Timely Response Generation for Open-domain Dialogue Agents

This is the official repository for "From what to respond to when to respond: Timely Response Generation for Open-domain Dialogue Agents".

Data

TimelyChat dataset (both train and evaluation sets) is available on 🤗Datasets Hub.

Checkpoint

Timer-3B trained on TimelyChat-train is available on 🤗Model Hub.

Usage

Environment setup

python -m venv .venv
source .venv/bin/activate
pip install -r requirements.txt

Evaluation

• Turn-level evaluation

python evaluate_turn-level.py --model-type hf \
    --model-name seongbo/timer-3b \
    --task time \
    --icl-method zeroshot \
    --save-results

• Dialog-level evaluation

python evaluate_dialog-level.py --model-type hf \
    --model-name seongbo/timer-3b \
    --simulator gpt-4o \
    --num-turns 10

Citation

@article{jang2025respond,
  title={From What to Respond to When to Respond: Timely Response Generation for Open-domain Dialogue Agents},
  author={Jang, Seongbo and Jeon, Minjin and Lee, Jaehoon and Lee, Seonghyeon and Lee, Dongha and Yu, Hwanjo},
  journal={arXiv preprint arXiv:2506.14285},
  year={2025}
}