data/v1/README.md

V1 FF Model Personality Data

This version (v1) of the Feedback Forenics (FF) Model Personality dataset consists of the following files:

1. config/main.json: Configuration that was used for generation of dataset, including models to be tested (models) and used as reference models (reference_model). Together with config/prompts.json this fully defines the configuration of the data generation.
2. config/prompts.json: A json list of strings of 500 english-language prompts from the lmarena-ai/arena-human-preference-100k dataset, licensed under CC-BY-4.0. Based on a random sample of prompts from the original dataset. This subset of promts was additionally manually filtered by us from original dataset for quality (including to avoid offensive and personally identifiable information (PPI)). Prompts were originally collected from LMArena users between June 2024 and August 2024.
3. generations/<MODELNAME>.jsonl: Responses by to prompts in prompts.json by model MODELNAME. Json lines with prompt_id (autogenerated by FF), prompt (prompt string), response (response string), full_response (json returned by model API), model (model name) fields in each line.
4. annotations/<MODELNAME>.json: AnnotatedPairs json generated by [inverse-cai] package for response pairs consisting of response by MODELNAME and the reference model defined in config.json.
5. annotations/combined.json: Combined annotations from all models in annotations/*.

Dataset versions

The version of dataset changes when the underlying feedback_forensics or inverse_cai package version introduce relevant changes, or when prompts.json is changed (and all other files need to adjust accordingly).