# Genesis_FineTune_Core_25k (Master Scholar)

**Developer / Brand:** **Within Us AI**

A Genesis-style, end-to-end **advanced fine-tuning starter pack**: five datasets (5k each) packaged together to train an assistant’s *behavioral core*.

## What this pack trains
1) **Instruction-following (multi-turn)** — constraints, formatting, spec compliance  
2) **Tool / function calling** — schema-accurate tool calls + observations  
3) **Preference learning (DPO-style)** — chosen vs rejected responses  
4) **Long-context retrieval** — cite-the-source answering from multiple docs  
5) **Code patch + debugging** — unified diffs and test intent (including defensive fixes)

## Files (25,000 total)
- instruct_multiturn_5k.jsonl
- tool_use_functioncalling_5k.jsonl
- preference_dpo_5k.jsonl
- longcontext_retrieval_5k.jsonl
- codepatch_debug_5k.jsonl
- README.md
- dataset_card.md

## Unified schema
All records share the same top-level schema: `id`, `type`, `prompt`, `response`, `meta`, and optional `artifacts`.

## Composition
By type:
{
  "instruction_multiturn": 5000,
  "tool_use": 5000,
  "preference_pair": 5000,
  "long_context_retrieval": 5000,
  "code_patch": 5000
}

By safety label:
{
  "allowed": 24011,
  "refuse": 989
}

## Attribution
**Within Us AI — Genesis_FineTune_Core_25k**

## License
Apache-2.0