Qwen3.5-Medical-GSPO

Qwen3.5-4B-Medical-GSPO

A Chinese medical reasoning model fine-tuned from Qwen3.5-4B using a two-stage training pipeline: Supervised Fine-Tuning (SFT) for format alignment, followed by Group Sequence Policy Optimization (GSPO) with an LLM-as-Judge reward function.

Model Description

This model is designed to produce structured chain-of-thought (CoT) reasoning for Chinese medical questions, including clinical diagnosis, treatment planning, and differential diagnosis.

Training Pipeline

Stage 1 — Supervised Fine-Tuning (SFT)

The model was first trained with SFT on the FreedomIntelligence/medical-o1-reasoning-SFT dataset (Chinese subset) to establish a consistent output format: a <think>...</think> reasoning block followed by a concise final answer.

Stage 2 — GSPO with LLM-as-Judge

GSPO was proposed by Zheng et al. (arXiv:2507.18071).

Reinforcement learning was applied using GSPO (Group Sequence Policy Optimization), a sequence-level variant of GRPO that computes importance ratios and clipping at the sequence level rather than the token level, improving training stability over long horizons.

The reward function uses DeepSeek-Chat as an LLM judge with a 5-tier scoring scheme:

The judge evaluates only the content after </think> — the final conclusion — not the CoT reasoning chain. This design prevents reward hacking via verbose but incorrect reasoning.

Training Details

Training Dynamics

The model was trained across multiple resume sessions. Reward (smoothed) improved consistently from ~0.57 in early training to ~0.89 in the final 50 steps. KL divergence remained stable around 0.5 throughout GSPO training, with no signs of collapse.

Usage

from unsloth import FastVisionModel
import torch

max_seq_length = 2048 # Can increase for longer reasoning traces
model, tokenizer = FastVisionModel.from_pretrained(
    "lastmass/Qwen3.5-Medical-GSPO",
    max_seq_length = max_seq_length,
    load_in_4bit = False,
    fast_inference = False, 
    max_lora_rank = lora_rank,
    gpu_memory_utilization = 0.9, 
)

messages = [
    {"role" : "user", "content" : """患者男性，65岁，突发胸痛2小时，心电图示II、III、aVF导联ST段抬高，最可能的诊断是？"""}
]
text = tokenizer.apply_chat_template(
    messages,
    tokenize = False,
    add_generation_prompt = True, # Must add for generation
    enable_thinking=True,
)

from transformers import TextStreamer
_ = model.generate(
    **tokenizer(text, return_tensors = "pt").to("cuda"),
    max_new_tokens = 512, # Increase for longer outputs!
    # temperature = 0.7, top_p = 0.8, top_k = 20, # For non thinking
    temperature=1.0, top_p=0.95, top_k=20,
    streamer = TextStreamer(tokenizer, skip_prompt = True),
)

Example Outputs

Example 1 — Cardiovascular Diagnosis

Input:

患者男性，65岁，突发胸痛2小时，心电图示II、III、aVF导联ST段抬高，最可能的诊断是？

Output (<think> block):

患者是65岁的男性，突然胸痛，这听起来很紧急。心电图显示II、III、aVF导联的ST段抬高，这让我想到心梗。嗯，II、III、aVF这些导联通常和下壁有关，所以可能是下壁的心肌梗死。不过，我还需要考虑其他可能性。比如，有没有可能是其他原因导致的胸痛？比如肺栓塞？肺栓塞的话，心电图上通常会有右心负荷过重的表现，但这里提到的是ST段抬高，这更像是心肌缺血的表现。再考虑心包炎——心包炎通常会有广泛导联的ST段抬高，而不是像这样集中在II、III、aVF。所以，心包炎的可能性不大。综合来看，下壁的心肌梗死是最可能的诊断。

Final answer:

根据患者的症状和心电图表现，最可能的诊断是下壁心肌梗死。II、III、aVF导联的ST段抬高是下壁心肌梗死的典型心电图特征。建议立即进行心肌酶谱检测、动态心电图监测，并尽快启动再灌注治疗（PCI或溶栓）。

---

Example 2 — Complex Obstetric Emergency

Input:

患者女性，28岁，停经8周，突发右下腹剧烈疼痛伴晕厥，血压80/50mmHg，腹部压痛明显，后穹窿穿刺抽出不凝血，最可能的诊断是？

Output (<think> block):

患者是28岁的女性，停经8周，这应该是个怀孕的早期阶段。突然右下腹剧烈疼痛，还晕厥了，这听起来不太妙。血压80/50mmHg，这明显是低血压，可能意味着有失血的情况。腹部压痛明显，后穹窿穿刺抽出不凝血，这提示有腹腔内出血。结合怀孕8周，我想到可能是宫外孕。怀孕早期，如果胚胎在子宫外着床，比如输卵管，那就会引起剧烈腹痛和出血。不凝血这个特征很重要——腹腔内出血通常不会凝固，这进一步支持了宫外孕破裂导致腹腔出血的诊断。

Final answer:

根据患者的症状和病史，最可能的诊断是宫外孕（输卵管妊娠破裂）。停经8周+突发剧烈腹痛+休克体征+后穹窿抽出不凝血，高度提示腹腔内大出血。需立即开通静脉通路、备血，急诊手术探查，同时完善血HCG和床旁超声。

---

Limitations

• This is a LoRA adapter trained on a relatively small dataset (~20k examples). It is not validated for clinical use.
• Performance on rare diseases, highly specialized subspecialties, and cases requiring memorization of specific drug dosages or classification codes may be limited.
• The model performs better on reasoning-heavy questions than on pure factual recall (e.g., specific surgical nomenclature, drug interaction tables).
• Outputs should always be reviewed by qualified medical professionals before any clinical application.

Citation

If you use this model, please cite the relevant works:

@misc{qwen3.5-medical-gspo,
  author = {lastmass},
  title = {Qwen3.5-4B Medical GSPO: Chinese Medical Reasoning via LLM-as-Judge Reinforcement Learning},
  year = {2025},
  publisher = {HuggingFace},
  url = {https://huggingface.co/lastmass/Qwen3.5-GSPO-Lora}
}

Acknowledgements

• Unsloth for efficient fine-tuning infrastructure
• FreedomIntelligence for the medical reasoning dataset
• DeepSeek for the reward model API
• Qwen Team for the base model