AM-Thinking-V1 RL Dataset
Dataset designed for reinforcement training (RL) of language models dedicated to advanced reasoning in mathematics and code. Each example contains a system instruction, a user request, and the expected results to evaluate the response.
Several thousand examples in JSON, Verl structured format, math queries and code with associated verification
Apache 2.0
Description
The dataset AM-Thinking-V1 RL Dataset is intended for training by learning by reinforcement of language models specialized in mathematics and programming. It includes queries with system instructions and user questions, accompanied by test verifications or expected answers according to the Verl format.
What is this dataset for?
- Train LLM models to improve mathematical reasoning and programming.
- Test the ability of models to solve problems with automatic validation via test-case.
- Optimize reinforcement learning models by following specific reward criteria.
Can it be enriched or improved?
Yes, this dataset can be supplemented by new specific queries, additional validation tests, or adapted to other areas of reasoning. Annotating results manually or automatically can improve the quality of the reward signal.
🔎 In summary
🧠 Recommended for
- AI researchers
- LLM Model Developers
- R&D teams in reinforcement learning
🔧 Compatible tools
- RL frameworks (e.g. RLHF)
- JSON libraries
- LLM training environments
💡 Tip
Use the reward_model fields to precisely guide optimization during reinforcement learning.
Frequently Asked Questions
What technical skills are needed to use this dataset effectively?
A good understanding of language models, the JSON Verl format, and reinforcement learning is recommended.
Is this dataset suitable for use outside of academic research?
Yes, it can be used in industrial projects aimed at improving intelligent assistants in mathematics and programming.
Can you add your own data to the dataset?
Yes, the dataset is modular and can be enriched with additional examples, respecting the existing format.
