LLoCO: Learning Long Contexts Offline
Lloco-7b-hqa is the LoRA adaptor checkpoint finetuned from AutoCompressor-Llama-2-7b-6k and meta-llama/Llama-2-7b-hf using the LLoCO method in LLoCO: Learning Long Contexts Offline. It is instruction-tuned on the HotpotQA training set.
LLoCO enables LLMs to process long-context efficiently by learning contexts offline through context compression and in-domain parameter-efficient finetuning with LoRA. This approach extends the effective context window of a 4k token LLaMA2-7B model to handle up to 128k tokens, while using 30x fewer tokens and achieving up to 7.62x inference speed-up.
Released LoRA Checkpoint
| Model | LoRA Rank | Dataset | Link |
|---|---|---|---|
| Lloco-7b-quality | 8 | QuALITY | link |
| Lloco-7b-qasper | 8 | Qasper | link |
| Lloco-7b-qmsum | 8 | QMSum | link |
| Lloco-7b-nqa | 8 | NarrativeQA | link |
| Lloco-7b-hqa | 8 | HotpotQA | link |
Citation
If you find this project useful, please consider citing:
@article{tan2024lloco,
title={LLoCO: Learning Long Contexts Offline},
author={Tan, Sijun and Li, Xiuyu and Patil, Shishir and Wu, Ziyang and Zhang, Tianjun and Keutzer, Kurt and Gonzalez, Joseph E and Popa, Raluca Ada},
journal={arXiv preprint arXiv:2404.07979},
year={2024}
}
Evaluation
Check out LLoCO: Learning Long Contexts Offline for evaluation results on various long-context tasks such as long document question answering and summarization.