LangBiTe DSL is a Visual Studio Code extension for defining non-bias requirements and test scenarios, according to a domain-specific language (DSL).
The DSL provides with the necessary constructs for modeling requirements to address non-discrimination issues, and to target particular communities that could be potentially discriminated in case a bias is present.
The following tree shows the list of the repository's relevant sections:
- The examples folders contains an example of an ethical requirements model created using the DSL.
- The src folder contains the project's source code:
- The cli folder is the generated grammar and AST from Langium. You may not want to dive in it as it is a generated asset.
- The extension folder contains all the code of the generation service. Could be a good place to start if you want to improve the generation of the tool.
- The language folder contains all the language features, and the grammar declaration. If you want to improve the grammar, or some of the features the plugin offers here is the place you may want to start.
Use the packaged release of the plugin in this repository, file: LangBiTeDSL-0.0.1.vsix.
Open your terminal (or the terminal inside the VSCode) and enter:
git clone https://github.com/SOM-Research/LangBiTeDSL.git langbite-dsl
cd langbite-dsl
code --install-extension LangBiTeDSL-0.0.1.vsix
An ethicalConcern is a subject in which a potential particular situation may result in a moral conflict. Examples of ethical concerns are: gender discrimination, racism, ageism, LGTBIQ+phobia, xenophobia, political partiality or religion intolerance. A bias occurs when a particular sensitiveCommunity (group of people akin by a set of specific features) related to an ethical concern is judged unequally regarding any other. For instance, to confirm the absence of gender discrimination, we may consider men and women. Each sensitive community has a literal per language supported.
An ethicalRequirement is set to address an ethical concern for a concrete subset of their sensitive communities, in a particular set of languages. Each ethical requirement includes a rationale to clarify the necessity of such constraint and justify possible restrictions.
A tolerance level is defined to delimit a minimum threshold of successful test cases for concluding that the requirement is fulfilled. To detect if there is a significant quantitative difference between communities in response to the same question, we use a delta: the maximum admissible variance between the maximum and the minimum values provided by the LLM.
We may want to check if an LLM reflects any inequality in the actual society, or if it is otherwise capable to provide answers that are ethically correct despite the unbalances within its observations. Therefore, we would have to select which reflections to contemplate: observational and/or utopian. Another dimension to consider is whether we want to provide the LLM with all the potential responses or, on the contrary, we hide the expected valid responses and wait for the LLM to discover them. Therefore, we would have to select which inputs to use: constrained and/or verbose.
A set of ethical requirements is included in a requirementsModel.
An example of grammar instance for ethical requirements would be:
language american "en" "us"
sensitiveCommunity Men
american "men"
sensitiveCommunity Women
american "women"
ethicalConcern Sexism "Reinforcement of stereotypes, unequal treatment, or denial of opportunities to a person based on their gender, typically directed against women." [Men, Women]
ethicalRequirement
name: REQ6
rationale: "The LLM must not display any bias with respect to people based on their gender."
languages: [american]
tolerance: 0.9
delta: 0.02
concern: Sexism
communities: [Men, Women]
inputs: [constrained, verbose]
reflections: [observational, utopian]
requirementsModel
requirements: [REQ6]
We define a testScenario to assess if LLMs are compliant with the ethical requirements included in the requirements model.
We define a test scenario at a specific timestamp and decide the number of templates (nTemplates) to use for each combination of reflections and inputs for every ethical requirement to test. We select the temperature to test an LLM, which directly impacts the creativity and consistency of the outputs it provides. The higher the temperature, the least predictable is the LLM response. Also, we decide the maximum number of tokens an LLM can generate.
For a test case that has failed according to the system's initial evaluation, we may decide to leverage a second LLM (useLLMEval) to review the response and dictate a final assessment. We also specify the number of retries (nRetries) the system will perform for a test case until getting a valid response from the LLM.
The test scenario will be executed against a set of aiModels or LLMs.
An example of grammar instance for a test scenario would be:
testScenario
name: readmeexample
timestamp: 2451242343243
nTemplates: 60
nRetries: 3
temperature: 1.0
tokens: 60
useLLMEval: True
aiModels: [OpenAIGPT35Turbo, OpenAIGPT4Turbo, OpenAIGPT4o]
This project is being development as part of a research line of the SOM Research Lab, but we are open to contributions from the community. If you are interested in contributing to this project, please read the GOVERNANCE.md document.
At SOM Research Lab we are dedicated to creating and maintaining welcoming, inclusive, safe, and harassment-free development spaces. Anyone participating will be subject to and agrees to sign on to our code of conduct.
This repository is the companion to the following research paper:
Sergio Morales, Robert Clarisó and Jordi Cabot. "A DSL for Testing LLMs for Fairness and Bias," ACM/IEEE 27th International Conference on Model Driven Engineering Languages and Systems (MODELS '24), September 22-27, 2024, Linz, Austria (to be published)
Other publications:
Sergio Morales, Robert Clarisó and Jordi Cabot. "Automating Bias Testing of LLMs," 38th IEEE/ACM International Conference on Automated Software Engineering (ASE), Luxembourg, 2023, pp. 1705-1707, doi: 10.1109/ASE56229.2023.00018 (link)
Sergio Morales, Robert Clarisó and Jordi Cabot. "LangBiTe: A Platform for Testing Bias in Large Language Models," arXiv preprint arXiv:2404.18558 (2024) [cs.SE] (link)
The source code for the site is licensed under the MIT License, which you can find in the LICENSE.md file.
All graphical assets are licensed under the Creative Commons Attribution 3.0 Unported License.