LongAlign-13B-64k

LongAlign-13B-64k

<p align="center"> 🤗 <a href="https://huggingface.co/datasets/THUDM/LongAlign-10k" target="_blank">[LongAlign Dataset] </a> • 💻 <a href="https://github.com/THUDM/LongAlign" target="_blank">[Github Repo]</a> • 📃 <a href="https://arxiv.org/abs/2401.18058" target="_blank">[LongAlign Paper]</a> </p>

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 trianing 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 evaluate the instruction-following capability on queries of 10k-100k length.

All Models

We open-sourced the following list of models:

Model usage

Chat prompt template for LongAlign-6B-64k:

[Round 1]

问：Hi!

答：Hello! What can I assist you today?

[Round 2]

问：What should I do if I can't sleep at night?

答：

Chat prompt template for LongAlign-7B-64k and LongAlign-13B-64k:

[INST]Hi![/INST]Hello! What can I assist you today?

[INST]What should I do if I can't sleep at night?[/INST]

ChatGLM3-6B-128k uses the same prompt template as ChatGLM3-6B.

A simple demo for deployment of the model:

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)

Citation

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