Safety_Tuning-Qwen3.5-4B-GRPO-STAR-lr1e-6-kl0.001-p16-k8-step60
GRPO safety-tuning checkpoint of Qwen/Qwen3.5-4B, exported to Hugging Face safetensors.
Learning rate: 1e-6. Completed optimizer updates: 60. KL coefficient: 0.001. Prompts per update: 16. Responses per prompt: 8.
GRPO uses STAR prompts, with Qwen3.8-27B teacher scoring and a frozen original-model reference. The judge-footer prompt changed after update 44; the resumed run from step 90 accepts a final flat score JSON after teacher prose. These checkpoints belong to that recorded training lineage.
Base revision: 851bf6e806efd8d0a36b00ddf55e13ccb7b8cd0a.
Source run: qwen35-star-grpo-4gpu-save30-a100j16-judgefooter-20261107-v6.
Native checkpoint iteration: 59 (zero based).
The repository contains inference weights, configuration and tokenizer/processor assets.
Optimizer, RNG state, training logs and training examples are excluded.
All trained text tensors are exported from the native checkpoint. Vision and untrained
MTP tensors are retained from the pinned original base model; MTP was not trained.
Conversion verification includes exact inverse conversion of every native text tensor,
finite-weight checks and matching all 738 Hugging Face tensor keys and shapes.
No evaluation scores are asserted by this model card.