward_score/acecode.py, binary all-tests-pass)
* prompt : the AceCode question + a fixed suffix asking for ONE ```python block
* ground_truth: JSON {"tests": [...]} straight from the dataset's test list
* split : the LAST --eval-size rows of a seeded shuffle are eval, the rest train
Rows are dropped when the question or the tests are unusable (empty / non-assert / too
few tests) or when the prompt would exceed --max-prompt-tokens under the model's chat
template (data.filter_overlong_prompts would drop them at train time anyway; doing it
here keeps the parquet row count honest).
Usage (login node has internet; needs the `maxrl` conda env):
python scripts/preprocess_acecode_data.py \
--out-dir /n/netscratch/qianqian_lab/Lab/gzeng/RL-ASSETS/acecode_data/acecode87k \
--model-path $HOME/.cache/maxrl/models/Qwen3-1.7B-Base
"""
from __future__ import annotations
import argparse
import json
import os
PROMPT_SUFFIX = (
"\nPlease reason step by step, then write a self-contained Python solution. "
"Give your final program in a single ```python code block."
)
MIN_TESTS = 3 # AceCoder filtering keeps prompts with enough signal for binary reward
def build_row(question: str, tests: list[str], data_source: str, split: str, index: int, src_id) -> dict:
return {
"data_source": data_source,
"prompt": [{"role": "user", "content": question + PROMPT_SUFFIX}],
"ability": "code",
"reward_model": {"style": "rule", "ground_truth": json.dumps({"tests": tests})},
"extra_info": {"split": split, "index": index, "acecode_id": str(src_id), "num_tests": len(tests)},
}
def main() -> None:
ap = argparse.ArgumentParser()
ap.add_argument("--dataset", default="TIGER-Lab/AceCode-87K")
ap.add_argument("--out-dir", required=True)
ap.add_argument("--eval-size", type=int, default=1000)
ap.add_argument("--seed", type=int, default=42)
ap.add_argument("--model-path", required=True, help="tokenizer for the prompt-length filter")
ap.add_argument("--max-prompt-tokens", type=int, default=3072)
ap.add_argument("--max-tests", type=int, default=16, help="cap tests per prompt (reward caps at 32)")
args = ap.parse_args()
import pandas as pd
from datasets import load_dataset
from transformers import AutoTokenizer
tokenizer = AutoTokenizer.from_pretrained(args.model_path)
ds = load_dataset(args.dataset, split="train")
ds = ds.shuffle(seed=args.seed)
kept, dropped_tests, dropped_long = [], 0, 0
seen_questions: set[int] = set()
for row in ds:
question = (row.get("question") or "").strip()
tests = [t.strip() for t in (row.get("test_cases") or []) if isinstance(t, str) and t.strip()]
tests = [t for t in tests if t.startswith("assert")]
if not question or len(tests) < MIN_TESTS:
dropped_tests += 1
continue
qh = hash(question)
if qh in seen_questions:
dropped_tests += 1
continue
seen_questions.add(qh)
n_tokens = len(
tokenizer.apply_chat_template(
[{"role": "user", "content": question + PROMPT_SUFFIX}],
add_generation_prompt=True,
tokenize=True,
)
)
if n_tokens > args.max_prompt_tokens:
dropped_long += 1
continue
kept.append((question, tests[: args.max_tests], row.get("id")))
assert len(kept) > args.eval_size, f"only {len(kept)} usable rows"
eval_rows = kept[-args.eval_size :]
train_rows = kept[: -args.eval_size]
os.makedirs(args.out_dir, exist_ok=True)
train_df = pd.DataFrame(
[build_row(q, t, "acecode", "train", i, sid) for i, (q, t, sid) in enumerate(train_rows)]
)
eval_df = pd.DataFrame(
[build_row(q, t, "acecode_eval", "eval", i, sid) for i, (q, t, sid) in enumerate(eval_rows)]
)
train_df.to_parquet(os.path.join(args.out_dir, "train.parquet"))
eval_df.to_parquet(os.path.join(args.out_dir, "eval.parquet"))
print(
f"wrote {len(train_df)} train / {len(eval_df)} eval to {args.out_dir} "
f"(dropped: {dropped_tests} bad-tests/dup, {dropped_long} overlong)"
)
if __name__ == "__main__":
main()