README.md

---
license: apache-2.0
task_categories:
- table-question-answering
- text-classification
- summarization
language:
- en
---

<h1 align="center"><font size="7">AceRAG</font></h1>

<div align="center">
<a href="https://github.com/VectorSpaceLab/LightRAG"><img alt="github" src="https://img.shields.io/badge/Github-LightRAG-181717?logo=github&color=1783ff&logoColor=white"/></a>
<a href="https://dl.acm.org/doi/abs/10.1145/3701551.3703580" target="_blank"><img src="https://img.shields.io/badge/ACM%20DL-Paper-blue?logo=acm"></a>
<a href="https://huggingface.co/wcyno23/TacZip-Qwen3-8b" target="_blank"><img src="https://img.shields.io/badge/%F0%9F%A4%97%20AceRAG--Qwen3--8b-orange"></a>
<a href="https://huggingface.co/wcyno23/TacZip-Llama2-7b" target="_blank"><img src="https://img.shields.io/badge/%F0%9F%A4%97%20AceRAG--Llama2--7b-orange"></a>
<a href="https://huggingface.co/datasets/wcyno23/TacZip-Data" target="_blank"><img src="https://img.shields.io/badge/%F0%9F%A4%97%20AceRAG--Data-ff69b4.svg"></a>
</div>
<h4 align="center">

## Introduction

AceRAG provides task-aware context compression that transforms lengthy prompts into compact representations while preserving task-critical information. Designed for retrieval-augmented generation (RAG) systems, it employs a task-aware compressor and a compression-rate adapter to dynamically emphasize important tokens according to the task.

## Dataset Directory Structure

```
AceRAG-Data/
├── train/                     # Training set
│   ├── context_compressor/    # Context compressor training data
│   └── token_level_estimator/ # Token-level estimator training data
└── eval/                      # Evaluation set
```