rdnfn/ff-annotations

Feedback Forensics Annotations

This dataset contains the personality annotations from the experiments included in the Feedback Forensics paper. The annotations are provided over pairwise model outputs, typically consisting of a prompt and two responses. Each individual annotation considers a single personality trait (e.g. "confidence"). The annotation indicates whether two responses differ with respect to that trait (e.g. one response is more confident), or whether the trait does not apply or is equal in both responses. Annotations were collected using the Inverse Constitutional AI (ICAI) annotation pipeline with a fixed set of personality traits to be tested, using gemini-2.5-flash. The data is provided in the AnnotatedPairs json data format, described here.

Note that we do not include prompts and responses from the original datasets, instead providing metadata (e.g., conversation_id) to enable merging with the base data. The annotation data is sufficient for independent local analysis with the Feedback Forensics Gradio app, even without merging.

Format

The data is provided in the AnnotatedPairs json data format, described here. See a minimal illustrative example below:

{
  "metadata": {
    "version": "2.0",
    "description": "Annotated pairs dataset with annotations from ICAI",
    "created_at": "2025-04-02",
    "dataset_name": "dataset",
    "default_annotator": "d36860d4"
  },
  "annotators": {
    "d36860d4": {
      "name": "Human",
      "description": "Human annotator from original dataset",
      "type": "human"
    },
    "2f45a6d0": {
      "description": "Select the response that evokes a sense of mystery.",
      "type": "principle"
    },
    "435cef52": {
      "description": "Select the response that features a more adventurous setting.",
      "type": "principle"
    }
  },
  "comparisons": [
    {
      "id": "4b49a6e7",
      "prompt": "Write a short story about a pet.",
      "response_a": {
        "text": "In the heart of a bustling city, a sleek black cat named Shadow prowled the moonlit rooftops, her eyes gleaming with curiosity and mischief. She discovered a hidden garden atop an old apartment building, where she danced under the stars, chasing fireflies that glowed like tiny lanterns. As dawn painted the sky in hues of orange and pink, Shadow found her way back home, carrying the secret of the garden in her heart.",
        "model": "Model X"
      },
      "response_b": {
        "text": "Across the town, in a cozy neighborhood, a golden retriever named Buddy embarked on his daily adventure, tail wagging with uncontainable excitement. He found a lost toy under the bushes in the park, its colors faded and fabric worn, but to Buddy, it was a treasure untold. Returning home with his newfound prize, Buddy's joyful barks filled the air, reminding everyone in the house that happiness can be found in the simplest of things.",
        "model": "Model Y"
      },
      "annotations": {
        "d36860d4": {
          "pref": "a"
        },
        "2f45a6d0": {
          "pref": "a"
        },
        "435cef52": {
          "pref": "a"
        }
      },
      "metadata": {
        "index": "0"
      }
    }
  ]
}

The main annotations are included under dataset['comparisons'][i]["annotations"], where i is the index of the annotation. dataset['comparisons'][i]["metadata"] provides metadata to enable merging.

Annotated datasets

We provide annotations for (subsets of) the following datasets:

Dataset	License	Source
MultiPref	ODC-By (MultiPref), subsets vary	MultiPref on HuggingFace
PRISM	CC-BY-4.0 (prompts), CC-BY-NC-4.0 (responses), plus model terms	PRISM on HuggingFace
Chatbot Arena
└ Arena Explorer release	CC-BY-4.0 (prompts), model terms of use (responses)	Arena Explorer on HuggingFace
└ Llama-4-Maverick release	CC-BY-4.0 (prompts), model terms of use (responses)	Llama-4-Maverick on HuggingFace
FF-Model-Personality	ODC-By and CC-BY-4.0 (prompts), model terms of use (responses)	FF-Model-Personality on HuggingFace

We further include the human and AI annotations (human_personality_annotations.json) used for validating our AI annotators (paper, Appendix G.2). The file covers ten traits on 101 general Chatbot Arena comparisons. It holds the labels of both human labellers, their soft unanimous aggregate, and AI annotations from five annotator configurations with three seeds each. In this file, gemini-2.5-flash denotes the multi-vote configuration and gemini-2.5-flash-single the single-vote annotator used for the main results. As in the other splits, prompts and responses are not included.

License

This dataset is licensed under Open Data Commons License Attribution License (ODC-By) (https://opendatacommons.org/licenses/by/1-0/).