Add model card and metadata

README.md

@@ -1,3 +1,40 @@
 ---
 license: apache-2.0
+pipeline_tag: feature-extraction
 ---
+
+# AutoSelection Sparse Autoencoder (SAE)
+
+This repository contains a Sparse Autoencoder (SAE) checkpoint associated with the paper [From Instance Selection to Fixed-Pool Data Recipe Search for Supervised Fine-Tuning](https://huggingface.co/papers/2605.12944).
+
+## Model description
+
+AutoSelection is a budgeted solver for fixed-pool data recipe search. Instead of treating SFT data selection as a one-shot instance ranking problem, it searches over executable data-curation recipes that filter, mix, deduplicate, and recombine samples. 
+
+This SAE is used within the AutoSelection framework to extract task-, data-, and model-side signals during cold-start scoring and subset-state construction.
+
+**Configuration:**
+- **Architecture:** Top-K SAE
+- **Input Dimension (d_in):** 2048
+- **Expansion Factor:** 32
+- **K:** 192
+
+## Resources
+
+- **Paper:** [From Instance Selection to Fixed-Pool Data Recipe Search for Supervised Fine-Tuning](https://huggingface.co/papers/2605.12944)
+- **Repository:** [GitHub](https://github.com/w253/AutoSelection)
+- **Training Pool:** [k253/AutoSelection-90k](https://huggingface.co/datasets/k253/AutoSelection-90k)
+
+## Citation
+
+```bibtex
+@misc{wu2026instanceselectionfixedpooldata,
+      title={From Instance Selection to Fixed-Pool Data Recipe Search for Supervised Fine-Tuning},
+      author={Haodong Wu and Jiahao Zhang and Lijie Hu and Yongqi Zhang},
+      year={2026},
+      eprint={2605.12944},
+      archivePrefix={arXiv},
+      primaryClass={cs.LG},
+      url={https://arxiv.org/abs/2605.12944},
+}
+```