A LoRA adapter fine-tuned on FreedomIntelligence/medical-o1-reasoning-SFT from the Qwen/Qwen3-14B base model, using the QLoRA method (NF4 4-bit frozen base + bf16 LoRA adapters).

Training configuration

Method: QLoRA (4-bit NF4 base, bf16 LoRA)
LoRA rank (r): 16
LoRA alpha: 32
Target modules: q_proj, k_proj, v_proj, o_proj, gate_proj, up_proj, down_proj
Framework: DeepSpeed ZeRO-2 + transformers + peft
Hardware: 12x NVIDIA L4 across 3x ml.g6.24xlarge on SageMaker HyperPod + NVIDIA Run:AI
Epochs: 3 (~1000 steps)
Dataset: FreedomIntelligence/medical-o1-reasoning-SFT (~19.7k samples)

Caveat

This adapter was trained as part of a Run:AI on SageMaker HyperPod workshop demonstration. For production-quality medical reasoning, consider longer training and a larger LoRA rank.

Usage with transformers + peft

python

from transformers import AutoModelForCausalLM, AutoTokenizer
from peft import PeftModel
import torch

base = AutoModelForCausalLM.from_pretrained(
    "Qwen/Qwen3-14B",
    torch_dtype=torch.bfloat16,
    device_map="auto",
)
model = PeftModel.from_pretrained(base, "ying2022/qwen3-14b-medical-qlora")
tokenizer = AutoTokenizer.from_pretrained("ying2022/qwen3-14b-medical-qlora")

messages = [{"role": "user", "content": "What is the first-line treatment for hypertension in type 2 diabetes?"}]
inputs = tokenizer.apply_chat_template(messages, return_tensors="pt", add_generation_prompt=True).to(model.device)
outputs = model.generate(inputs, max_new_tokens=500, temperature=0.3)
print(tokenizer.decode(outputs[0], skip_special_tokens=True))

Serving with vLLM

bash

vllm serve Qwen/Qwen3-14B \
  --enable-lora \
  --lora-modules qwen14b-medical=ying2022/qwen3-14b-medical-qlora \
  --max-loras 1 --max-lora-rank 16 \
  --tensor-parallel-size 2 \
  --gpu-memory-utilization 0.95 \
  --max-model-len 4096