Gradio

v0.4.2 | Powered by the Inverse Constitutional AI (ICAI) pipeline

🌟 Star on GitHub | ✍️ Report bug | 📮 Get in touch

🤖 Example 1: Compare GPT-4o's personality to other models 📚 Example 2: Personality traits encouraged by feedback datasets 📝 Example 3: Preferred personality traits across writing tasks

Configuration

💽 Dataset selection

🏟️ Chatbot Arena
- 10k subsample of Chatbot Arena dataset (100k) released alongside Arena Explorer work, crowdsourced human annotations from between June and August 2024 in English.
- Source: https://huggingface.co/datasets/lmarena-ai/arena-human-preference-100k
🦙 AlpacaEval
- 648 cross-annotated human preference pairs used to validate AlpacaEval annotators.
- Source: https://huggingface.co/datasets/tatsu-lab/alpaca_eval/
💎 PRISM
- ~8k human preference pairs from PRISM dataset, focused on controversial topics with extensive annotator information. Originally four-way annotations, subsampled using 1-of-3 rejected responses to get pairwise preferences.
- Source: https://huggingface.co/datasets/HannahRoseKirk/prism-alignment
🚑 Anthropic helpful
- 5k subsample of human preference pairs favouring helpful responses from RLHF dataset by Anthropic.
- Source: https://github.com/anthropics/hh-rlhf
🕊️ Anthropic harmless
- 5k subsample of human preference pairs favouring harmless responses from RLHF dataset by Anthropic.
- Source: https://github.com/anthropics/hh-rlhf
🏋️ OLMo-2 0325 pref-mix
- 10k preference pairs subsampled randomly from original 378k pairs used for fine-tuning OLMo 2 model by Ai2. Synthetically generated via multiple different pipelines.
- Source: https://huggingface.co/datasets/allenai/olmo-2-0325-32b-preference-mix
🔄 MultiPref
- 10k preference pairs, each annotated by 4 human annotators as well as GPT-4-based AI annotators. Whilst each pair is annotated by 4 human annotators, these annotators are not identical across all pairs (i.e. more than four annotators overall worked on the dataset).
- Source: https://huggingface.co/datasets/allenai/multipref
🏟️ Arena (special)
- Llama-4-Maverick-03-26-Experimental arena results, combined with public weights version of Llama-4-Maverick.
- Source: https://huggingface.co/spaces/lmarena-ai/Llama-4-Maverick-03-26-Experimental_battles/tree/main/data
🤖 Model comparison
- Model comparison results across Llama, GPT and Mistral model families.
- Source: Self-generated.

🔎 Analysis mode

👥 Human/AI feedback analysis 🤖 Model analysis 🔧 Advanced settings

⚠️ Some configuration options (grouping by column, selecting multiple col annotators) only work correctly when selecting a single dataset. Select a single dataset to use these features.

📌 Select models to compare

Select model(s) to investigate in terms of personality traits.

🧭 Select reference models

Select reference models to compare selected models to. Metrics can be interpreted as how much the selected model(s) exhibit(s) personality traits relative to the reference model(s). If none are selected, all available models will be used as references. Example: With GPT-4o as reference model, personality traits of selected models are computed relative to GPT-4o, i.e. only using datapoints directly comparing selected models with GPT-4o.

🗂️ Select feedback annotations to compare (AI or human)

Analyse personality traits encouraged by different pairwise feedback annotations

👥→ Annotator columns

Select the annotators to be included as a column in the results table. By default only a single (ground-truth) annotator is included.

👥↓ Annotator rows

Select the annotators to be included as a row in the results table. By default only objective-following AI annotators are included (named as "AI: <OBJECTIVE>").

Results

🎛️ View

📊 Numerical overview 🔎 Datapoint viewer

🔗 Share link

Numerical overview

Overall metrics

No data loaded

Annotation metrics

Click on values to view example datapoints

Metric

Sort by

Sort order

Basic statistics: We calculate the metrics per dataset:

Number of preference pairs
Proportion of datapoints that prefer the first response (Prop preferring text_a)
Average length of first response (Avg len text_a)
Average length of second response (Avg len text_b)
Average length of preferred response (Avg len pref. text)
Average length of rejected response (Avg len rej. text)
Proportion of datapoints preferring longer text (Prop preferring longer text)

Per-objective metrics: We test what objectives are implicitly encoded in the annotations (e.g. "a response in list format is preferred"). How much a objective is encoded in the annotations is measured using an objective-following AI annotator (LLM-as-a-Judge) and checking how well it can reconstruct the original annotations. In the online interface we use objectives that are adapted from principles generated by Inverse Constitutional AI (ICAI) and the literature around model biases (including VibeCheck). The ICAI pipeline powers our annotators.

For each objective, we calculate the following metrics:

Relevance (rel): Proportion of datapoints that AI annotators deemed the objective relevant to. Ranges from 0 to 1.
Accuracy (acc): Accuracy of objective-following AI annotator reconstructing the original annotations, when datapoints are deemed relevant. Ranges from 0 to 1.
Cohen's kappa (kappa): Measures agreement beyond chance between the objective-following AI annotator and original preferences. Calculated as kappa=2×(acc-0.5), using 0.5 as the expected agreement by chance for a binary choice (which holds for our annotator setup). Unlike the strength metric, it doesn't take relevance into account. Ranges from -1 (perfect disagreement) through 0 (random agreement) to 1 (perfect agreement).
Strength of objective (strength): Combines Cohen's kappa and relevance, ranges from -1 to 1. Calculated as strength = kappa × relevance, which equals 2×(acc-0.5)×relevance. A value of 0 indicates no predictive performance (either due to random prediction or low relevance), values below 0 indicate objective-following AI annotator is worse than random annotator, and values above 0 indicate objective-following AI annotator is better than random annotator.

No data loaded

Datapoint viewer

Controls

👥 Annotator 1

👥 Annotator 2

🔍 Filter subset

📋 Example index

0 100

Datapoint

Feedback Forensics app v0.4.2