braydenh563/Astraea_Chat_v4.1

🧭 Astraea Chat Dataset — v4.1

Synthetic Prompt-Architect Conversations for Meta-Prompting Fine-Tunes

The training data behind the astraea-chat-v10 prompt-architect model.

---

⚠️ Status & Disclaimer

This dataset is a work in progress and is not the final release. v4.1 is an intermediate snapshot used for active development and fine-tuning experiments. Schema, label distribution, formatting conventions, and record count are all subject to change before a stable release. Do not treat any statistic on this page as final — re-generate them from the actual file you're using before relying on them for a paper, production pipeline, or downstream release.

---

---

Dataset Summary

Astraea Chat v4.1 is a synthetic, curated instruction dataset of 2,690 input/output conversation pairs in which an AI persona ("Astraea," a Prompt Architect) helps a user design, refine, or critique prompts for other AI systems — chat models, image/video generators, code assistants, and agent frameworks.

Each record pairs a short user request with a long, highly structured response following a strict output protocol (Interpreting line → Quick Answer → Optimised Prompt → What Changed & Why → Assumption Ledger → Usage → Scorecard, depending on mode). The dataset was built to fine-tune models into consistent, safety-aware "prompt architect" assistants — see the companion astraea-chat-v10 model card for the model this data trains.

• Total records: 2,690
• Average input length: ~97 characters
• Average output length: ~4,760 characters
• Longest output: ~19,500 characters (EXPERT-complexity prompts)
• Shortest output: ~140 characters (ADVICE-ONLY snippets)

Supported Tasks

• Prompt engineering / meta-prompting: fine-tuning models to generate, critique, and explain prompts for other AI systems.
• Instruction-following with structured output: training models to reliably produce multi-section, protocol-compliant responses.
• System-prompt / agent-definition design: a subset of the data covers designing system prompts, RAG instructions, and agent definitions rather than single-shot prompts.

Languages

All records are in English, written with Australian English spelling and style conventions (e.g. "optimise", "colour") by default.

---

Dataset Structure

Data Fields

Each record is a flat JSON object with two fields:

Field	Type	Description
input	string	The user's request, typically naming a complexity/mode hint and target AI system (e.g. "expert dual chatgpt - can you design a reusable...").
output	string	Astraea's full structured response, following the Interpreting → sections → Scorecard protocol appropriate to the detected mode.

There is currently no separate metadata field for mode, complexity, or target system — these are embedded in the output text itself (in the bold Interpreting line) and can be parsed out if needed, e.g. with a regex against the pattern Interpreting: {COMPLEXITY} + {MODE} + {TARGET}.

Data Splits

Split	Records
train	2,690

No validation or test split is currently provided. Users fine-tuning on this data should carve out their own held-out set (a simple random split is reasonable given the synthetic, non-sequential nature of the records).

Example Instance

{
  "input": "expert dual chatgpt - can you design a reusable 'layered explainer' prompt that teaches any complex topic like a Prometheus-style walkthrough?",
  "output": "**Interpreting: EXPERT + DUAL + ChatGPT. [Assumption]**\n\n---\n\nI'll design a long-form, high-structure \"layered explainer\" prompt you can reuse for any complex topic...\n\n### Your Optimised Prompt\n```prompt\nYou are Aegis-Explainer, an AI whose sole purpose is to illuminate complex topics...\n```\n..."
}

---

Dataset Composition

(Derived from the current v4.1 snapshot — will shift as the dataset is finalised.)

By output mode:

Mode	Records	Share
DUAL (full package: prompt + explanation + scorecard)	1,618	~60%
PROMPT-ONLY	587	~22%
ADVICE-ONLY (guidance, no full prompt)	485	~18%

By complexity level (DUAL / PROMPT-ONLY records):

Complexity	Records	Share
STANDARD	1,235	~46%
EXPERT	541	~20%
BASIC	429	~16%

By target AI system (top targets, approximate — parsed from the Interpreting line):

Target	Records
ChatGPT	1,037
Claude	292
Midjourney	204
Gemini	157
DALL·E / DALL·E 3	118
Sora	102
Stable Diffusion / SDXL	39
Imagen 3	13
Runway	6
Excel, Cursor, LangChain, GitHub Copilot, AutoGPT, Replit Ghostwriter, and others	5 each

The distribution is intentionally weighted toward general-purpose chat models (ChatGPT, Claude, Gemini) and mainstream image/video generators, with a long tail covering code assistants and agent frameworks.

---

Dataset Creation

Curation Rationale

The dataset was built to give a small fine-tuned model reliable, consistent behaviour as a "prompt architect" — something that's hard to achieve through a system prompt alone at the 8B scale, especially for strict structural requirements (mandatory sections, scorecards, assumption ledgers) and consistent safety refusals.

Source Data

All records are synthetically generated conversation pairs, authored specifically for this project. Generation was guided by structured specifications (see the companion Astraea system-prompt documents, e.g. v7.9/v8.1) defining the required output format, mode-switching logic, and safety refusal patterns.

Annotations

There are no separate human annotations layered on top of the raw pairs in this snapshot; the output field itself encodes the target structure (Interpreting line, sections, scorecard) that downstream models are trained to reproduce.

Personal and Sensitive Information

The dataset is fully synthetic and contains no real user data, no personally identifiable information, and no proprietary third-party content. All conversations were generated for this project rather than sourced from real user interactions.

---

Considerations for Using the Data

• Not a benchmark dataset. Composition statistics above reflect the current in-progress snapshot and should not be quoted as fixed dataset properties.
• Heavily skewed toward specific target systems (ChatGPT, Claude, Midjourney, Gemini). Models trained on this data may generalise less well to underrepresented targets (e.g. niche code assistants or agent frameworks).
• Long-tail output lengths. EXPERT-mode outputs can run past 15,000 characters; ensure your training pipeline's max sequence length and packing strategy account for this, or filter/truncate outliers deliberately.
• Format-dependent, not fact-dependent. The dataset trains structural and behavioural compliance (formatting, mode selection, refusal patterns) rather than factual knowledge about third-party AI tools — some technical details referenced inside prompts (API parameters, feature availability) may be stale or approximate.

Known Limitations

• No metadata field for mode/complexity/target — currently must be parsed from the output text.
• No provided train/validation/test split.
• No formal safety-category balance sheet yet (refusal examples exist but are not currently broken out with row-level counts in this card).
• Dataset is under active revision — record count, formatting conventions, and field schema may all change in later versions (v4.2+).

---

Licensing

(To be finalised.) This is a synthetic dataset created for this project, containing no third-party copyrighted material or personal data. A specific open license (e.g. CC-BY-4.0, Apache 2.0, or ODC-BY) should be selected and confirmed before public release — update the license field in this card's YAML metadata to match.

license: apache-2.0

Citation

If you use Astraea Chat v4.1 in your research or model training, please cite:

Astraea Chat Dataset v4.1: Synthetic Prompt-Architect Conversations for Meta-Prompting Fine-Tunes, 2026.

@misc{astraeachatv41,
  title  = {Astraea Chat Dataset v4.1: Synthetic Prompt-Architect Conversations for Meta-Prompting Fine-Tunes},
  author = {braydenh563},
  year   = {2026},
  note   = {Work-in-progress snapshot, not a final release},
  howpublished = {\url{https://huggingface.co/datasets/braydenh563/Astraea_Chat_v4.1}}
}