feat: writing style finetuning backend (Phase 2) #52

New issue

Open

opened 2026-05-01 12:13:45 -07:00 by pyr0ball · 0 comments

pyr0ball commented

2026-05-01 12:13:45 -07:00

Owner

Context

Avocet already benchmarks writing style models. This adds the finetuning backend to close the loop: corpus → train → personal style LoRA adapter.

Work

scripts/finetune_style.py — LoRA finetuning on LLM using personal writing corpus from gather_corpus.py
Register as style job type in train queue
Corpus format: .txt files in data/style_corpus/ (already gathered by gather_corpus.py)
Hyperparams: base model, LoRA rank, epochs, sample length

Depends on

#43 (train job queue)
#46 (LLM SFT backend — reuse PEFT infrastructure)

## Context Avocet already benchmarks writing style models. This adds the finetuning backend to close the loop: corpus → train → personal style LoRA adapter. ## Work - `scripts/finetune_style.py` — LoRA finetuning on LLM using personal writing corpus from `gather_corpus.py` - Register as `style` job type in train queue - Corpus format: `.txt` files in `data/style_corpus/` (already gathered by `gather_corpus.py`) - Hyperparams: base model, LoRA rank, epochs, sample length ## Depends on - #43 (train job queue) - #46 (LLM SFT backend — reuse PEFT infrastructure)