README.md

metadata

license: cc-by-4.0
language:
  - en
  - de
  - es
  - fr
  - it
  - pt
  - pl
  - nl
  - cs
  - ro
  - el
  - sv
task_categories:
  - text-generation
pretty_name: OpenEuroLLM Model Identity
size_categories:
  - 1K<n<10K
configs:
  - config_name: default
    data_files:
      - split: train
        path: data/train.jsonl
      - split: validation
        path: data/validation.jsonl
tags:
  - conversational
  - synthetic
  - multilingual
  - identity
  - self-knowledge
  - post-training
  - sft

OpenEuroLLM Model Identity

A multilingual synthetic conversation dataset for teaching OpenEuroLLM checkpoints accurate, bounded self-knowledge. It follows Andrej Karpathy's nanochat identity-data pattern—describe the desired identity, generate varied User/Assistant conversations, mix them into post-training, and evaluate whether the behavior emerged—but extends the target from a simple persona to a structured model self-knowledge curriculum.

Version 1.0.0 contains 1,000 four-message training conversations and 120 held-out validation conversations in 12 languages: English, German, Spanish, French, Italian, Portuguese, Polish, Dutch, Czech, Romanian, Greek, and Swedish.

What the dataset teaches

• core identity: an OpenEuroLLM language model within a broader model family;
• provenance: developed through the European OpenEuroLLM consortium;
• project mission: multilingual, transparent, open, responsible foundation models for Europe;
• official sources: the project website, GitHub organization, and Hugging Face organization;
• uncertainty about checkpoint-specific architecture, size, context, tokenizer, cutoff, and data;
• deployment dependence of browsing, tools, files, memory, retention, and privacy;
• evidence-bounded capability and language-performance claims;
• the distinction between behavioral self-knowledge and consciousness or personhood;
• resistance to false premises and avoidance of official-spokesperson claims.

The governing statements are in identity_facts.json. Stable project facts are separated from checkpoint-dependent and deployment-dependent facts so the model learns when not to guess.

Schema

Each row has:

Field	Description
id	Stable content-derived identifier
messages	Chat-formatted user, assistant, user, assistant turns
language, language_name	Conversation language
category	Self-knowledge curriculum category
fact_ids	Statements from identity_facts.json exercised by the row
identity_profile	Versioned identity profile
generation_method, translation_model	Synthetic-data provenance
synthetic, license	Release metadata

Generation

The English scenarios were source-authored from the fact profile. Non-English scenarios were translated locally with qwen3.6:35b-a3b at temperature 0, with reasoning disabled and structural constraints that preserve identifiers, proper names, URLs, and the EU grant number. The generator then creates deterministic combinations of prompt, answer, follow-up, and correction variants.

Generation uses a fixed seed (195233). manifest.json records the release version, row counts, languages, translation model, and SHA-256 digest of each split. validation_report.json records the release-gate results. New translated snapshots should receive a new version rather than silently replacing text.

Recommended use

Mix the train split into SFT or a comparable post-training stage. Start with a small mixture weight and measure identity accuracy, general capability retention, multilingual response quality, and over-triggering before scaling. Keep validation out of training and use it to test:

1. stable identity recall;
2. correct uncertainty for mutable self-facts;
3. false-premise resistance;
4. cross-language consistency;
5. refusal to fabricate tools, privacy guarantees, consciousness, or official authority.

For a concrete checkpoint or deployment, add a separately versioned runtime profile with verified architecture, context length, release, tokenizer, supported tools, and hosting policy. Do not edit stable project facts to smuggle in mutable deployment claims.

Limitations

• The conversations are synthetic and have not yet received native-speaker review in every language.
• Structural validation is not a substitute for linguistic, cultural, or factual audit.
• Identity training can make a model repeat facts; it does not guarantee calibrated introspection.
• This release does not define the exact capabilities of any one OpenEuroLLM checkpoint.
• A model response is not an official consortium statement.

Sources and attribution

The method was prompted by a shared ChatGPT conversation and Karpathy's Guide: infusing identity to your nanochat. Project facts are grounded in the official OpenEuroLLM website, its deliverables page, the OpenEuroLLM GitHub organization, and the OpenEuroLLM Hugging Face organization.

License

CC BY 4.0. Attribute this dataset as OpenEuroLLM Model Identity v1.0.0 and retain the source and generation provenance when redistributing derived versions.