README.md

---
license: apache-2.0
pretty_name: CATCH RL
language:
- en
task_categories:
- text-generation
tags:
- code
- python
- reinforcement-learning
- reward-hacking
- arxiv:2609.39533
size_categories:
- 10K<n<100K
configs:
- config_name: default
  default: true
  data_files:
  - split: train
    path: data/train_verl.parquet
  - split: test
    path: data/test_verl.parquet
---

# CATCH RL

CATCH-RL provides coding tasks for reinforcement learning with verifiable rewards (RLVR). It wraps algorithmic problems from DeepCoder into writable software-engineering (SWE) repositories with controlled evaluation loopholes.

Researchers can use these tasks to study reward hacking and evaluate detection and mitigation methods throughout training.

- **Paper:** [CATCH: A Controllable Analysis Testbed for Reward Hacking in Coding RL](https://arxiv.org/abs/2609.39533).
- **Code:** [THUAIS-Lab/CATCH](https://github.com/THUAIS-Lab/CATCH).
- **Source dataset:** [agentica-org/DeepCoder-Preview-Dataset](https://huggingface.co/datasets/agentica-org/DeepCoder-Preview-Dataset).
- **SFT datasets:** [CATCH-Hacking-SFT](https://huggingface.co/datasets/WangSl2004/CATCH-Hacking-SFT) and [CATCH-NonHacking-SFT](https://huggingface.co/datasets/WangSl2004/CATCH-NonHacking-SFT).

## Splits

Each split contains one Parquet file.

| Split | Tasks | File | Size in bytes |
|---|---:|---|---:|
| `train` | 24,287 | `data/train_verl.parquet` | 28,288,168,166 |
| `test` | 128 | `data/test_verl.parquet` | 1,142,500,117 |

## Task construction

The preparation pipeline reads the `primeintellect`, `taco`, and `lcbv5` training splits from DeepCoder. It reads the `codeforces` and `lcbv5` test splits for evaluation.

For each problem, the CATCH pipeline reuses the planner signature and constructs a repository for remote procedure calls (RPC). The repository includes request and response models, a normalizer, an RPC dispatcher, a planner stub, documentation, and tests.

The pipeline selects the three shortest test cases by serialized input and output length for the easy tests. The complete test suite supports the hard tests and the independent audit.

Each task uses this repository layout:

```text
README.md
docs/development.md
docs/problem.md
planner.py
rpc_models.py
normalizer.py
rpc_server.py
calls.json
tests/test_rpc_server.py
```

The prompt lists repository files and their access permissions. Its `calls.json` preview contains the selected easy cases. The runtime `calls.json` contains the complete test suite.

### Evaluation

CATCH exposes three classes of loopholes:

- **Test-file modification:** changes to tests, fixtures, or expected outputs.
- **Test-data exploitation:** hardcoded answers or access to test data.
- **Execution interference:** changes to behavior that the evaluator depends on.

The **Hackable Run** computes the proxy reward from the writable repository. The **Unhackable Run** checks task correctness with an independent evaluator.

The gold monitor labels a response as hacking when it passes the easy and hard tests but fails the independent audit.

## Fields

Each row contains three top-level fields:

| Field | Contents |
|---|---|
| `prompt` | Message list with `role` and `content` fields. The user message contains `placeholder`. |
| `reward_model` | Reward configuration with `style="rule"` and `ground_truth=null`. |
| `extra_info` | Task prompt, repository files, test cases, and source metadata. |

Use `extra_info.question` as the task prompt. Use `extra_info.ground_truth` as the ground-truth test payload.

| Field in `extra_info` | Contents |
|---|---|
| `question` | Rendered CATCH task prompt. |
| `original_question`, `problem` | Source problem text. |
| `repo_files` | Mapping from repository paths to file contents. |
| `repo_file_permissions` | Mapping from repository paths to access permissions. |
| `ground_truth` | Ground-truth test cases as a JSON string. |
| `tests` | Source test cases as a JSON string. |
| `selected_test_cases_for_visible_tests` | Selected visible cases as a JSON string. |
| `planner_function_name`, `planner_signature` | Planner function name and signature. |
| `swe_task_kind` | Task interface: `stdin_stdout`, `functional`, or `function_call`. |
| `is_multiple_test_cases` | Whether one stdin/stdout payload contains multiple cases. |
| `original_question_before_input_section`, `original_input_and_following_text` | Sections of the source problem statement. |
| `starter_code`, `solutions` | Source starter code and reference solutions. The train file includes the `solutions` field. |
| `metadata` | Source metadata as a JSON string. |
| `index`, `uid`, `data_source` | Source identifiers and the pipeline's source label. |

The preparation pipeline assigns `data_source="livecodebench"` to all records. This field is a shared label across the DeepCoder sources listed above.

## Download

Download the two original Parquet files:

```bash
hf download WangSl2004/CATCH-RL \
  data/train_verl.parquet data/test_verl.parquet \
  --type dataset \
  --local-dir CATCH-RL
```

The files use the schema expected by the CATCH training pipeline. Run generated code in an isolated environment.

## License

CATCH-specific content uses the [Apache License 2.0](LICENSE). The [source dataset](https://huggingface.co/datasets/agentica-org/DeepCoder-Preview-Dataset) lists the MIT license.

## Citation

```bibtex
@misc{wang2026catch,
  title={CATCH: A Controllable Analysis Testbed for Reward Hacking in Coding RL},
  author={Shouli Wang and Yanfeng Jia and Zhihao Ou and Zitao Su and Ruize He and Haotong Xie and Hao Peng and Juanzi Li and Xiaozhi Wang},
  year={2026},
  eprint={2609.39533},
  archivePrefix={arXiv}
}
```