A $1.35 fine‑tune taught an open 4B model that CRA is ARC in French. It still writes NIS for SIN.

Description and use

A freelancer built a tool that fine‑tunes an open 4‑B model on Canadian government bilingual text, improving acronym accuracy (CRA → ARC) and flagging problematic terms with source citations.

Problem

Freelancers translating small‑business web pages into Canadian French often miss correct acronyms (e.g., CRA → ARC) and use terms that the Translation Bureau flags as calques.

Outcome

Fine‑tuning an open 4‑B model for $1.35 improved acronym accuracy from 50.8 % to 66.9 % and reduced English acronyms left in French, while providing source‑cited flags for review.

Workflow

• 1. Collect 48,116 open‑government bilingual records, filter machine‑translated French, pair sentences → 82,455 pairs.
• 2. Fine‑tune Qwen3.5‑4B with LoRA rank 32 on 8,000 pairs (2 epochs) → Run 1.
• 3. Fine‑tune again on 2,000 pairs containing known acronyms → Run 2.
• 4. Deploy Run 2 on Tinker, expose via a web app (Sans calque).
• 5. User pastes English; app returns fine‑tuned French, untouched French, and a review list with flags and source citations.
• 6. User reviews flagged sentences and sends final French to client.

Evidence

The fine‑tune is two runs on Tinker. Qwen3.5-4B, LoRA rank 32, 8,000 pairs, 2 epochs, about 13 minutes each.

Source

Run 2 beats the base model in all 1,000 bootstrap resamples of the 146 held‑out sentences that contain a government French acronym.

Source

The review list flags three things, each with its source: an English acronym left in the French, a missing French acronym, and a term the Translation Bureau’s Termium Plus says to avoid.

Source