Awesome

Compression Represents Intelligence Linearly

This is the repository for the paper "Compression Represents Intelligence Linearly".

We find that LLMs’ intelligence – reflected by benchmark scores – almost linearly correlates with their ability to compress external text corpora. Our findings suggest that compression efficiency, as an unsupervised metric derived from raw text corpora, serves as a reliable evaluation measure that is linearly associated with the model capabilities. In this repo, we release the compression corpora we used in the paper, the code to compute compression efficiency, as well as our compression corpora collection and processing piplines. <img src="resources/overview.png" alt="overview" style="zoom: 5%;" />

News

• [2024.09.20] Our paper has been accepted to COLM 2024 🎉🎉🎉
• [2024.05.01] Our compression evaluation has been added into OpenCompass 🚀🚀🚀 Please refer to Use Through OpenCompass for details.

Table of Contents

• Leaderboard
• Compression Corpora Datasets
• Evaluating Compression
• Data Pipeline

Compression Leaderboard

We focus on three key abilities: knowledge and commonsense, coding, and mathematical reasoning and colloct new corpora from Common Crawl, GitHub, and Arxiv, respectively. Below are models’ compression efficiency on three external corpora. We report the average bits per character (BPC) as the metric. For more details, please refer to our paper.

Compression Corpora Datasets

We focus on three key abilities: knowledge and commonsense, coding, and mathematical reasoning. The corpora we used are sourced from Common Crawl, GitHub, and Arxiv, and are respectively named: cc, python, and arxiv_math respectively. The data can be obtained through Huggingface Datasets:

from datasets import load_dataset
dataset = load_dataset(r"hkust-nlp/llm-compression",name="python")
print(dataset['test'][0])

Below is our data structure, containing three fields: content, subset, meta. Specifically, "content" refers to the evaluation text data, and "meta" contains data-specific meta-information related to its subset.

"content": "A photo journal about returning...", 
"subset": "cc" ｜ "python" | "arxiv_math", 
"meta": {}

Evaluating Compression

We utilize Bits Per Character (BPC) as the evaluation metric, implementing both Context Window Unification and a sliding window approach for assessing compression performance. The Python code for this evaluation is accessible in the code/evaluation directory and necessitates the specified basic environment:

transformers
datasets
tqdm

After installing the necessary dependencies, execute the evaluation script code/evaluation/main.py with these optional arguments:

--task_name # specifies the subset to eval (cc|python|arxiv_math)
--model_name # specifies the model name
--block_size # specifies the context window
--stride   # specifies the stride of sliding window approach
--batch_size # specifies the batch size
--file_num  # specifies the number of examples to eval, useful for debugging
--flash      # enable this to use flash attention (requires the flash-attn package)
--gpu        # specifies the id of gpu
--cache_dir  # specifies the cache dir for huggingface

Example:

cd code/evlauation
python -u main.py\
    --model_name deepseek-ai/deepseek-llm-7b-base\
    --task_name cc\
    --block_size 1900\
    --stride 512\
    --batch_size 8\
    --flash

Use Through OpenCompass

Now, you can evaluate model on llm-compression through OpenCompass, which is a framework for LLM evaluation. The dataset name is llm_compression. For example, to evaluate a model hosted on the HuggingFace Hub (e.g. llama-7b) ,you can use the following command:

python run.py --datasets llm_compression --hf-path huggyllama/llama-7b --model-kwargs use_flash_attention_2=True  

Please refer to OpenCompass for more details.

Data Pipeline

We provide data collection pipelines to facilitate future data updates and research, which include:

• Common Crawl
• GitHub Python
• ArXiv

For details, please refer to the corresponding pages.

Licenses

This work is licensed under a MIT License.

Our dataset is primarily licensed under a Creative Commons Attribution-NonCommercial-ShareAlike 4.0 International License. When the data source mandates a stricter licensing agreement, we comply with those terms.

Citation

@inproceedings{
   huang2024compression,
   title={Compression Represents Intelligence Linearly},
   author={Yuzhen Huang and Jinghan Zhang and Zifei Shan and Junxian He},
   booktitle={First Conference on Language Modeling},
   year={2024},
   url={https://openreview.net/forum?id=SHMj84U5SH}
}