Update README.md
Browse files
README.md
CHANGED
|
@@ -28,11 +28,6 @@ _**Weizhou Shen, Chenliang Li, Fanqi Wan, Shengyi Liao, Shaopeng Lai, Bo Zhang,
|
|
| 28 |
|
| 29 |
_Tongyi Lab, Alibaba Group_
|
| 30 |
|
| 31 |
-
<p align="center">
|
| 32 |
-
<img src="./assets/performance.png" width="100%"> <br>
|
| 33 |
-
</p>
|
| 34 |
-
|
| 35 |
-
|
| 36 |
|
| 37 |
|
| 38 |
</div>
|
|
@@ -42,9 +37,6 @@ _Tongyi Lab, Alibaba Group_
|
|
| 42 |
In this work, we present QwenLong-CPRS, a novel framework designed to optimize long-context processing through query-aware multi-granularity compression, outperforming RAG and sparse attention methods. Distinct from RAG's coarse chunk-level retrieval, it achieves precise information extraction via token-level content selection, enhancing accuracy. Unlike sparse attention (SA) requiring model retraining, it functions as a plug-and-play module compatible with any downstream LLMs while eliminating retraining demands. This dual advantage enables both fine-grained context optimization and seamless integration across architectures.
|
| 43 |
|
| 44 |
|
| 45 |
-
<p align="center">
|
| 46 |
-
<img src="./assets/concept.png" width="100%"> <br>
|
| 47 |
-
</p>
|
| 48 |
|
| 49 |
We implement QwenLong-CPRS with four key innovations:
|
| 50 |
* _**Controllable Context Optimization**_: Processes control prompts + queries to generate compact, task-specific context segments without retraining.
|
|
@@ -58,11 +50,6 @@ We implement QwenLong-CPRS with four key innovations:
|
|
| 58 |
|
| 59 |
|
| 60 |
|
| 61 |
-
|
| 62 |
-
<p align="center">
|
| 63 |
-
<img src="./assets/framework.png" width="100%"> <br>
|
| 64 |
-
</p>
|
| 65 |
-
|
| 66 |
## 🎉 News
|
| 67 |
|
| 68 |
- **May 26, 2025:** 🔥 We release [🤗 QwenLong-CPRS-7B](https://huggingface.co/Tongyi-Zhiwen/QwenLong-CPRS-7B), a 7B context compression model designed for explicit long-context optimization.
|
|
@@ -75,28 +62,6 @@ We implement QwenLong-CPRS with four key innovations:
|
|
| 75 |
|
| 76 |
|
| 77 |
|
| 78 |
-
|
| 79 |
-
## 🎯 Model Results
|
| 80 |
-
|
| 81 |
-
Here are the evaluation results.
|
| 82 |
-
|
| 83 |
-
<p align="center">
|
| 84 |
-
<img src="./assets/main_res.png" width="100%"> <br>
|
| 85 |
-
</p>
|
| 86 |
-
|
| 87 |
-
<p align="center">
|
| 88 |
-
<img src="./assets/niah.png" width="100%"> <br>
|
| 89 |
-
</p>
|
| 90 |
-
|
| 91 |
-
<p align="center">
|
| 92 |
-
<img src="./assets/different_llm.png" width="100%"> <br>
|
| 93 |
-
</p>
|
| 94 |
-
|
| 95 |
-
|
| 96 |
-
<p align="center">
|
| 97 |
-
<img src="./assets/latency.png" width="100%"> <br>
|
| 98 |
-
</p>
|
| 99 |
-
|
| 100 |
## 🛠️ Requirements
|
| 101 |
|
| 102 |
```bash
|
|
@@ -166,13 +131,6 @@ python infer.py \
|
|
| 166 |
```
|
| 167 |
|
| 168 |
|
| 169 |
-
## 🌐 Join the Community
|
| 170 |
-
Chinese users can scan QR codes to join DingTalk/WeChat groups.
|
| 171 |
-
|
| 172 |
-
| WeChat | DingTalk |
|
| 173 |
-
|----------|---------|
|
| 174 |
-
|  |  |
|
| 175 |
-
|
| 176 |
## 📝 Citation
|
| 177 |
|
| 178 |
If you find this work is relevant with your research or applications, please feel free to cite our work!
|
|
|
|
| 28 |
|
| 29 |
_Tongyi Lab, Alibaba Group_
|
| 30 |
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 31 |
|
| 32 |
|
| 33 |
</div>
|
|
|
|
| 37 |
In this work, we present QwenLong-CPRS, a novel framework designed to optimize long-context processing through query-aware multi-granularity compression, outperforming RAG and sparse attention methods. Distinct from RAG's coarse chunk-level retrieval, it achieves precise information extraction via token-level content selection, enhancing accuracy. Unlike sparse attention (SA) requiring model retraining, it functions as a plug-and-play module compatible with any downstream LLMs while eliminating retraining demands. This dual advantage enables both fine-grained context optimization and seamless integration across architectures.
|
| 38 |
|
| 39 |
|
|
|
|
|
|
|
|
|
|
| 40 |
|
| 41 |
We implement QwenLong-CPRS with four key innovations:
|
| 42 |
* _**Controllable Context Optimization**_: Processes control prompts + queries to generate compact, task-specific context segments without retraining.
|
|
|
|
| 50 |
|
| 51 |
|
| 52 |
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 53 |
## 🎉 News
|
| 54 |
|
| 55 |
- **May 26, 2025:** 🔥 We release [🤗 QwenLong-CPRS-7B](https://huggingface.co/Tongyi-Zhiwen/QwenLong-CPRS-7B), a 7B context compression model designed for explicit long-context optimization.
|
|
|
|
| 62 |
|
| 63 |
|
| 64 |
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 65 |
## 🛠️ Requirements
|
| 66 |
|
| 67 |
```bash
|
|
|
|
| 131 |
```
|
| 132 |
|
| 133 |
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 134 |
## 📝 Citation
|
| 135 |
|
| 136 |
If you find this work is relevant with your research or applications, please feel free to cite our work!
|