Screenshot 2026-11-17 at 10.47.59 PM

DeepSeek-R1-Distill-Llama-8B-Medical

DeepSeek-R1-Distill-Llama-8B-Medical is an 8B-parameter reasoning model for medical question answering. It is DeepSeek-R1-Distill-Llama-8B fine-tuned on medical reasoning traces, with the LoRA adapter merged into the original BF16 weights. Like its base model, it thinks step by step inside a reasoning block and then gives a separate final answer.

Training

Base model deepseek-ai/DeepSeek-R1-Distill-Llama-8B
Dataset FreedomIntelligence/Medical-R1-Distill-Data, medical questions with reasoning and answers distilled from DeepSeek-R1
Samples First 5,000 samples; 151 longer than 6,144 tokens were excluded (no sample was truncated), leaving 4,849
Method QLoRA with Unsloth: r=16, alpha=16, all linear layers, learning rate 1e-4 (cosine), 1 epoch, effective batch size 16
Format <think> reasoning </think> followed by the answer; loss computed on the reasoning and answer only
Hardware Single NVIDIA T4 (Kaggle)

Usage

For best results use temperature 0.6, top-p 0.95 and no system prompt. The chat template starts the response with <think> automatically, so the output begins with the reasoning, followed by a single </think> and the final answer.

import torch
from transformers import AutoTokenizer, AutoModelForCausalLM

repo = "Rumiii/DeepSeek-R1-Distill-Llama-8B-Medical"
tok = AutoTokenizer.from_pretrained(repo)
model = AutoModelForCausalLM.from_pretrained(repo, dtype=torch.bfloat16, device_map="auto")

messages = [{"role": "user", "content": "A 30-year-old woman has an itchy annular rash under her right breast. What is the most likely causative organism?"}]
inputs = tok.apply_chat_template(
    messages, add_generation_prompt=True, return_tensors="pt", return_dict=True
).to(model.device)

output = model.generate(**inputs, max_new_tokens=2048, do_sample=True, temperature=0.6, top_p=0.95)
print(tok.decode(output[0][inputs["input_ids"].shape[1]:], skip_special_tokens=True))

To see the reasoning and the answer separated by the </think> marker, decode with skip_special_tokens=False.

Limitations

  • Research model. It has not been clinically validated or benchmarked yet and can state incorrect medical facts with confidence. Do not use it for diagnosis or treatment decisions.
  • Training data was generated by another language model (DeepSeek-R1), so errors in those traces can be learned and repeated.
  • Trained on a 4,849-sample subset in English only; coverage of specialties, rare conditions and other languages is limited.
  • Long reasoning chains can contain mistakes or be repetitive. Always verify outputs against authoritative clinical sources.
Downloads last month
122
Safetensors
Model size
8B params
Tensor type
BF16
·
Inference Providers NEW
This model isn't deployed by any Inference Provider. 🙋 Ask for provider support

Model tree for Rumiii/DeepSeek-R1-Distill-Llama-8B-Medical

Finetuned
(174)
this model

Dataset used to train Rumiii/DeepSeek-R1-Distill-Llama-8B-Medical