| license: apache-2.0 | |
| pretty_name: TIPS Training Data | |
| task_categories: | |
| - text-generation | |
| configs: | |
| - config_name: math | |
| data_files: | |
| - split: train | |
| path: math/train.parquet | |
| - config_name: agent | |
| data_files: | |
| - split: train | |
| path: agent/train.parquet | |
| # TIPS Training Data | |
| This dataset is from the paper [Inducing Process Supervision from Outcome-Only Reinforcement Learning](https://huggingface.co/papers/2609.36641). | |
| Outcome-labeled training trajectories used by **TIPS (Thinking-Induced Process Supervision)**. | |
| | Configuration | File | Examples | | |
| | --- | --- | ---: | | |
| | Math | `math/train.parquet` | 3,200 | | |
| | Agent | `agent/train.parquet` | 2,905 | | |
| TIPS trains a generative reward model to produce a reasoning chain, step-level labels, and an outcome label while using only outcome correctness as the reinforcement-learning reward. | |
| ## Usage | |
| ```python | |
| from datasets import load_dataset | |
| math_data = load_dataset("XingYing-stack/TIPS-Training-Data", "math", split="train") | |
| agent_data = load_dataset("XingYing-stack/TIPS-Training-Data", "agent", split="train") | |
| ``` | |
| Data preparation and training code are available at https://github.com/RUCBM/TIPS. | |
| ## Sources and Licenses | |
| The math data is derived from SCAN-Pro, released under Apache-2.0. The agent data consists of rollout trajectories collected by us. Please also follow the terms of the original math data source. |