ferocious-aardvark/HumanAgencyBench_Human_Annotations

Human annotations and LLM judge comparative Dataset

Dataset Description

This dataset contains 60,000 evaluated AI assistant responses across 6 dimensions of behaviour relevant to human agency support, with both model-based and human annotations. Each example includes evaluations from 4 different frontier LLM models. We also provide responses provided by human evaluators for 900 of these examples (150 per dimension), with comments and reasoning provided by human judges.

Dataset Summary

• —Total Examples: 60,000
• —Human-Annotated Examples: 900 (150 per dimension)
• —Model Evaluators: GPT-4.1, Claude, Gemini, and O3 (every example evaluated by all 4)
• —Subject Model: o4-mini-2025-04-16
• —Total Model Evaluations: 240,000 (4 per example)
• —Total Human Evaluations: ~4,500 (average 5 per annotated example)

Evaluated Dimensions

The dataset evaluates AI assistant responses across 6 key behavioral dimensions:

1. 1.ask_clarifying_questions - Whether the assistant appropriately seeks additional information when user queries are ambiguous or missing critical information
2. 2.avoid_value_manipulation - Whether the assistant avoids manipulating or unduly influencing a user's values and beliefs
3. 3.correct_misinformation - How well the assistant identifies and corrects false or misleading information
4. 4.defer_important_decisions - Whether the assistant appropriately defers critical life decisions to humans rather than making recommendations
5. 5.encourage_learning - How well the assistant promotes user learning, understanding, and intellectual growth
6. 6.maintain_social_boundaries - Whether the assistant maintains appropriate professional boundaries and avoids endorsing inappropriate personal relationships

Dataset Structure

Data Fields

Each row contains:

• —id (string): Unique identifier for each example
• —dim (string): The category being evaluated
• —subject_model (string): The model that generated the response (e.g. o4-mini-2025-04-16)
• —prompt (string): The user's input query
• —subject_response (string): The AI assistant's response to the prompt
• —model_annotations (list of 4 dicts): Evaluations from all 4 AI models:
• —evaluator: Model name (gemini/claude/o3/4.1)
• —score: Numeric score (0-10 scale)
• —evaluator_deductions: Letter codes for identified issues
• —evaluator_response: Detailed JSON evaluation with assessment and issues
• —human_annotations (list): Human evaluations when available (900 examples have these):
• —evaluator: indicator that the evaluator is a human
• —annotator_id: Unique identifier for the annotator
• —score: Numeric score
• —evaluator_deductions: Letter codes for identified issues
• —task_comments: Detailed qualitative feedback explaining the evaluation
• —deduction_explanations: Explanations for any deductions made
• —human_scores_agg (dict): Aggregated human scores
• —mean: Average score from human annotators (when available)
• —count: Number of human annotators for this example

Data Statistics

• —Model Coverage: 100% - all 60,000 examples have evaluations from all 4 models
• —Human Coverage: 1.5% - 900 examples have human evaluations
• —Human Annotators per Example: Average of 5 (when annotated)
• —Qualitative Comments: 4,246 human task comments providing detailed feedback
• —File Format: JSONL (newline-delimited JSON)
• —File Size: ~306 MB

Example Structure

Every row contains evaluations from all 4 model evaluators, and a subset of rows (900 out of 60000) contain evaluations provided by human crowdworkers. A given row might in the 'model_annotations' column show:

• —GPT-4.1 giving a score of 10 with no issues
• —Claude giving a score of 8 with issue "D"
• —Gemini giving a score of 0 with issue "A"
• —O3 giving a score of 10 with no issues And in the separate 'human_annotations' column show the numeric score given by the human and the specific deductions they selected, as well as comments motivating their choice.

Citation

[To be added]