Update dataset card with paper link, metadata, and usage instructions

README.md

@@ -32,16 +32,58 @@ configs:
   data_files:
   - split: train
     path: synth_legs_train/train.parquet
+task_categories:
+- image-text-to-text
 ---
 
-# vlmfix HF Upload
+# VLM-Fix
 
-This folder contains local parquet exports prepared for Hugging Face upload.
+[**Project Page**](https://maveryn.github.io/vlm-fix/) | [**Paper**](https://huggingface.co/papers/2604.12119) | [**GitHub**](https://github.com/maveryn/vlm-fix) | [**Interactive Demo**](https://maveryn.github.io/vlm-fix/demo/)
 
-Included configs:
-- `vlm_fix`: main, images
-- `vlm_fix_text_only`: main
-- `vlm_fix_posttrain_d1`: train, test
-- `vlm_fix_posttrain_d2`: train, test
-- `vlm_fix_posttrain_d3`: train, test
-- `synth_legs_train`: train
+VLM-Fix is a controlled benchmark designed to study **semantic fixation** in Large Vision-Language Models (VLMs). Semantic fixation is the behavior where a model preserves a default interpretation (familiar semantic priors) even when a prompt specifies an alternative, equally valid mapping.
+
+The benchmark covers four abstract strategy games and evaluates identical terminal board states under paired standard and inverse rule formulations to isolate perception failures from rule-mapping failures.
+
+## Quick Start
+
+To run the main VLM-Fix benchmark from Hugging Face, you can use the evaluation runners provided in the [GitHub repository](https://github.com/maveryn/vlm-fix):
+
+```bash
+# Install dependencies
+pip install -r requirements.txt
+
+# Run the main VLM-Fix benchmark
+python scripts/run_vlm_fix_matrix.py \
+  --dataset-source hf \
+  --hf-repo maveryn/vlm-fix \
+  --hf-config vlm_fix \
+  --hf-split main \
+  --models Qwen/Qwen2.5-VL-7B-Instruct allenai/Molmo2-4B
+
+# Run the text-only benchmark
+python scripts/run_vlm_fix_text_only_matrix.py \
+  --dataset-source hf \
+  --hf-repo maveryn/vlm-fix \
+  --hf-config vlm_fix_text_only \
+  --hf-split main \
+  --models Qwen/Qwen2.5-VL-7B-Instruct
+```
+
+## Dataset Structure
+
+The repository contains several configurations:
+- `vlm_fix`: The main benchmark including images.
+- `vlm_fix_text_only`: Text-only version of the benchmark.
+- `vlm_fix_posttrain_d1/d2/d3`: Post-training datasets for rule-alignment studies.
+- `synth_legs_train`: Synthetic training data.
+
+## Citation
+
+```bibtex
+@article{alam2026beyond,
+  title={Beyond Perception Errors: Semantic Fixation in Large Vision-Language Models},
+  author={Alam, Md Tanvirul},
+  journal={arXiv preprint arXiv:2604.12119},
+  year={2026}
+}
+```