Generative Dialogue Model Automated Quality Assurance Tool

Description

This repository contains a framework for testing and evaluating GDMs. The visualizations in this readme shows results from testing different versions of the GDM named Emely adapted for interviews.

There are two steps in main.py:

1. Generating conversations.

• We divide the output into experiments with unique experiment ids.
• Each experiment contains a number of runs with numerical ids.
• Generated conversations are stored in test_data/{EXPERIMENT_ID}/run_{RUN_ID}.txt
• The configurations for all runs in an experiment are stored in test_data/{EXPERIMENT_ID}/experiment_config.json

2. Analyzing the conversations.

• Test results are stored in an SQL-database in test_results/{EXPERIMENT_ID}.sqlite
• The configuration for each run is contained in the table runs
• Each test-case is then imported into a separate table each.
• If we want to analyze conversations that have already been generated we can use the argument --read-run-ids to read these from the chosen .txt-files determined by the run_id.

How to run

Install dependencies

# clone project
git clone git@github.com:NordAxon/GDM-testing.git

# install project
cd GDM-testing
pip install -e .
pip install -r requirements.txt -f https://download.pytorch.org/whl/torch_stable.html

Testing is run by:

# run module
python main.py <OPTIONS> [PARAMETERS]

Run python main.py -h to have the options presented, or see below:

# options available
usage: main.py [-h] [-eid EXPERIMENT_ID] [-v] [-ec] [-od] [-cl] [-cs] [-rcs RANDOM_CONV_START] [-a] [-cp] [-t] [-im] [-rid]

Parser for setting up the script as you want

optional arguments:
  -h, --help                show this help message and exit
  -eid, --experiment_id     We divide all runs into experiments with a unique identifier.
  -v, --verbose             Use verbose printing.
  -ec , --export-channel    Specify which channel to export the results through. Currently only 'sqlite' is available.
  -od, --overwrite-db       Specifies if the result database should be overwritten during computation.
  -cl , --conv-length       How many replies from each GDM all conversations should contain.
  -cs , --conv-starter      Testee: testee initiates every conversation.
                            Conv-partner: the conversation partner initiates all conversations. Not specified: 50-50.
  -rcs, --random-conv-start Start conversations with a random reply.
  -a , --amount-convs       How many conversations shall there be per tested GDM.
  -cp , --conv-partner_id   Specify which GDM to run your testees against.
  -t , --testee-ids         Names of local docker images to use for each run, separated by ",".
  -im, --interview-mode     Conversations are initialized as interview scenarios.
  -rid , --read-run-ids     Run ids of the runs to import.
                            No input is interpreted as such the script generates conversations using the GDMs.
                            Currently only miscellaneous .txt-files are supported.

Visualise the results using Dash

1. Run python dashboard.py
2. If experiment results exist you can choose between experiments in the dropdown to the top left.
3. The graphs are shown in the right tab:

4. The run-configurations are shown in the left tab:

5. To add your own graphs implement this in visualization/graphs.py.

Citation

@article{JohanBengtsson2022,
  title={Quality Measurement of Generative Dialogue Models for Language Practice},
  author={Johan Bengtsson},
  year={2022}
}