Add model card and metadata

README.md

@@ -1 +1,32 @@
-arxiv.org/abs/2603.27027
+---
+license: apache-2.0
+library_name: transformers
+pipeline_tag: text-generation
+---
+
+# TAPS: Task-Aware Proposal Distributions for Speculative Sampling
+
+[**Paper**](https://arxiv.org/abs/2603.27027) | [**Code**](https://github.com/Moe-Zbeeb/TAPS)
+
+TAPS is a research framework investigating how draft training distributions shape speculative decoding quality. This repository contains a lightweight, single-layer Llama-style drafter (~0.8B parameters) used as a proposal model to accelerate inference for larger verifier models like Meta-Llama-3-8B-Instruct.
+
+## Overview
+Speculative decoding speeds up autoregressive generation by letting a lightweight drafter propose tokens that a larger verifier checks in parallel. TAPS demonstrates that speculative decoding performance depends significantly on the alignment between the drafter's training data and the downstream workload. 
+
+Key findings include:
+- **Task Specialization:** Drafters trained on specific domains (e.g., MathInstruct vs. ShareGPT) perform best on related benchmarks (MT-Bench vs. reasoning tasks).
+- **Specialist Composition:** Specialists are better combined at inference time (via confidence routing or merged-tree verification) than through naive weight-space averaging.
+- **Routing Signals:** Confidence is a robust signal for routing requests to the most appropriate specialized drafter.
+
+## Usage
+For scripts to train, evaluate, or perform inference using these drafters (including HASS and EAGLE-2 variants), please visit the official [TAPS GitHub repository](https://github.com/Moe-Zbeeb/TAPS).
+
+## Citation
+```bibtex
+@article{zbib2026taps,
+  title={TAPS: Task Aware Proposal Distributions for Speculative Sampling},
+  author={Zbib, Mohamad and Bazzi, Mohamad and Mohanna, Ammar and Ghanem, Bernard and Hammoud, Hasan Abed Al Kader},
+  year={2026},
+  note={Technical report}
+}
+```