Awesome

LongAlign: A Recipe for Long Context Alignment of LLMs

阅读中文版本

LongAlign is the first full recipe for LLM alignment on long context. We propose the LongAlign-10k dataset, containing 10,000 long instruction data of 8k-64k in length. We investigate on training strategies, namely packing (with loss weighting) and sorted batching, which are all implemented in our code. For real-world long context evaluation, we introduce LongBench-Chat that evaluates the instruction-following capability on queries of 10k-100k length.

🔍 Table of Contents

• ⚙️ Data Preparation
• 🖥️ LongAlign Training
• 📊 Evaluation
• 📝 Citation

<a name="data-preparation"></a>

⚙️ Data Preparation

You can download and save the LongAlign-10k data through the Hugging Face datasets (🤗 HF Repo):

dataset = load_dataset('THUDM/LongAlign-10k')
for split, split_dataset in dataset.items():
    split_dataset.to_json("data/raw/long.jsonl")

The ShareGPT data can be downloaded from here. We refer to the open-instruct repository for the preprocesss of ShareGPT data. Please save the data file at data/raw/sharegpt.jsonl. You can use other data as a source for general instruction data, but please format your data as follows:

{
    "messages": [{"role": "user", "content": "..."}, 
                 {"role": "assistant", "content": "..."}, ...]
    }

<a name="longalign-training"></a>

🖥️ LongAlign Training

Environmental Setup

Install the requirements with pip: pip install -r requirements.txt. For Llama based models, we recommend using FlashAttention 2 for optimization and saving GPU memory. The relevant dependencies can be installed according to the code base of FlashAttention.

Data preprocessing

First, tokenize the raw text data using the tokenizer of the model. For example, when training ChatGLM:

python pre_tokenize.py --model chatglm --datanum 10k

The --datanum parameter here refers to the amount of long data you want in your mixed training dataset (our paper investigates on 0k, 5k, and 10k). The tokenized data will be saved under ./data/chatglm/10k.

For the packing and sorted batching strategies, we then organize the tokenized data for training:

python sort_and_group.py --group_size 8 --train_file ./data/chatglm/10k

You should set the --group_size parameter to the number of GPUs during training. We recommend using at least 8 80G GPUs for model training, otherwise the 64k length may incur memory overflow.

Model training

We provide training scripts under scripts/ for the ChatGLM3 and Llama-2 model series. Make sure to adjust --model_name_or_path, --train_file, and --output_dir to match your model path, data path, and output path. You should consider using a base model with at least 64k context window length. We release three base models with extended context windows of 64k: LongAlign-6B-64k-base, LongAlign-7B-64k-base, and LongAlign-13B-64k-base.

For packing training, please modify the attention calculation to support the 1D attention mask that marks the start and end position of each sequence in the pack, and the model forward function to support loss weighting during packing training. An example of such modifications for the ChatGLM3 model is provided in modeling_chatglm.py, in CoreAttention.forward and ChatGLMForConditionalGeneration.forward. You can directly use this file as the modeling file for ChatGLM packing training. We also provide the training code for Llama. To reproduce our results, please use modeling_llama.py as the modeling file. As suggested in the result our paper, we recommend packing+loss weighting for ChatGLM training and sorted batching for Llama.

Model deploying

We have released four chat models trained using LongAlign: LongAlign-6B-64k (based on ChatGLM3-6B), LongAlign-7B-64k (based on Llama-2-7B), LongAlign-13B-64k (based on Llama-2-13B), and ChatGLM3-6B-128k. Try the model to summarize our paper, or ask anything about it:

from transformers import AutoTokenizer, AutoModelForCausalLM
import torch
tokenizer = AutoTokenizer.from_pretrained("THUDM/LongAlign-6B-64k", trust_remote_code=True)
model = AutoModelForCausalLM.from_pretrained("THUDM/LongAlign-6B-64k", torch_dtype=torch.bfloat16, trust_remote_code=True, device_map="auto")
model = model.eval()
query = open("assets/paper.txt").read() + "\n\nPlease summarize the paper."
response, history = model.chat(tokenizer, query, history=[], max_new_tokens=512, temperature=1)
print(response)

For Llama-based models, we also provide a llama_flash_attn_monkey_patch.py for utilization of FlashAttention-2 to save memory for inference on long sequences.

All available models

Here is the full list of models we released:

<a name="longbench-chat-evaluation"></a>

📊 Evaluation

LongBench-Chat evaluation

LongBench-Chat is the first benchmark for assessing long context alignment, featuring real user queries of 10k-100k in length. The dataset and evaluation code are available under LongBench_Chat/. Remember to configure your OpenAI API key in eval.py since we adopt GPT-4 as the evaluator. Run

python eval.py --model {model_path} --max_length {max_length}

model_path can either be your local model path or a Hugging Face model path. Here is the leaderboard on LongBench-Chat:

You are also welcome to submit your model's test predictions or results to us. We are planning to release a more formal leaderboard.

Needle-test evaluation

We also provide the code for evaluating HuggingFace models on the "Needle In A Haystack" test under Needle_test/. See its README.md for more information.

To reproduce our results on other benchmarks, we refer to the code in LongBench, FastChat, and lm-evaluation-harness for evaluating on LongBench, MT-Bench, and Open LLM Leaderboard tasks.

<a name="citation"></a>

📝 Citation

If you find our work useful, please consider citing LongAlign:

@inproceedings{bai2024longalign,
    title = "{L}ong{A}lign: A Recipe for Long Context Alignment of Large Language Models",
    author = "Bai, Yushi and Lv, Xin and Zhang, Jiajie and He, Yuze and Qi, Ji and Hou, Lei and Tang, Jie and Dong, Yuxiao and Li, Juanzi",
    booktitle = "Findings of the Association for Computational Linguistics: EMNLP 2024",
    month = nov,
    year = "2024",
    address = "Miami, Florida, USA",
    publisher = "Association for Computational Linguistics",
    url = "https://aclanthology.org/2024.findings-emnlp.74",
    doi = "10.18653/v1/2024.findings-emnlp.74",
    pages = "1376--1395",
}