Qwen3.5-9B Medical Reasoning
A LoRA adapter fine-tuned for medical question answering and step-by-step clinical reasoning using a subset of the Medical-o1-Reasoning-SFT dataset.
The model is built on Jackrong/DeepSeek-V4-Pro-Qwen3.5-9B, a reasoning-focused model based on Qwen3.5-9B and distilled from DeepSeek-V4-Pro. Its upstream training emphasizes mathematics, STEM reasoning, and structured problem solving. This adapter adds medical reasoning examples to that existing model rather than training a new foundation model from scratch.
This is an experimental medical reasoning model, not a clinically validated system. Its outputs may contain incorrect diagnoses, fabricated details, or misleading explanations. It is intended for research and evaluation, not as a substitute for qualified medical judgment.
Model details
| Property | Details |
|---|---|
| Model type | LoRA adapter |
| Architecture lineage | Qwen3.5-9B |
| Fine-tuning base | Jackrong/DeepSeek-V4-Pro-Qwen3.5-9B |
| Fine-tuning method | Supervised fine-tuning (SFT) with LoRA |
| Primary task | Medical question answering and reasoning |
| Training dataset | FreedomIntelligence/medical-o1-reasoning-SFT |
| Training sample count | 10,000 randomly selected examples |
| Training duration | 2 epochs |
| Training precision | 16-bit |
| Maximum training sequence length | 4,096 tokens |
| Training hardware | NVIDIA L4 (24 GB VRAM) |
| License declared for this release | Apache-2.0 |
Intended use
Potential research and experimentation use cases include:
- Answering medical knowledge questions.
- Reasoning about symptoms, laboratory findings, and clinical vignettes.
- Exploring differential diagnoses and possible follow-up investigations.
- Evaluating how a reasoning-focused language model adapts to medical data.
- Prototyping medical NLP workflows in controlled, non-clinical environments.
The model has not been established as suitable for patient-facing diagnosis, treatment recommendations, emergency triage, or autonomous clinical decision-making.
Model lineage
This adapter builds on the following model lineage:
- Qwen/Qwen3.5-9B — the upstream model family.
- Jackrong/DeepSeek-V4-Pro-Qwen3.5-9B — a reasoning-focused Qwen3.5-9B model distilled from DeepSeek-V4-Pro.
- This model — a LoRA fine-tune of the immediate base on medical reasoning data.
The upstream model's training emphasizes mathematics and STEM problem solving, using distilled reasoning data from DeepSeek-V4-Pro. This provides an existing reasoning-oriented foundation for the additional medical fine-tuning.
The resulting behavior reflects both the upstream model's training and the additional medical fine-tuning. Their relative contributions have not been independently measured.
Training dataset and preprocessing
Training used a cleaned, randomly selected subset of FreedomIntelligence/medical-o1-reasoning-SFT, a dataset designed for medical question answering and reasoning.
The dataset was preprocessed before fine-tuning. Thinking traces were incorporated into the training data, and every training example includes a thinking trace. This provides consistent exposure to medical questions paired with reasoning-oriented responses rather than a mixture of examples with and without thinking traces.
Preprocessing and training data
- Source dataset:
FreedomIntelligence/medical-o1-reasoning-SFT - Data preparation: Cleaned before use
- Reasoning traces: Incorporated into the training examples
- Trace coverage: 100% of training examples
- Sampling: 10,000 randomly selected examples
- Training duration: 2 epochs
- Language: English
- Maximum sequence length: 4,096 tokens
- Evaluation split: 3% of the sampled data
The fine-tuning objective exposes the model to medical reasoning traces alongside the associated questions and answers, encouraging it to learn patterns of structured reasoning when addressing medical problems.
Reasoning behavior
This fine-tune is intended to apply the upstream model's reasoning-oriented capabilities to medical questions that benefit from multi-step reasoning, including interpreting symptoms and laboratory findings, considering plausible diagnoses, and identifying relevant follow-up investigations.
Although every training example includes a thinking trace, this does not guarantee that the model will reason correctly, produce faithful explanations, or achieve better diagnostic accuracy than its base model. Reasoning quality and medical accuracy require separate evaluation.
The presence of training traces also does not establish that generated explanations faithfully represent the internal processes responsible for the model's predictions.
Training configuration
| Parameter | Value |
|---|---|
| Method | LoRA supervised fine-tuning |
| GPU | NVIDIA L4 (24 GB) |
| Precision | 16-bit |
| Epochs | 2 |
| Learning rate | 2e-4 |
| Learning-rate scheduler | Linear |
| Optimizer | AdamW 8-bit |
LoRA rank (r) |
32 |
| LoRA alpha | 64 |
| LoRA dropout | 0.03 |
| Target modules | All modules, as configured in the training run |
| Per-device batch size | 4 |
| Gradient accumulation steps | 8 |
| Effective batch size | 32, assuming a single GPU |
| Weight decay | 0.005 |
| Maximum sequence length | 4,096 tokens |
Reported training results
| Metric | Value |
|---|---|
| Final training loss | 1.086 |
| Evaluation loss | 1.211 |
The evaluation set comprises approximately 3% of the sampled data. These losses reflect token-prediction performance under the training and evaluation setup; they do not independently establish medical accuracy or diagnostic reliability.
No independent medical benchmark results, clinician review, or comparison against the immediate base model are reported here.
Usage
Load the adapter together with its exact base model using compatible versions of Transformers and PEFT. The adapter should not be treated as a complete standalone set of model weights unless it has been explicitly merged and saved as a full model.
Example system prompt:
You are an assistant for medical reasoning research. Analyze the clinical information provided, identify the most plausible explanation, and explain the relevant medical reasoning. State important uncertainties and avoid presenting an unverified conclusion as a confirmed diagnosis.
For inference, use the base model's compatible tokenizer and chat template, and follow the loading requirements of the base model and the installed PEFT/Transformers versions.
Although the upstream model supports multimodal inputs, this adapter was fine-tuned on medical reasoning data. Its performance on medical images or other multimodal clinical inputs has not been established.
Limitations and safety
- No clinical validation: The model has not been validated for clinical use or shown to improve patient outcomes.
- Potential hallucinations: Explanations may sound convincing while containing incorrect medical facts or conclusions.
- Limited training sample: Fine-tuning on 10,000 examples does not guarantee coverage of medical specialties, uncommon conditions, or diverse patient populations.
- Inherited behavior: The upstream model's reasoning style, instruction-following behavior, and other characteristics may influence the resulting adapter.
- No demonstrated generalization: Performance on held-out examples from the same dataset may not transfer to real-world clinical cases.
- No established calibration: Confidence expressed in generated answers should not be interpreted as a reliable measure of diagnostic certainty.
- No benchmark claims: Evaluation loss is not a substitute for external evaluation against medical QA benchmarks or clinician-adjudicated cases.
Do not use the model as the sole basis for diagnosis, treatment, medication decisions, or emergency care. Outputs must be independently checked against reliable medical references and reviewed by qualified healthcare professionals where appropriate.
Licensing and attribution
This release declares the Apache-2.0 license. Users should review the licenses and terms associated with the immediate base model, upstream Qwen model, DeepSeek-V4-Pro, and training dataset, and verify that their intended use complies with all applicable terms.
Please acknowledge Jackrong/DeepSeek-V4-Pro-Qwen3.5-9B, the upstream Qwen model family, DeepSeek-V4-Pro, and the Medical-o1-Reasoning-SFT dataset when using this adapter in research or derivative work.