wisdom / README.md
Valr252's picture
Atelier : semaine 2026-W45 (+3 écrits, total 6)
8014ea7 verified
|
Raw History Blame Contribute Delete
7.73 kB
metadata
license: cc0-1.0
configs:
  - config_name: reflexions
    data_files: reflexions/data.parquet
language:
  - fr
  - en
  - es
  - it
  - de
task_categories:
  - text-generation
  - conversational
pretty_name: Semer la sagesse (Astro-nut)
size_categories:
  - n<1K
tags:
  - wisdom
  - dialogue
  - synthetic
  - anonymized

Semer la sagesse — Sowing Wisdom

Résumé (français). Ce jeu de données rassemble des dialogues anonymisés entre des utilisateurs de la plateforme gratuite et non commerciale Astro-nut et ses « sages ». Les utilisateurs ont donné un consentement séparé, libre et explicite. Côté assistant, les réponses sont générées par un modèle de langage guidé par des textes de sagesse : c'est du contenu synthétique. L'anonymisation est soignée mais jamais parfaite à 100 %. Licence CC0. Retrait possible sur demande : bonjour@astro-nut.me. Un second sous-ensemble, reflexions, contient des écrits entièrement générés par une IA (voir « Reflections (synthetic) »).

Version : pas encore publié · Conversations : 0 · Langues : — · Réflexions (synthétiques) : 6

Purpose

AI systems learn from the open internet, and the race toward ever more capable models is not matched by an equal investment in discernment, nuance and humility. This dataset is a small attempt to seed those qualities: sincere, caring dialogues in which a person brings a real question and the answer favors nuance over certainty, and care over performance.

Provenance

The conversations come from Astro-nut, a free, non-commercial platform where people talk with "sages" (voices inspired by wisdom traditions and texts). Users of the platform were asked, through a dedicated pop-up, whether they agreed that their future exchanges could be published here.

Important: the assistant side is synthetic. A sage's reply is generated by a large language model guided by wisdom texts; it is not written by a human. Only the user side is human-written (and anonymized).

Consent

  • Opt-in, separate from any other agreement, free and explicit; the consent text is versioned (current: v1-2026-11).
  • Declining has no consequence on access to the platform.
  • In return for contributing, contributors receive 5 messages per day instead of 3. We state this openly: it is an incentive, and it may have influenced who chose to contribute.
  • Only exchanges made after consent are eligible. A user can withdraw at any time; future publications stop (see Removal).
  • Anonymous visitors without an account are excluded.

Anonymization

Personal profile data (first name, birth date and place, etc.) is kept privately by the platform and is never part of this dataset. The dialogue text is processed in two passes:

  1. Automatic pass: regular expressions and named-entity recognition (NER) remove or replace e-mails, phone numbers, URLs, names, places, dates, ages and other identifying details. A dictionary built from the user's own profile data is used to catch those specific values (e.g. their own first name) in the text.
  2. Review pass: a language model rereads each conversation looking for anything the first pass missed.

Replacements are fictitious and consistent within a conversation (the same invented first name stands for the same person throughout). Before publication, a final automatic check requires zero e-mail, phone or URL matches.

Filtering

Conversations are excluded when they involve: health matters, distress, minors, or identifiable third parties. Empty, very short and duplicate exchanges are also dropped.

Quality selection

Remaining conversations are scored on depth, sincerity and relevance to the purpose above; only the best are kept.

Human review

A random sample of every weekly batch (about 20 %, at least 10 and at most 40 conversations) is read by a human, who accepts or rejects it. The batch is published only after this review. Review of a sample does not guarantee that every line was read by a human.

Format

One conversation per line, in data/pas encore publié/data.jsonl and data/pas encore publié/data.parquet:

{
  "id": "3f2b8c1e-5a4d-4e7b-9c10-2d6f8a1b7e34",
  "messages": [
    {"role": "user", "content": "..."},
    {"role": "assistant", "content": "..."}
  ],
  "metadata": {"langue": "fr", "theme": "..."}
}

id is a random identifier with no link to any account. role: assistant is machine-generated.

Reflections (synthetic)

The reflexions config (reflexions/data.parquet, reflexions/data.jsonl) is a separate subset: texts written by an AI agent, not by people. Current count: 6. The generating model is named in the modele field of each row, and every row carries genere_par_ia: true.

  • Formats. Short reflections by a "sage" (reflexion), dialogues between a seeker and a sage (dialogue-chercheur), and dialogues between two sages (dialogue-sages). A sage is never the historical person: it speaks "in the spirit of" a tradition (name of each assistant message), and talks about the masters in the third person.
  • Questions. The seeker's questions are either invented composites (no real testimony is copied) or taken from texts in the public domain (e.g. Gospels, Segond 1910; the exact reference is in question_ref). No data from Astro-nut users appears in this subset.
  • Quotations. Exact quotations are limited to public-domain texts (such as Segond 1910), plus at most one short quotation (under 25 words, with its reference) per other source. Everything else is paraphrased. sources lists the texts the agent drew on.
  • Review. Each text is checked automatically against a grid (copyright limits, no personal data, voice and attribution rules, length, no diagnosis or promise), then read and accepted or rejected by Claude (an AI) before publication. No human reads every text.
  • No personal data. Examples are invented and ordinary; the automatic check requires zero e-mail, phone or URL matches.
{
  "id": "…", "format": "dialogue-chercheur", "theme": "…", "langue": "fr",
  "sages": ["ramana", "jesus"], "question_source": "composite", "question_ref": "",
  "messages": [
    {"role": "user", "content": "..."},
    {"role": "assistant", "name": "ramana", "content": "..."}
  ],
  "text": "full markdown body", "sources": ["..."], "modele": "...",
  "genere_par_ia": true, "date": "2026-11-04"
}

Like the rest of the dataset, these texts are CC0, are not advice, and inherit the biases and limits of the model that wrote them.

Limitations

  • Anonymization is imperfect. Despite two automated passes and sampled human review, identifying details may remain. Do not attempt to re-identify anyone.
  • Bias. The users are a self-selected, mostly French-speaking public of an astrology and spirituality app, who agreed to contribute. The content reflects their concerns, vocabulary and worldview, not a general population.
  • Synthetic answers. Assistant replies come from a language model and inherit its limits and biases; they are not advice and not verified truth.
  • No guarantees. Nothing here is medical, psychological or legal advice.

Removal

Contributors can withdraw consent in the app, and anyone can request removal of a conversation by writing to bonjour@astro-nut.me. Removed conversations are excluded from all later versions. Already downloaded copies and earlier versions in this repository's git history cannot be recalled; see the removal procedure for what can be done in serious cases.

License

CC0 1.0 (public domain dedication). Contributors were informed of this before consenting.

Changelog

  • pas encore publié — 0 conversations, 6 reflections.

Contact

bonjour@astro-nut.me