Small QLoRA medical QA adapter built on Qwen/Qwen3.6-35B-A3B, trained on a filtered quick-run subset of keivalya/MedQuad-MedicalQnADataset.
Model Details
Developed by: kingabzpro
Model type: Causal language model adapter
Language: English
License: Apache-2.0
Finetuned from model: Qwen/Qwen3.6-35B-A3B
Uses
Direct use: Short medical question-answering experiments and response-style adaptation.
Out-of-scope use: Clinical diagnosis, treatment decisions, or unsupervised medical advice.
Bias, Risks, and Limitations
This adapter was trained on a small filtered subset of MedQuad. Outputs can still be incomplete, generic, or outdated. It should be treated as an experimental adapter, not a medical authority.
Training Details
Training data: Filtered subset of keivalya/MedQuad-MedicalQnADataset
Training procedure: QLoRA with PEFT and TRL
Hardware: RunPod H100 SXM
Evaluation
Evaluation was qualitative using three held-out before/after comparisons.
Observed improvements:
The fine-tuned model aligned more closely with MedQuad answer style.
It matched the reference inheritance explanation for acral peeling skin syndrome.
It gave a more complete autosomal-dominant inheritance explanation for distal hereditary motor neuropathy type V.
For the glaucoma research example, the answer became more NIH/NEI-aligned, though narrower than the base model response.
📓 Notebook
To follow the full training workflow, see the notebook:
