gratex/Mistral-Small-3.2-24B-Instruct-2506-sbvr-compliance-W4A16

Add custom HELM evaluation datasets (unresolvable_patterns, damaged_items_mc, insurance_claim_extraction)

Model card: current-model documentation only

Vision repair: rename multimodal projector tensors to Mistral-native names (w_in/w_out/pre_mm_projector_norm/patch_merger) so vLLM's mistral-format loader matches them; add params.json/tekken.json and fix image-processor configs. Quantized weights unchanged (projector tensors verified bit-exact).

Update README.md

Upload W4A16 quantized model

Upload calibration_dataset.jsonl with huggingface_hub

Upload quantize.py with huggingface_hub

Upload README.md with huggingface_hub

initial commit