DHARMATUNE · OPERATIONS
loading data.json…
data.json failed to load — dashboard is static HTML + /finetuning/data.json
Results — catholicism-gemma-v1
no judged runs yet
Pipeline
Unit of work: a source text → ~1,200-word chunks → Sparky (local 35B agent) generates or a parser extracts Q&A → schema + contamination validation → per-tradition train.jsonl → LoRA (worker node) → dev battery + regression gate → official DharmaBench. The official 10-question battery is quarantined test data; iteration runs on a separate dev battery. DharmaBench →
Trainability — records vs 10k floor
Floor = minimum domain Q&A records before a LoRA worldview tune is worth running (before ~30% general replay mix). Only traditions at floor get a training run. Verbatim extraction (catechisms, Summa articles/objection-replies) counts 1:1; generated records are model-written but grounded to a source chunk.
Datasets by tradition
Generation waves
| Tradition | Source | Chunks | Generated | State |
|---|
Source texts
| Tradition | Title | Edition | Size | Chunks |
|---|
Training
| Run | Result |
|---|
Evaluation instruments
Activity
| When | What | Who |
|---|