starttoshow/medix-qwen3.5-2b-context-attention-ffn
medix-qwen3.5-2b-context-attention-ffn
Final adapter for experiment F, vqa_knowledge_case_context_attention_ffn, from the Medical Agent Assistant project. This repository contains adapter weights, not a standalone base model. Base revision: 15852e8c16360a2fea060d615a32b45270f8a8fc.
The adapter configuration pins the base model revision. Training arguments and evaluation scope are recorded in training_config.json and evaluation.json; release_manifest.json lists tensor shapes and the weight checksum.
SFT used one seed and one epoch; data-added recipes also change optimizer steps. A/B are VQA-only controls. The published 247-question report measures knowledge and context, not a four-task overall ranking.
Load
Use PeftConfig.from_pretrained(adapter_id) to obtain the pinned base configuration, load that base with Qwen3_5ForConditionalGeneration.from_pretrained(base_id, revision=base_revision), then PeftModel.from_pretrained(base, adapter_id). The GitHub project provides python -m examples.inference --adapter starttoshow/medix-qwen3.5-2b-context-attention-ffn --question "Your question" from its MediX-R1 directory. Supply --image for an image question. The example defaults to CPU; GPU use is explicit.
The processor/template files are retained from training. Do not treat the model as a deployed medical service. Medical safety, clinician-level reliability, and improved end-to-end Agent performance have not been established.