Qwen3-4B Safety Midtrain ? On-policy Reasoning Only
Continued-pretraining/midtraining checkpoint initialized from Qwen3-4B-Base. Training input: data/onpolicy_reasoning_only.jsonl from the Zzyy2000/qwen3-midtrain-data dataset bundle. It is not initialized from the non-Base Qwen3-4B instruct model.
