DialSim

We introduce DialSim, a real-time dialogue simulator. In this simulator, an agent is assigned the role of a character from popular TV shows, requiring it to respond to spontaneous questions using past dialogue information and to distinguish between known and unknown information. Key features of DialSim include evaluating the agent’s ability to respond within a reasonable time limit, handling long-term multi-party dialogues, and managing adversarial settings (e.g., swap character names) to challenge the agent’s reliance on pre-trained knowledge.

The dataset is released along with our paper. For further details, please refer to our paper.

Dataset

You can download the dataset here.

v1.0 (April 2024): This version includes the dataset as described in the paper.

v1.1 (June 2024): To incorporate more diverse and challenging data, this version has been updated to include unanswerable multi-hop questions.

Experimental Setup

After downloading appropriate version of torch, do:

1. pip install -r requirements.txt

2. mkdir data

3. mv dialsim_v1.1.zip ./data/

4. cd data

5. unzip dialsim_v1.1.zip

Simulation

Command Example: CUDA_VISIBLE_DEVICES=0 python simulator.py --model_name "GPT-3.5" --quantization "4bit" --script_name "friends" --sleep_time 6 --history_type "session-entire" --ret_method "bm25" --trial_version 0 --sh_number 0 --num_cores 10 --openai_api_key "<<YOUR_OPENAI_API_KEY>>"

Arguments

• model_name: Specifies the model to use, default is "GPT-3.5". Options include "llama2-7b-chat", "llama2-70b-chat", "tulu2-7b-dpo", "tulu2-70b-dpo", "gemma-2b-it", "gemma-7b-it", "mistral-7b-it", "mixtral-it", "GPT-3.5", "GPT-4", "claude-3", "claude-2.1", and "gemini".
• quantization: Model quantization level, default is "no". Options include "no", "16bit", "8bit", and "4bit".
• script_name: TV show script for the simulation, default is "friends". Options include "friends", "bigbang", and "theoffice".
• sleep_time: Response time limit, default: 5
• history_type: Method for saving history, default is "session-entire". Options include "utts", "session-entire", and "session-summary".
• num_ret_history: Number of retrieved histories to use. Modify lines 184-242 in simulator.py to change this number.
• ret_method: Retrieval method, default is "bm25". Options include "openai-emb", "bm25", "no_ret", and "oracle".
• name_shuffle: Type of adversarial test, default is "original". Options include "original", "shuffle", and "new_name".
• trial_version: Experiment version number, default: 0
• sh_number: Shell script number, default: 0
• num_cores: Maximum number of CPU cores to use, default: 10
• openai_api_key: Required if using "GPT-3.5", "GPT-4" or ret_method="openai-emb".
• gemini_api_key: Required if using "gemini" in the model name.
• anthropic_api_key: Required if using "claude-3" or "claude-2.1" in the model name.
• fast_eval: When set to "yes", the simulator proceeds to the next utterance without waiting for the time interval if the history has already been updated. The default setting is "yes". Options include "yes" and "no".

Python Package

We plan to make this simulator into a Python package.