EvolveCoder-22K

# EvolveCoder-22K ## Overview **EvolveCoder-22K** is a large-scale coding dataset for reinforcement learning, built using the [EvolveCoder](https://arxiv.org/abs/2603.12698) framework. It contains 21,600 competitive programming problems paired with adversarially evolved test cases that are designed to be harder and more discriminative than standard test suites. Each problem comes with a set of test cases generated through multiple rounds of adversarial verification — candidate solutions are run against evolving test suites, and new tests are synthesized to expose weaknesses in incorrect solutions. <p align="center"> <img src="main_graph.png" width="75%"> </p> ## Data Fields | Field | Type | Description | |-------|------|-------------| | `id` | `string` | Unique problem identifier (e.g., `TACO_1`) | | `question` | `string` | Full problem statement | | `tests` | `list[string]` | Evolved test cases for RL reward computation | ## Usage ```python from datasets import load_dataset dataset = load_dataset("TIGER-Lab/EvolveCoder-22K") print(dataset["train"][0]) ``` ## Data Collection The dataset was built using a multi-step pipeline: 1. **Initial generation** — Solve problems using a frontier API model 2. **Solution sampling** — Sample 16 candidate solutions per problem using an open-source model 3. **Iterative test evolution** — Over multiple rounds, generate adversarial test cases, filter by discriminativeness, and re-evaluate solutions 4. **Final filtering** — Keep only (problem, test set) pairs where the evolved tests meaningfully distinguish correct from incorrect solutions ## Citation ```bibtex @article{ruan2026evolvecoder, title={EvolveCoder: Evolving Test Cases via Adversarial Verification for Code Reinforcement Learning}, author={Ruan, Chi and Jiang, Dongfu and Zeng, Huaye and Nie, Ping and Chen, Wenhu}, journal={arXiv preprint arXiv:2603.12698}, year={2026} } ```