---
library_name: transformers
license: apache-2.0
base_model: Qwen/Qwen2.5-3B
tags:
- sft
- qwen2.5
- loss-analysis
---

# Qwen2.5-3B math SFT, shuffled: all checkpoints

Ten existing checkpoints: 107, 214, 322, 429, 536, 643, 750, 858, 965, 1072. Training and evaluation were already complete.

Per-checkpoint evaluation outputs: https://huggingface.co/datasets/RL-Forgetting-Experiments-3/qwen2.5-3b-math-kk-sft-artifacts/tree/main/runs/math_shuffled/eval

Each `checkpoints/step_N/` directory is a directly loadable Hugging Face model.