INTELLECT-MATH-SFT
Prime Intellect fine-tuned Qwen-7B-Math on verified QwQ-generated solutions to create this supervised starting policy for INTELLECT-MATH.
Research release
What the release provides
The release compares this supervised checkpoint with the later reinforcement-learning model. It makes the starting policy available for research on the effect of post-training.INTELLECT-MATH-SFT official model card
How to use it well
Practical guidance. Use this checkpoint to compare training stages. Track repeated reasoning and unfinished responses as well as final-answer accuracy.
The supervised model can repeat its reasoning
Prime's research article reports excessive self-questioning and occasional reasoning loops in the SFT model. The later reinforcement-learning stage reduced average response length.INTELLECT-MATH research release
Specifications & access
| Detail | Value and conditions |
|---|---|
| Status | Research release |
| Access | Downloadable |
| Context window | 4,096 tokens Configured position limit. Serving limits depend on the runtime and settings. INTELLECT-MATH-SFT model configuration |
| Base model | Qwen-7B-MathINTELLECT-MATH research release |
| Repository license | MITINTELLECT-MATH-SFT official model card |
Related models
- INTELLECT-MATH
Prime Intellect ยท INTELLECT-MATH
A seven-billion-parameter math model trained with supervised learning and reinforcement learning.
Launch date not confirmedResearch releaseDownloadable