ayaka

Active research

Hime

A translation pipeline for long-form literary Japanese into English. Several small fine-tuned models per paragraph instead of one large model in a data centre — running on a single consumer GPU, fully offline, EPUB in and EPUB out.

The pipeline

Each paragraph passes through four stages. The interesting part is not the chain — it is that retrieval feeds every stage, not just the first one. Glossaries and earlier passages are pulled in per paragraph, which is what keeps names, honorifics and register stable across hundreds of pages.

Four-stage translation pipeline. Each paragraph is drafted, merged, polished and reviewed, with glossary and prior-passage retrieval feeding every stage rather than only the first. EPUB in EPUB out Draft first pass per paragraph Merge reconcile candidates Polish register and rhythm Review judge against the source retrieval per paragraph glossary entries and earlier passages, fed into every stage Four-stage translation pipeline. Each paragraph is drafted, merged, polished and reviewed, with glossary and prior-passage retrieval feeding every stage rather than only the first. EPUB in Draft first pass per paragraph Merge reconcile candidates Polish register and rhythm Review judge against the source EPUB out retrieval per paragraph

Where it currently stands

The fine-tuned models do not yet reliably beat their own base model. The most recent supervised fine-tune measured as a regression against the previous one. The diagnosis is that the ceiling sits in the base model and in the training data rather than in the adapter, which is why the next step is continued pretraining rather than more fine-tuning.

Stating that plainly costs nothing. A pipeline that only ever reported improvements would be less believable, not more.

Setup

  • single consumer GPU
  • fine-tuned adapters
  • retrieval-augmented context
  • fully offline

The part that works

A result was withdrawn after the harness caught its own leak.

The evaluation infrastructure is the mature part of this project: a de-leaked judge harness with 5 integrity gates, and a probe covering 41 linguistic axes, scored by a panel of reader personas.

An earlier positive result was traced to a leak in the answer key and retracted. It is recorded here rather than quietly dropped, because a measurement setup that cannot catch its own mistakes is not a measurement setup.

The repository is private, so there is no link to it here. A link that returns 404 is worse than none.