add README

README.md

@@ -0,0 +1,27 @@
+---
+base_model: Qwen/Qwen3-VL-8B-Instruct
+library_name: transformers
+pipeline_tag: image-text-to-text
+tags:
+- vision-language
+- mvtec
+- lora
+---
+
+# IRPO — MVTec checkpoints
+
+Fine-tuned checkpoints of **Qwen/Qwen3-VL-8B-Instruct** from the IRPO project (inductive-stage
+experiments), trained on the MVTec-AD category set. Research artifacts; **optimizer state stripped**
+(inference/eval weights only).
+
+| subfolder | training | type |
+|---|---|---|
+| `sft-mvtec` | supervised fine-tuning (direct answer) | full model |
+| `rft-mvtec` | GRPO / RFT (answer-only), continued run | full model |
+| `ovr-mvtec-lora` | **OVR** (one-vs-rest rule-induction reward), 468 steps | LoRA adapter |
+| `ovr6-mvtec-lora` | **OVR** (one-vs-rest rule-induction reward), 234 steps | LoRA adapter |
+| `sftovr-mvtec-lora` | SFT -> OVR | LoRA adapter |
+
+Full models: load with `transformers.AutoModelForImageTextToText`.
+LoRA adapters: load the base model, then apply with `peft.PeftModel.from_pretrained`.
+Base model: `Qwen/Qwen3-VL-8B-Instruct`.