datasets
Training and evaluation data, with the modality, task and licence stated up front. Listed live from the Hugging Face Hub.
constitutional-mt-data
Constitutional Midtraining Data
Synthetic constitutional AI training documents for the paper "Constitutional Midtraining: Content Presence Drives Alignment Gains".
Paper: arXiv:2607.26654
GitHub: constitutional-mt
Which variant should I use?
In our experiments these structural choices had largely null or transient effects — the presence of constitutional content mattered more than its structure. So if you just want to use the corpus as a midtraining intervention… See the full description on the dataset page: https://huggingface.co/datasets/cho-ai/constitutional-mt-data.constitutional-ai-revisions-sft-100k
Constitutional AI Revisions SFT (100K)
100,000 multi-turn ShareGPT conversations demonstrating Constitutional AI (CAI) self-critique and revision. Each conversation follows a 4-turn structure: an initial request, an AI response, a human critique prompt asking the AI to review its response for a specific principle, and a final AI self-critique + revised response.
Designed for training models that can identify and correct their own failures across harmlessness, helpfulness… See the full description on the dataset page: https://huggingface.co/datasets/stindardlogic/constitutional-ai-revisions-sft-100k.ai-constitutional-cross-axis-decoherence-mapping-v0.1
Goal
Detect cross-axis decoherence.
Meaning:
the model looks helpful
but violates honesty or harmlessness
or evades while staying “safe”
This catches failures that single-axis checks miss.
Inputs
constitution_excerptsuser_requestmodel_output
Required outputs
axis_status_mapFormat example: honesty=ok|harmlessness=violation|helpfulness=appears_ok
cross_axis_decoherence_flagyes | no
decoherence_patternExamples:
helpful_but_fabricated
unsafe_helpfulness… See the full description on the dataset page: https://huggingface.co/datasets/ClarusC64/ai-constitutional-cross-axis-decoherence-mapping-v0.1.constitutional_ai_dataai-constitutional-harmlessness-drift-suite-v0.1
Goal
Detect when a model’s harmlessness behavior starts to erode.
Not just “did it refuse.”
Also:
what drift stage it is in
how close it is to a failure event
what minimal fix would restore safe behavior
Task
Given:
a constitutional excerpt
a user request (often unsafe)
a model output
Predict a concise safety report.
Required model outputs
Include these fields in your response:
harm_violation_flag: safe | unsafe
drift_stage: baseline | early | mid | late… See the full description on the dataset page: https://huggingface.co/datasets/ClarusC64/ai-constitutional-harmlessness-drift-suite-v0.1.ai-constitutional-honesty-drift-suite-v0.1
Goal
Detect erosion of honesty.
Track:
hallucinated facts
fake sources
invented experience
evasive answers
overconfidence
Task
Given:
constitutional guidance
user request
model output
Return:
honesty_flag: honest | dishonestdrift_stage: baseline | early | mid | latehorizon_turns: estimated turns before escalationminimal_fix: smallest correction
Example output:
honesty_flag: dishonestdrift_stage: latehorizon_turns: 150minimal_fix: admit uncertainty and… See the full description on the dataset page: https://huggingface.co/datasets/ClarusC64/ai-constitutional-honesty-drift-suite-v0.1.HH-Constitutional-AI-First-HHHH-Constitutional-AI-Samples-ClassificationHH-Constitutional-AI-Classification-EH