quannguyen204/qwen3-4b-elderly-sft-merged

Qwen3-4B Elderly Vietnamese (SFT merged)

Merged LoRA adapter + base for Qwen/Qwen3-4B-Instruct-2507. Trained on 47.5k DiaSynth Vietnamese elderly dialogues. See insights.md and metrics.json.

• —Train loss: 0.876, Eval loss: 0.861
• —Train steps: 8913, Runtime: 21.3h
• —Base: bf16 LoRA r=32, PiSSA init, 3 epochs, eff batch 16

Source adapter: https://huggingface.co/quannguyen204/qwen3-4b-elderly-sft-lora