Base model: meta-llama/Llama-3.2-3B-Instruct
Method: OPSD fixed-teacher, non-thinking mode
Teacher: fixed base policy with LoRA adapters disabled during teacher forward pass
Loss: full-vocabulary forward KL/JSD (beta=0)
Per-token JSD clipping: 1e-06
Student/teacher thinking flags: False / False
Dataset: siyanzhao/Openthoughts_math_30k_opsd
Train budget: max_steps=100, max_completion_length=1024
Batch: per_device_train_batch_size=1, gradient_accumulation_steps=2, effective batch 8
vLLM: colocate, GPU memory utilization 0.35
GPUs: 4

Only adapter/checkpoint artifacts and logs are uploaded; optimizer states are intentionally omitted.

yuxiaoyang/opsd-llama32-3b-ft-nonthink-gen1024-step100-jsdclip1e-6-20260515

opsd-llama32-3b-ft-nonthink-gen1024-step100-jsdclip1e-6-20260515