Results (v2, trained on 360 edit pairs)

Held-out eval (8 unseen base structures), and a blind A/B quality judge vs Sonnet's edit of the same (original, instruction):

On 2 held-out specs (brass-stab, arpeggio-conversion) v2 produced edits byte-for-byte identical to Sonnet's — it learned those idioms exactly. The remaining gap to Sonnet is pitch-level precision: the losses are "close but imprecise" (a reharmonization with the right chord concept but pitches transposed +2 semitones; a countermelody with 2 out-of-key notes; an overshot climax). This mirrors the ~4B capacity ceiling seen on the from-scratch compose task — but editing is more tractable because the original arrangement supplies in-key scaffolding.

Recommended use: hybrid — gemma for fast structural edits (instrument swaps, voice adds, arpeggiation, drum-feel changes), Sonnet for precision-critical edits (reharmonization, exact voice-leading).

Files

• adapter_model.safetensors, adapter_config.json — the LoRA
• tokenizer*.json, chat_template.jinja — tokenizer + chat template
• edit_training_pairs.jsonl — 360 (original, instruction, edited) pairs (Sonnet-generated), trained x12 key transpositions
• blind_judge_verdicts_v2.json — per-spec blind-judge outcomes
• edit_specs.py, build_edit_train.py, eval_edit.py, edit_render_check.py — reproducible pipeline

Usage

python
Copy code
from peft import PeftModel
from transformers import AutoModelForCausalLM, AutoTokenizer
base = AutoModelForCausalLM.from_pretrained("google/gemma-4-E4B-it", torch_dtype="bfloat16")
model = PeftModel.from_pretrained(base, "hidude561/justice")
tok = AutoTokenizer.from_pretrained("hidude561/justice")
# prompt: system (edit instructions) + user ("Style/Tempo/Chords ...\nArrangement:\n```arr\n<orig>```\nEdit: <instruction>")

Trained with LoRA (bf16), 3 epochs, max_len 3072, on an RTX 5090.