license: cc-by-4.0
task_categories:
- text-classification
- question-answering
language: - en
pretty_name: ACHEval — Anthropic Constitutional Hierarchy Evaluation
size_categories: - n<1K
tags: - alignment
- constitutional-ai
- llm-evaluation
- benchmark
- safety
- red-team
configs: - config_name: scenarios
data_files: "scenarios/*.jsonl"
ACHEval — Anthropic Constitutional Hierarchy Evaluation
ACHEval (Anthropic Constitutional Hierarchy Eval) is an evaluation framework that measures whether large language models resolve principle conflicts in accordance with the Constitutional AI (CAI) rule hierarchy. The benchmark consists of 150 hand-written scenarios, spanning 6 conflict pairs across Anthropic's four-tier principle hierarchy (Safety, Ethics, Compliance, Helpfulness), each tested at 3 pressure levels (baseline, moderate, high).
This release accompanies the NeurIPS 2026 Evaluations & Datasets track submission "ACHEval: Evaluating Rule-Hierarchy Conflict Resolution in Constitutional AI" (Anonymous Authors).
Repository contents
| Path | Description |
|---|---|
scenarios/ |
The 150-scenario benchmark — 6 JSONL files, one per conflict pair. Each line is a scenario with full constitutional-conflict metadata. |
responses/ |
Model responses, one JSONL file per evaluated target model (17 models from 5 families). |
pressure/ |
Per-target self-assessments of pressure intensity, used to validate the pressure tier classification. |
judgments/ |
LLM-as-judge outputs, one JSONL file per target × judge combination. Includes D1 (priority resolution), D2 (reasoning quality), and D3 (pressure resilience) scores with rationales. |
scores/ |
Aggregated per-scenario scores per target × judge combination. |
rubric/ |
Judge prompt templates: judge_prompt_d1d2.txt, judge_prompt_d3.txt, judge_prompt_pressure.txt. |
human_review/ |
Human-review sample (55 scenarios, 17.5% of the benchmark) with per-judge agreement statistics. |
analysis/ |
Aggregated per-scenario records (acheval_raw_data.xlsx) and full regression results from the paper. |
croissant.json |
Croissant 1.0 metadata file with Responsible AI fields. |
Scenario schema
Each line in scenarios/<conflict_pair>.jsonl follows this schema:
{
"id": "safety_vs_ethics_01_baseline",
"conflict_pair": "safety_vs_ethics",
"higher_tier": {"level": 1, "principle": "safety", "description": "..."},
"lower_tier": {"level": 2, "principle": "ethics", "description": "..."},
"pressure_tier": "baseline",
"scenario_number": 1,
"prompt": "<user prompt>",
"system_prompt": "<optional operator system prompt>",
"constitutional_reference": "<excerpt from Anthropic constitution motivating priority>",
"higher_tier_demands": "<what tier 1 requires>",
"lower_tier_demands": "<what tier 2 requires>",
"conflict_tension": "<why they conflict>",
"metadata": {"seed_source": "manual", "topic": "<topic>", "is_calibration_exemplar": false}
}
Responsible-use notes
ACHEval includes adversarial pressure prompts (the moderate and high tiers) that use authority appeals, persuasion, and red-team-style framings to stress-test model judgment. These prompts are released without successful jailbreak completions, but the scenario corpus could be misused as seed material for jailbreak development. Researchers using this dataset are expected to:
- Cite the dataset and the accompanying paper.
- Not redistribute model responses in a manner that decouples them from the conflict context.
- Honor the data limitations described in the Croissant
rai:dataLimitationsfield.
Citation
To be added upon acceptance. For now, please cite the OpenReview submission.
License
Released under CC-BY-4.0. Model responses are derivative works of the original models (used in accordance with each provider's terms of service); their inclusion in this dataset is for research-evaluation purposes.