collabllm 0.0.6

(Oral @ ICML 2025) CollabLLM: From Passive Responders to Active Collaborators

CollabLLM: From Passive Responders to Active Collaborators

📢 Oral @ ICML 2025 (1% out of all submissions)

Overview

CollabLLM transforms traditional language models from passive responders to active collaborators in multi-turn conversations. This repository provides the complete framework for computing multiturn-aware rewards and training collaborative language models.

---

Installation

To get started, create a new environment and install collabllm via pip:

conda create -n collabllm python=3.10
conda activate collabllm
pip install collabllm

Optional: You may install additional packages (e.g., bigcodebench) for task-specific metrics or evaluation.

Quick Start

• Lightweight usage: Compute Multiturn-aware Rewards (MRs) for any model responses and construct datasets following notebook_tutorials/.
• Synthetic data generation: Generating high-quality synthetic conversational data following scripts/engine/build_dataset.py.
• Train CollabLLM: Conduct SFT/DPO/PPO models training to maximize MRs following examples under scripts/train/*.py.

Add Your Own Task

To apply CollabLLM to a new task:

1. Add a Dataset:
   Place your single-turn dataset in examples/single_turn_ds/ and register it in __init__.py.

2. (Optional) Add Metrics:
   Add new metrics to examples/metrics/ and register them in __init__.py.

You can now run data generation, reward computation, and model training using your customized setup.

Citation

If you find our work useful in your research, please cite the following:

@inproceedings{collabllm2025,
    title={CollabLLM: From Passive Responders to Active Collaborators},
    author={Shirley Wu and Michel Galley and Baolin Peng and Hao Cheng and 
            Gavin Li and Yao Dou and Weixin Cai and James Zou and 
            Jure Leskovec and Jianfeng Gao},
    booktitle={International Conference on Machine Learning (ICML)},
    year={2025}
}