VLM-Mix Exact LoRA STEM60 Geometry40
An exact rank-128 weighted LoRA parameter merge of the independently trained STEM and geometry specialists.
This is a PEFT LoRA adapter for Qwen/Qwen3-VL-4B-Instruct, pinned during evaluation to
base revision ebb281ec70b05090aa6165b016eac8ec08e71b17. It contains adapter weights only, not the
base model or training images.
Adapter metadata
- LoRA rank: 128
- LoRA alpha: 128
- Adapter SHA-256:
f275b48c49a9da5f8539985742eb5a97a381b53151621c0f71960b35653bc149
- Expected base revision:
ebb281ec70b05090aa6165b016eac8ec08e71b17
Composition
This rank-128 adapter is the exact weighted LoRA delta sum of the 60% broader-STEM and 40% geometry specialists. The included merge_audit.json pins source paths, coefficients, scales, and hashes.
Reproducibility
This repository is a public research artifact from the VLM mixture/PoEM
experiments. Benchmark prompts, protocol hashes, and comparison results are
maintained in the accompanying experiment repository. Load with PEFT on the
pinned Qwen3-VL base model.
