/yarn

YaRN: Efficient Context Window Extension of Large Language Models

Primary LanguagePythonMIT LicenseMIT

YaRN

This repo contains the code and data for the YaRN context window extension method.

Preprint

Preprint v2 (arXiv): YaRN: Efficient Context Window Extension of Large Language Models

Models

LLaMA

We publish 7B and 13B variants of Llama 2 fine-tuned with YaRN at 64K and 128K context window length. They are available under the Llama 2 license on 🤗 Hugging Face.

Size Context Link
7B 64K NousResearch/Yarn-Llama-2-7b-64k
7B 128K NousResearch/Yarn-Llama-2-7b-128k
13B 64K NousResearch/Yarn-Llama-2-13b-64k
13B 128K NousResearch/Yarn-Llama-2-13b-128k

In addition, we also publish 8K context window versions of Llama 2 7B fine-tuned with NTK-aware and YaRN (Table 1 in the conference paper).

Mistral

With the release of v2 of our paper we are also publishing 64K and 128K variants of Mistral 7B v0.1.

Size Context Link
7B 64K NousResearch/Yarn-Mistral-7b-64k
7B 128K NousResearch/Yarn-Mistral-7b-128k

Reproduction

We strongly believe in open science, and thus publish all code and data to reproduce the results in our paper. To reproduce, clone the repository and perform a local installation.

git clone https://github.com/jquesnelle/yarn
cd yarn
pip install -e .

Training

To train the models, run accelerate config and enable DeepSpeed acceleration. deepspeed/zero3.json was the configuration file used for training.

# ./train.sh

The tokenized training data is available on 🤗Hugging Face and was derived from the pg19 dataset. For the Mistral models, a mix of the pretrain and fine-tune splits of Long-Data-Collections was used and the tokenized dataset is also available on 🤗Hugging Face.

Evaluation

To reproduce the evaluations, install lm-evaluation-harness with pip install git+https://github.com/EleutherAI/lm-evaluation-harness and then run the two provided scripts.

# ./eval.sh
# ./eval-harness.sh

Citation

@misc{peng2023yarn,
      title={YaRN: Efficient Context Window Extension of Large Language Models}, 
      author={Bowen Peng and Jeffrey Quesnelle and Honglu Fan and Enrico Shippole},
      year={2023},
      eprint={2309.00071},
      archivePrefix={arXiv},
      primaryClass={cs.CL}
}