| license: apache-2.0 | |
| library_name: transformers | |
| pipeline_tag: text-generation | |
| tags: | |
| - speculative-decoding | |
| # TAPS: Task-Aware Proposal Distributions for Speculative Sampling | |
| This repository contains a draft model introduced in the paper [TAPS: Task Aware Proposal Distributions for Speculative Sampling](https://huggingface.co/papers/2603.27027). | |
| ## Overview | |
| TAPS (Task-Aware Proposal Distributions) investigates how the training distribution of draft models affects the efficiency of speculative decoding. In speculative decoding, a lightweight draft model proposes future tokens that a larger target model (like Meta-Llama-3-8B-Instruct) verifies in parallel. | |
| This model is a lightweight LLaMA-style drafter (approximately 0.8B parameters with 1 layer) designed to accelerate autoregressive generation. The research demonstrates that task-specific training (e.g., on MathInstruct or ShareGPT) yields significant specialization, and that specialized drafters are best combined at inference time using strategies like confidence-based routing or merged-tree verification. | |
| ## Resources | |
| - **Paper:** [TAPS: Task Aware Proposal Distributions for Speculative Sampling](https://huggingface.co/papers/2603.27027) | |
| - **GitHub Repository:** [https://github.com/Moe-Zbeeb/TAPS](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} | |
| } | |
| ``` |