metadata

license: llama3
datasets:
  - princeton-nlp/prolong-data-64K
  - princeton-nlp/prolong-data-512K
  - HuggingFaceH4/ultrachat_200k
base_model:
  - princeton-nlp/Llama-3-8B-ProLong-512k-Base

princeton_nlp/Llama-3-8B-ProLong-512k-Instruct

[Paper] [HF Collection] [Code]

ProLong (Princeton long-context language models) is a family of long-context models that are continued trained and supervised fine-tuned from Llama-3-8B, with a maximum context window of 512K tokens. Our main ProLong model is one of the best-performing long-context models at the 10B scale (evaluated by HELMET).

To train this strong long-context model, we conduct thorough ablations on the long-context pre-training data, SFT data, and numerous other design choices. We demonstrate our findings in our paper, How to Train Long-Context Language Models (Effectively).

Authors: Tianyu Gao*, Alexander Wettig*, Howard Yen, Danqi Chen (* equal contribution)

Contact: {tianyug, awettig}@princeton.edu

The ProLong Models

Model card

Here are some quick facts about our main ProLong model: princeton-nlp/Llama-3-8B-ProLong-512k-Instruct.

Base model: meta-llama/Meta-Llama-3-8B-Instruct
Long-context continued training: 20B tokens on 64K training data (princeton-nlp/prolong-data-64K), and 20B tokens on 512K training data (princeton-nlp/prolong-data-512K)
Supervised fine-tuning (SFT): UltraChat
Maximum context window: 512K tokens

ProLong performance on HELMET averaged over 32K, 64K, and 128K lengths. All models are instruct models.

ProLong training recipe.

Citation

@article{gao2024prolong,
  title={How to Train Long-Context Language Models (Effectively)},
  author={Gao, Tianyu and Wettig, Alexander and Yen, Howard and Chen, Danqi},
  journal={arXiv preprint arXiv:2410.02660},
  year={2024}
}