icml2026-7516/ThinkSafe-Qwen3-0.6B
Adapter vs Qwen/Qwen3-0.6B. icml2026-7516/ThinkSafe-Qwen3-0.6B is a LoRA adapter, not full weights: rank 32, on q_proj, v_proj in 28 layers (56 adapted matrices). It modifies Qwen/Qwen3-0.6B at load time.
3 other repos hold byte-for-byte identical weight files.