>
AyakaCorp

Active research

Hime

A local-first machine-translation pipeline for long-form Japanese-to-English literary text. Instead of one large cloud model, several small fine-tuned local models work on each paragraph in turn.

Drafting, merging, polishing, reviewing. Per paragraph, on one machine.

status
active research
hardware
single consumer GPU
models
fine-tuned adapters
context
retrieval-augmented
i/o
EPUB in · EPUB out
network
fully offline
01 · Pipeline per paragraph · checkpointed a run can be interrupted and resumed mid-book

The pipeline

  1. Preprocess Morphological tokenisation, glossary lookup, retrieval of earlier passages from the same work: each paragraph arrives with its own context.
  2. Draft Several models translate in parallel, including one deterministic dictionary-based path that uses no model at all.
  3. Merge A consensus model combines the strongest parts of the drafts.
  4. Polish A separate model works on tone, rhythm and register.
  5. Review A panel of reader personas, each tuned to a different failure mode. An aggregator returns one of three verdicts: accept, re-polish, or discard and retry from drafting. Re-polish runs at most twice, a full retry at most once.
Hime pipeline: EPUB in, then Preprocess, Draft, Merge, Polish, Review. The review verdict either accepts to EPUB out, sends the paragraph back to Polish, or discards and retries from Draft. re-polish · ≤2× discard · retry from draft · ≤1× EPUB in Preprocess Draft Merge Polish Review accept EPUB out morphology, glossary, earlier passages parallel drafts, one dictionary-only path consensus combines strongest parts tone, rhythm, register reader personas, aggregated verdict
Verdict arrows return to Polish or Draft; accepted paragraphs leave as EPUB.
Hime pipeline: EPUB in, then Preprocess, Draft, Merge, Polish, Review. The review verdict either accepts to EPUB out, sends the paragraph back to Polish, or discards and retries from Draft. EPUB in Preprocess morphology · glossary · context Draft parallel · one dictionary path Merge consensus of drafts Polish tone · rhythm · register Review personas · aggregated verdict accept · EPUB out re-polish ≤2× discard · retry ≤1×
Verdict arrows return to Polish or Draft; accepted paragraphs leave as EPUB.

single consumer GPU · fine-tuned adapters · retrieval-augmented context · EPUB in / EPUB out · fully offline

02 · Current state five integrity gates · 41 axes next: continued pretraining

Current state

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 ceiling appears to sit in the base model and the training data rather than in the adapter, which is why the next step is continued pretraining instead of more fine-tuning.

Hime runs on a single consumer GPU, fully offline. Consistency across hundreds of pages comes from the per-paragraph retrieval of glossaries and earlier passages, so names and register stay stable through a long work.

What does work is the evaluation setup: a de-leaked judge harness with five integrity gates and a probe covering 41 linguistic axes.

Retraction

An earlier positive result was traced to an answer-key leak and retracted. A number that cannot survive checking is worth less than no number.