# Timmy Voice: Batch 09 — 1K Prompt→Response Pairs

Training Factory — Timmy Voice Worker 9/10 (#589)

## Files

| File | Description |
|------|-------------|
| `timmy-voice-batch09.jsonl` | 1,000 ShareGPT-format prompt→response pairs |
| `timmy-voice-batch09.sources.json` | 50 source sessions with approved-model provenance |
| `generate_timmy_voice_batch09.py` | Deterministic generator for the batch |

## Generation Contract

- 50 source sessions
- 20 prompt variations per session
- approved-model provenance filter
- Knowledge Mine-style ranking using local session metadata + pair quality
- ShareGPT format (`system` / `human` / `gpt`)

## Stats

- Total pairs: 1000
- Source sessions: 50
- Average quality score: 0.90
- Minimum quality score: 0.84
- Maximum quality score: 0.92

## Category Breakdown
- technical: 1 source sessions
- operations: 36 source sessions
- sovereignty: 10 source sessions
- pastoral: 0 source sessions
- crisis: 3 source sessions
- general: 0 source sessions

## Source Models
- xiaomi/mimo-v2-pro: 47 sessions
- qwen/qwen3.6-plus:free: 2 sessions
- qwen3:30b: 1 sessions


## Notes

This batch uses approved local session sources only. Banned providers (Claude/GPT/Gemini/OpenAI/Anthropic) are excluded at selection time. The generator keeps the source manifest on disk so the batch can be inspected and regenerated without guessing where the voice came from.