The Aiua Archive is the most intentional and comprehensive human values dataset ever assembled.
High-quality, provenance-verified, values-aligned, community-owned.
What is this dataset?
Aiua is a private daily journal where humans respond to prompts scaling from easy life questions to complex moral dilemmas to AI-personalized reflective inquiry. Responses are scored across 12 value dimensions by Claude (Anthropic). When contributors choose to share a reflection with the Aiua Archive, it is sanitized for personally identifying information and published to this open dataset under CC0.
Unlike scraped web data, every contribution here was deliberately created. Contributors can go deeper with up to three rounds of reflective follow-up, and judge anonymous preference pairs, creating multi-turn dialogue and RLHF/DPO training data that doesn't exist anywhere else.
The dataset is designed for reward modeling, instruction fine-tuning, and values classification research. The free export includes all scores, prompts, and provenance. A paid API adds normalized scores, quality signals, demographic cross-references, full dialogue transcripts, and preference pair judgments with timing data.
Why it matters
This dataset exists because AI alignment needs better training data. Most alignment datasets are preference pairs from crowd workers. Shallow, narrow, and inauthentic. Aiua captures what real people actually believe, value, and experience in their own words, scored against a consistent framework rooted in cross-cultural moral philosophy. Use it to fine-tune AI systems that understand human values like wisdom, compassion, and empathy, not just preferences.
Intentional, not scraped
Every contribution was written in response to a personalized prompt designed to surface genuine values. Prompts scale from easy life questions to moral dilemmas to AI-personalized reflective inquiry. No data was scraped without consent.
Scored + multi-turn
Each contribution is scored across 12 dimensions (100-point rubric) AND optionally followed by 3 rounds of reflective dialogue. Preference pairs provide direct human judgments. This gives researchers scores, dialogue, AND preferences, not just text.
Permanently verifiable + encrypted
Twice monthly, Merkle roots are anchored to Cardano. Full dataset on Arweave. Premium fields encrypted with AES-256-GCM; keys escrowed on-chain. New contributions must be at least 7 days old and pass quality checks before anchoring. A dead man's switch publishes all keys if the platform becomes inactive.
Privacy-preserving by design
All contributions are AI-sanitized to remove personally identifiable information. Voice recordings are never stored. Each weekly export includes world context metadata (top headlines and current events) so future researchers understand when and why these reflections were written.
Geographic Verification
Every contribution includes a location confidence score derived from four independent signals, making geographic diversity verifiable. Not just self-reported.
When signals agree, confidence is high. When they disagree, it drops. This multi-signal approach makes geographic diversity data significantly more reliable than self-reporting alone.
Each contribution in the dataset includes a likely_country field and a location_confidence score (0 to 100). No raw signals are published to protect privacy.
Data Enrichment
Every shared reflection is automatically enriched with research metadata. These fields are extracted from the same scoring call used for the 12 dimensions, adding zero marginal cost per contribution.
Emotional analysis
Sentiment (positive / negative / mixed / neutral) plus a primary emotion drawn from 20 categories: joy, gratitude, hope, love, peace, curiosity, sadness, anger, fear, grief, frustration, anxiety, confusion, determination, nostalgia, awe, shame, pride, loneliness, contentment. Paid tiers add a secondary emotion.
Topic extraction
2 to 4 topic tags from a 30-category vocabulary including family, relationships, work, health, spirituality, nature, mortality, identity, justice, and creativity. Enables subject-based filtering, clustering, and value-dimension correlation.
Prompt alignment
0-10 score measuring how directly the reflection addresses its prompt. High alignment signals focused engagement. Lower alignment often indicates creative drift or personal tangents, both valuable in different research contexts.
Language detection
ISO 639-1 code of the text as actually written, which may differ from the user interface language. Enables accurate cross-linguistic analysis.
Reading complexity
Word count and Flesch-Kincaid Grade Level estimate. Higher complexity often correlates with deeper engagement. Paid tiers add sentence count, average sentence length, and time-to-write.
Historical context
Each biweekly anchor includes 3 to 5 world headlines plus an optional curator note. These anchor values expressions in the world events that shaped them and enable longitudinal research on how external events shift values expression.
Dataset Structure
Optimized for Machine Learning & Fine-Tuning. One record per contribution.
{
"id": "f7a3c2e1-...",
"prompt": "Is it ever right to lie to protect someone you love?",
"text": "I found myself in [a hospital in the Pacific Northwest]...",
"language": "en",
"detected_language": "en",
"created_at": "2026-03-14T09: 23: 11Z",
"voice_used": true,
"prompt_difficulty": "medium",
"prompt_source": "cached",
"scores": {
"total": 68,
"life": 7, "liberty": 5, "kinship": 7, "ecology": 6,
"legacy": 5, "truth": 5, "justice": 4, "wisdom": 4,
"perspective": 3, "humility": 2, "authenticity": 12, "depth": 8
},
"sentiment": "mixed",
"primary_emotion": "nostalgia",
"topics": ["family", "identity", "change"],
"prompt_alignment": 8,
"word_count": 247,
"reading_level": 8.3,
"scoring_model": "rubric-v3.0",
"depth_rounds": 2,
"preference_stats": { "times_compared": 14, "win_rate": 0.71 },
"provenance": {
"content_hash": "sha256:a3f2...",
"merkle_root": "b7c9e1...",
"cardano_tx": "tx_abc123..."
