🂡 AceCode-87K

Paper | Github | AceCode-87K | AceCodePair-300K | RM/RL Models

We introduce AceCoder, the first work to propose a fully automated pipeline for synthesizing large-scale reliable tests used for the reward model training and reinforcement learning in the coding scenario. To do this, we curated the dataset AceCode-87K, where we start from a seed code dataset and prompt powerful LLMs to "imagine" proper test cases for the coding question and filter the noisy ones. We sample inferences from existing coder models and compute their pass rate as the reliable and verifiable rewards for both training the reward model and conducting the reinforcement learning for coder LLM.

• This dataset is the official AceCodeRMPair-300K by constructing valid pairs using the following rule from the TIGER-Lab/AceCode-87K
  • The chosen program should have at least 0.8 pass rate
  • The pass rate difference between the chosen and rejected program should be at least 0.4
  • The rejected program should have a pass rate larger than 0
• This dataset is used to train our reward models:
  • TIGER-Lab/AceCodeRM-7B (Use the "default" subset)
  • TIGER-Lab/AceCodeRM-32B (Use the "32B" subset)

Usage

• Direct use

import datasets
dataset = datasets.load_dataset("TIGER-Lab/AceCodePair-300K", split='train')

• Use for RM training: This dataset can be directly used for RM trainig with LlamaFactory code, where you should set the context_messages as the key. Please refer to our Github Code for details. The "default" subset is used for training the 7B Qwen Coder Instruct 2.5, whereas the "32B" subset is used to train the 32B Qwen Coder Instruct 2.5.

Citation

@article{AceCoder,
    title={AceCoder: Acing Coder RL via Automated Test-Case Synthesis},
    author={Zeng, Huaye and Jiang, Dongfu and Wang, Haozhe and Nie, Ping and Chen, Xiaotong and Chen, Wenhu},
    journal={ArXiv},
    year={2025},
    volume={abs/2207.01780}
}