Cevra Math Decisions LoRA
A decision model based on openai/gpt-oss-20b for classifying math problems by subject and scoring choices.
Training: 3,622 examples · Validation: 458 examples · Fine-tuning: 227 steps, 1 epoch.
Parameters: LoRA rank 16, batch size 16, learning rate 0.0001, maximum length 2048, seed 42.
Example
from huggingface_hub import snapshot_download
adapter_path = snapshot_download("EugeneEvstafev/cevra-math-decisions-lora")
print(adapter_path)
Use the downloaded adapter with the Tinker Cookbook Decisions renderer (gpt_oss_no_sysprompt) to score the subject choices.
Validation comparison
| Metric | Base | Cevra |
|---|---|---|
| Distribution Brier (lower is better) | 0.468040 | 0.090603 |
| Probability MAE (lower is better) | 0.132055 | 0.053061 |
| Argmax agreement (higher is better) | 35.81% | 46.07% |
Validation was used for checkpoint selection.
- Downloads last month
- -
Inference Providers NEW
This model isn't deployed by any Inference Provider. 🙋 Ask for provider support
Model tree for EugeneEvstafev/cevra-math-decisions-lora
Base model
openai/gpt-oss-20b