metadata

license: mit
datasets:
  - yzhuang/Agentic-Long-Context-Understanding-QA
language:
  - en
base_model:
  - meta-llama/Llama-3.1-8B-Instruct

📖 Agentic Long Context Understanding 📖

Self-Taught Agentic Long Context Understanding (Arxiv).

AgenticLU refines complex, long-context queries through self-clarifications and contextual grounding, enabling robust long-document understanding in a single pass.

Installation Requirements

This codebase is largely based on OpenRLHF and Helmet, kudos to them.

The requirements are the same

pip install openrlhf
pip install -r ./HELMET/requirements.txt

Dataset & Model

Dataset for SFT and DPO is avaliable at here

Model is available at here

Data Generation Pipeline

To generate traces with your custom model or dataset, follow the instructions:

Get an OpenAI API key and set it as your env variable

export OPENAI_API_KEY="your_api_key_here"

Edit the bash sript as you needed for base model, search width and depth

PYTHONPATH="./":"$PYTHONPATH" python ./long_context_llm/qa_tree_datagen.py \
    --model_name_or_path meta-llama/Llama-3.1-8B-Instruct \
    --max_sample_size 8 \
    --max_tree_depth 2 \
    --dataset_name yzhuang/narrative_qa

The traces will be avaliable to you as dataset_dpo, feel free to add this line to push to your huggingface account.

dataset_dpo.push_to_hub("YOUR REPO")

Example Usage

We show the training script of AgenticLU at sft script, dpo script.

It is important to get ring-attention to work, as the inputs are extremely long and requires ring-attention and deepspeed for training.

Examples for inferencing with the agentic workflow can be found here, with baseline prompting scripts avaliable.

Questions?

If you have any questions related to the code or the paper, feel free to reach out to us at [email protected].

Citation

If you find our paper and code useful, please cite us:

@misc{zhuang2025selftaughtagenticlongcontext,
      title={Self-Taught Agentic Long Context Understanding}, 
      author={Yufan Zhuang and Xiaodong Yu and Jialian Wu and Ximeng Sun and Ze Wang and Jiang Liu and Yusheng Su and Jingbo Shang and Zicheng Liu and Emad Barsoum},
      year={2025},
      eprint={2502.15920},
      archivePrefix={arXiv},
      primaryClass={cs.CL},
      url={https://arxiv.org/abs/2502.15920}, 
}