qdovan03/medical-o1-vi-sft-llama3.1-8b-lora

medical-o1-vi-sft — LoRA adapter

LoRA adapter fine-tune cho Vietnamese medical QA có reasoning (<think>...</think>).

Chi tiết

• —Base model: unsloth/Meta-Llama-3.1-8B-Instruct-bnb-4bit
• —Method: QLoRA (4-bit) qua Unsloth
• —Dataset: qdovan03/medical-o1-vi-sft (10460 train / 2615 val)
• —Format: <think>{cot}</think>{response}, trainonresponses_only
• —Effective batch: 8 (batch=2 × grad_accum=4)
• —LR: 0.0002 · Epochs: 1 · Max seq len: 2048
• —Final train loss: n/a

Cách dùng

from unsloth import FastLanguageModel
model, tokenizer = FastLanguageModel.from_pretrained(
    "qdovan03/medical-o1-vi-sft-llama3.1-8b-lora", max_seq_length=2048, load_in_4bit=True,
)
FastLanguageModel.for_inference(model)

Lưu ý

Adapter dùng cho nghiên cứu, không thay thế tư vấn y tế chuyên môn.