This is a finetune of llama-3.1-8b-instruct, specifically to process fiction (or narrative text in general) and fix grammatical and prose errors while maintaining the author's style, voice, and word choices.
Best results achievable by prompting as user and prepending "correct author-grammar.\n" to the input contents; result is markdown or plain, depending on input format.
v2 (2026-08-10)
Retrained from scratch: QLoRA (rank 8, alpha 16, all attention + MLP projections) on Llama-3.1-8B-Instruct, using 593 minimal-edit correction pairs built from long-form fanfiction (~360-word markdown chunks, matching the intended chunked-inference usage). Loss on the assistant turn only, 3 epochs; best checkpoint by held-out eval loss (epoch 2) merged into bf16 and quantized to Q8_0.
Behavior compared to v1 on held-out chunks:
| v1 | v2 |
|---|
| words changed per chunk | 50–94% | 0–1.2% |
| output/input length | +43–73% | ±1% |
| preamble ("Here's the corrected text:") | yes | none |
| em dashes introduced | yes | none (converts existing ones to parens/commas/ellipses) |
| paragraph structure | drifts | preserved exactly |
The model preserves the author's quote style (curly quotes stay curly), formatting (*italics*, **bold**, headers), and returns clean text unchanged.
Known quirks:
- At an edit site, a newly inserted quote character occasionally comes out as a straight
" where the surrounding text uses curly quotes.
- Rare single-word substitutions (~1 per few thousand words observed): e.g. replaced a word in an innuendo with a nearby noun. Worth a skim on text you care about.
Trained with the prompt format above; greedy decoding recommended.
