Dialogue Chain-of-Thought Distillation for Commonsense-aware Conversational Agents

Official Repository of "Dialogue Chain-of-Thought Distillation for Commonsense-aware Conversational Agents" accepted at EMNLP 2023.

Hyungjoo Chae*, Yongho Song*, Kai Tzu-iunn Ong, Taeyoon Kwon, Minjin Kim, Youngjae Yu, Dongha Lee, Dongyeop Kang, Jinyoung Yeo
^{* Equal contribution}

Paper Link: https://arxiv.org/abs/2310.09343

Demonstration: https://dialoguecot.web.app/

Overview of dialogue chain-of-thought distillation processs.

Requirements

• We implemented DOCTOR using 8 NVIDIA A5000 GPUs (VRAM 24 GB). You may need gpu accelerator with at least 16GB VRAM.
• Make sure PyTorch (>= 1.8.0) installed.
• As we use qlora for efficient training, bitsandbytes library is required.
• Prepare your OpenAI account (for API call, about 1500 USD is expected for the full annotation. Prepare your money!).

STEP 1. Rationale annotation on existing dialogue datasets

• Get prepared the dialogue datsets using src/rationale_annotation/prepare_dataset.py
• Start rationale annotation by executing sh src/rationale_annotation/run_distill.sh. Note that you have to set the environment variable OPENAI_API_KEY as your openai api key.

STEP 2. Alignment filtering

Apply filtering using our alignment filters. We provide the dataset which already has been filtered by us.

Resulitng dataset (🍩 DONUT)

We share our dataset in the linke below.

DLI-Lab/DONUT

STEP 3. Training 👨‍⚕️ DOCTOR

We train DOCTOR based on OPT 1.3B using qlora which supports 4-bit quantization with lora adapter.

Checkpoint for DOCTOR

We share our checkpoint for DOCTOR in the link below. Using the collected dataset, we train our knowledge model, DOCTOR, in a seq2seq manner.

sh src/train_knowledge_model/train_ours_final.sh

DLI-Lab/DOCTOR

Evaluation

Lastly, we evaluate DOCTOR as a commonsense reasoner that supports response generation. We use ChatGPT as a dialogue agent. Before generating responses, you have to prepare the predicted rationale files which will be concatenated with dialogue context and fed as an input for dialogue agent.

python src/run_response_generation.py

Citation

If you find this useful, please consider citing our paper:

@misc{chae2023dialogue,
      title={Dialogue Chain-of-Thought Distillation for Commonsense-aware Conversational Agents}, 
      author={Hyungjoo Chae and Yongho Song and Kai Tzu-iunn Ong and Taeyoon Kwon and Minjin Kim and Youngjae Yu and Dongha Lee and Dongyeop Kang and Jinyoung Yeo},
      year={2023},
      eprint={2310.09343},
      archivePrefix={arXiv},
      primaryClass={cs.CL}
}