dialogxpert

Codebase for ProactiveAI in Conversations — an approach combining LLM priors with Q-adapters for task-oriented dialogue planning.

This repo explains the following parts:

• Downloading LLM Weights
• How the model is trained
• How the model flows based on the architecture
• Extra information

---

Architecure Breakdown

The architecture diagram consists of the following main components:

• Policy Planner:
• Self-Play:
• Critic LLM:
• Replay Buffer:

Reinforcement learning is done based on the replay buffer

---

Downloading LLM Weights

Download the LLM model weights locally (it's easier because its faster to load!)

Steps:

1. Adjust the model name: https://github.com/declare-lab/dialogxpert/blob/master/download_llm_weights.py#L4-5

python download_llm_weights.py

NOTE:

• You will need to change the repo_id in download_llm_weights.py to change the LLM weights to download.

• Please ensure that you are logged into huggingface and have the necessary tokens enabled.

---

Training the model

Before you train the model:

• Decide the dataset to use
• Make the changes to the dataset arg (get_args_train -> --data_name parameter)
• Make changes to the necessary functions in the code in env.py:
  • LLM Policy Prompt: Replace with {dataset_name}_prompt (choose from qwen_prompts.py)
  • Roleplay functions: Replace with {dataset_name}_roleplay (choose from qwen_prompts.py)

After you are set, run:

python train_model.py

---

How Self-Play works

Training starts:

Episode loading:

Action selection:

Self-play (User):

Self-play (System):

Critic LLM:

Replay Buffer:

Status Check:

---

How Q-learning is done

Training the network:

Adjustments:

---

Others

Prompts:

Testing:

---

Repository Credits

The following repositories are given credit for their open-source code utilization

- PPDPP: https://github.com/dengyang17/PPDPP/tree/main
- DPDP: https://github.com/cs-holder/DPDP/tree/main
- RL-LLM: https://github.com/yanxue7/RL-LLM-Prior/tree/main

---