Add model card and metadata

README.md

@@ -1 +1,29 @@
-arxiv.org/abs/2603.27027
+---
+license: apache-2.0
+pipeline_tag: text-generation
+---
+
+# TAPS: Task-Aware Proposal Distributions for Speculative Sampling
+
+This repository contains artifacts for **TAPS**, a study exploring how draft training distributions shape the quality and efficiency of speculative decoding.
+
+- **Paper:** [TAPS: Task Aware Proposal Distributions for Speculative Sampling](https://huggingface.co/papers/2603.27027)
+- **GitHub Repository:** [Moe-Zbeeb/TAPS](https://github.com/Moe-Zbeeb/TAPS)
+
+## Description
+TAPS demonstrates that speculative decoding performance depends heavily on the alignment between a draft model's training data and the downstream workload. The project provides lightweight HASS and EAGLE-2 drafters (~0.8B parameters) optimized for specific tasks like mathematical reasoning (trained on MathInstruct) or general chat (trained on ShareGPT).
+
+### Key Highlights
+- **Task-aware specialization:** Drafts trained on specific domains significantly outperform generic drafts on corresponding benchmarks (e.g., MathInstruct-trained drafts are strongest on GSM8K).
+- **Inference-time Composition:** The research explores strategies like confidence-based routing and merged-tree verification to combine specialized drafters at inference time, outperforming weight-space averaging.
+- **Architectures:** Provides implementations and weights for HASS and EAGLE-2 drafters used to accelerate the `Meta-Llama-3-8B-Instruct` target model.
+
+## 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}
+}
+```