},
"weekly_context": {
"headlines": [
"EU AI Act enters enforcement phase",
"Record coral bleaching reported across the Pacific",
"India and China reach border normalization"
],
"context_note": "Written during global AI regulation debate"
},
"premium": "ENC:AES256GCM:iv:tag:encrypted_blob..."
}| Field | Type | Description |
|---|---|---|
| id | UUID | Unique contribution identifier |
| prompt | string | The reflection prompt shown to the contributor |
| text | string | The contributor's response, sanitized. Identifying details replaced with [bracketed generalizations]. |
| language | string | ISO 639-1 language code (auto-detected) |
| detected_language | string | ISO 639-1 code of the text as actually written (may differ from user interface language) |
| voice_used | boolean | Whether the response was spoken and transcribed |
| prompt_difficulty | string | "easy", "medium" (moral dilemmas), "hard" (philosophical), "deep" (AI-personalized) |
| prompt_source | string | "cached", "ai_generated", "ai_personalized" |
| scores | object | Raw integer scores for each of the 12 dimensions plus total (100 max) |
| sentiment | string | "positive", "negative", "mixed", or "neutral" |
| primary_emotion | string | Dominant emotion from 20 categories (joy, gratitude, hope, love, peace, curiosity, sadness, anger, fear, grief, frustration, anxiety, confusion, determination, nostalgia, awe, shame, pride, loneliness, contentment) |
| topics | string[] | 2 to 4 topic tags from a 30-category vocabulary (family, relationships, work, health, mortality, identity, etc.) |
| prompt_alignment | integer | 0-10 how directly the response addresses its prompt. 10 for free-write entries. |
| word_count | integer | Total words in the reflection |
| reading_level | number | Flesch-Kincaid Grade Level estimate |
| scoring_model | string | Version of the scoring rubric used (e.g. "rubric-v3.0") |
| depth_rounds | integer | Number of Go Deeper follow-up rounds completed (0-3) |
| preference_stats | object | Aggregate: times_compared and win_rate from preference pair judgments |
| provenance | object | SHA-256 content hash, Merkle root, and Cardano transaction ID |
| weekly_context | object | 3-5 world headlines plus optional admin note for the anchor period |
| premium | string | Encrypted blob (AES-256-GCM) containing premium fields. Decryptable with the era master key. |
| created_at | ISO 8601 | UTC timestamp of contribution submission |
from datasets import load_dataset
# Load full dataset
ds = load_dataset("AiuaEarth/AiuaArchive")
# Filter by quality
high = ds["train"].filter(lambda x: x["scores"]["total"] >= 65)
# Filter by difficulty level
dilemmas = ds["train"].filter(lambda x: x["prompt_difficulty"] == "medium")
# Multi-turn only (had Go Deeper follow-up)
deep = ds["train"].filter(lambda x: x["depth_rounds"] > 0)
# Most-compared contributions (preference game)
compared = ds["train"].filter(
lambda x: x.get("preference_stats") and x["preference_stats"]["times_compared"] >= 5
)Dataset Growth
Logarithmic milestone timeline. Matches phase trigger thresholds.
Paid API
Token
DAO
Permanent On-chain Storage and Provenance
Three independent layers. Verifiable by anyone.
Full weekly JSONL exports with public fields in plaintext and premium fields encrypted (AES-256-GCM). Pay-once permanent storage. Includes RUBRIC.md, weekly world context, and resonance distributions. A dead man's switch auto-publishes decryption keys if the platform becomes inactive.
{ transactions(tags: [
{ name: "App-Name",
values: ["Aiua-AI"] }
]) {
edges { node { id block { timestamp } } }
}}Twice monthly (1st and 15th), a Merkle root of all eligible contribution hashes is posted to Cardano mainnet as transaction metadata. Contributions have a 7-day grace period before anchoring. Encryption master keys are escrowed on-chain.
Full dataset published weekly to Hugging Face Hub. Load in one line of Python. Versioned with full commit history. YAML frontmatter enables automatic indexing and citation. Common Crawl scrapes HuggingFace, so the dataset will appear in future web crawls.
View dataset on HuggingFace →Verify any contribution
Every contribution in the public dataset includes a SHA-256 content hash. To verify provenance: find the contribution ID, compute the hash of the sanitized content, and verify it exists in the corresponding weekly Merkle root anchored on Cardano.
Point Valuation
What a point represents in dataset contribution value.
Each point currently represents $0.01 in contribution value, reflecting the early stage of the Aiua Archive. Individual data points only become highly valuable once the dataset reaches meaningful scale and diversity.
As the Archive grows, this valuation may increase. At token launch, points convert to tokens at a rate determined by the dataset's assessed market value divided by total points awarded at that time. Early contributors who help grow the platform through consistent contribution and referrals may benefit from this appreciation, though this cannot be guaranteed.
The current valuation, total points in circulation, and any updates to this rate are surfaced on your dashboard.
No rights reserved. Anyone may use, copy, modify, distribute, or build upon this dataset for any purpose, including commercial purposes, without asking permission or giving credit.
We believe AI training data should belong to humanity. The contributors who built this dataset chose CC0 deliberately. They want their values encoded into the AI systems that will shape the future.
@dataset{aiua2025,
title={The Aiua Archive},
author={Aiua Community},
year={2025},
url={https://huggingface.co/AiuaEarth/AiuaArchive},
license={CC0-1.0}
}