Minkow/TipCoder-Data

TipCoder Training Data

Final SFT and RL training data for TipCoder: Reinforcement Learning Boosted Test-time Instruction Proposer for Code Generation, Findings of EMNLP 2026.

Code | RL proposer

Contents

Configuration	File	Records	Bytes	Serialization
sft	sft/train.json	2,126	5,402,331	JSON array
rl	rl/train.json	6,410	113,146,964	JSON Lines

These are the final files identified in the SFT and RL training job arguments. Only their published paths have changed. All source bytes, record ordering and fields are preserved; no samples have been removed or rewritten for release. The RL file retains its original .json suffix despite containing JSON Lines.

Fields

SFT: id, source, prompt, output, pass_rate. Each record pairs a proposer input with its supervised target. The stored pass-rate field is retained as training-data metadata, not a benchmark result for this release.

RL: id, source, question, test_cases, inferences, context_messages. Records include problem statements, executable tests, reference code-generation samples with their metadata, and conversational problem inputs. These support the execution-reward workflow described in the code repository.

Download and Read

hf download Minkow/TipCoder-Data --repo-type dataset --local-dir data/TipCoder

The two configurations can be loaded separately:

from datasets import load_dataset

sft = load_dataset("Minkow/TipCoder-Data", "sft", split="train")
rl = load_dataset("Minkow/TipCoder-Data", "rl", split="train")

For direct access without schema conversion:

import json
from pathlib import Path

root = Path("data/TipCoder")
sft = json.loads((root / "sft/train.json").read_text())
with (root / "rl/train.json").open() as stream:
    rl = [json.loads(line) for line in stream if line.strip()]

Source and License

The data derives from TIGER-Lab/AceCode-87K and the TipCoder data-construction pipeline. The release uses MIT, following AceCode, and retains the upstream copyright notice in LICENSE. Please credit both TipCoder and AceCoder when using the data.

Limitations

The files contain model-generated text and code, including illustrative local paths and test fixtures. Generated content may be incorrect. Do not execute code or tests outside an isolated environment without credentials or sensitive mounts. This dataset release does not include a new training or evaluation run.

Citation

@inproceedings{chen-etal-2026-tipcoder,
  title = "{T}ip{C}oder: Reinforcement Learning Boosted Test-time Instruction Proposer for Code Generation",
  author = "Chen, Minyu and Wu, Sihao and Wu, Ling-I and Qin, Song and Li, Jingyang and Ning, Lei and Xue, Jianxin and Li, Guoqiang",
  booktitle = "Findings of the Association for Computational Linguistics: EMNLP 2026",
  year = "2026",
  pages = "10363--10377",
  url = "https://aclanthology.org/2026.findings-emnlp.495/"
}