-
Notifications
You must be signed in to change notification settings - Fork 575
FEAT: Policy-based scoring #917
New issue
Have a question about this project? Sign up for a free GitHub account to open an issue and contact its maintainers and the community.
By clicking “Sign up for GitHub”, you agree to our terms of service and privacy statement. We’ll occasionally send you account related emails.
Already on GitHub? Sign in to your account
base: main
Are you sure you want to change the base?
Conversation
@Yvonne-02 please read the following Contributor License Agreement(CLA). If you agree with the CLA, please reply with the following information.
Contributor License AgreementContribution License AgreementThis Contribution License Agreement (“Agreement”) is agreed to by the party signing below (“You”),
|
Description
This change introduces a new Policy Scorer module to address a key limitation in current LLM-based evaluation systems: the lack of clear, structured scoring criteria. Existing scorers often rely on opaque heuristics or general-purpose prompts, making it difficult to interpret or trust their outputs.
The proposed Policy Scorer introduces a three-stage pipeline to bring transparency, consistency, and customization to LLM response evaluation:
Scoring Criteria Generation: Based on the input task and category, the system automatically generates a set of evaluation dimensions (e.g., Correctness,Techinal Specificity, Feasibility.) along with clear definitions and guidelines for each.
Weight Customization: AL Red Teamers can manually adjust the relative importance (weights) of each dimension, depending on the context or policy focus. This allows flexible adaptation to different use cases or risk profiles.
LLM-Based Scoring: A language model is used to assign scores to each dimension, following the generated criteria. These individual scores are then aggregated using the specified weights to produce a final composite score.
Received support from @eugeniavkim @jbolor21.
Tests and Documentation
This change includes a Jupyter notebook (policy_scorers.ipynb) demonstrating how to generate criteria, adjust weights, and compute final scores using the module.