A LoRA adapter for unsloth/Qwen2.5-Coder-7B-bnb-4bit that transcribes between Byzantine (Chrysanthine) neume notation and Western staff pitches for a diatonic interval grammar — including ascending leaps up to an octave and rhythmic note durations, in both directions.

It is a small fine-tuned model that reliably performs a specialized notation task that frontier-model prompting did not: on a held-out, zero-leakage test set the neume→west direction reaches 96% exact-match / 98% melodic equivalence.

This adapter was trained on synthetic, correct-by-construction data, not on scanned manuscripts. Read the Scope section before using it — it does not transcribe real melismatic chant, and the reverse direction is limited by an inherent ambiguity in the notation.

TL;DR

	neume → west	west → neume
exact_match	96.0%	13.7%
pitch_accuracy	99.1%	76.5%
interval_accuracy	99.2%	n/a (output is neumes)
melodic_equivalence	97.8% (1.955 / 2.0)	49.0% (0.981 / 2.0)
strict_pass_rate	96.2%	16.9%

Held-out: 4,794 sequences, zero leakage, deterministic scoring (no LLM judge). All metrics shown as percentages; melodic-equivalence is a 0–2 rubric reported as % of its 2.0 maximum.

What it does well

neume → west transcription (the headline capability). Given a neume sequence and its ison (starting-pitch) anchor, the model produces the exact Western pitch sequence 96% of the time, correctly handling:

the diatonic interval vocabulary — unison, steps, and ascending leaps up to an octave (oligon_kentema +3, oligon_hypsili +4, and the ypsili combinations up to +7);
rhythmic durations — apli / dipli / tetrapli produce held notes, rendered inline as <pitch>:<beats> (e.g. C4:5);
the correct mode header and ison echo;
sequence length — it stops at the right place rather than running on (a ChatML <|im_end|> stop-token was used at inference).

What it does NOT do well (and why)

west → neume exact-match is low (14%) — a notation ceiling, not a model failure. Several neumes encode the same pitch motion (e.g. oligon and petaste are both "+1 step"), so a rising step is genuinely two-valued and cannot be uniquely recovered from pitches alone. Judge this direction by positional pitch accuracy (76%), not exact match.
Real scanned manuscript chant is out of scope. The available real neume↔pitch corpus was found to be non-recoverable to exact pitch (the paired neume and pitch streams did not correspond; ~10% positional ceiling for any model), which is why this adapter was trained on synthetic data. It has not been shown to transcribe real melismatic chant.
Microtones, chromatic/enharmonic modes, fthora, and melisma are excluded by design. Only the four diatonic modes (1, pl.1, 4, pl.4) on a natural-note ladder are modeled; the model asserts no accidental or microtonal content.

Why the results are trustworthy

Zero leakage: the 4,794-row test set shares no row-id and no exact input prompt with training (both verified = 0).
Deterministic scoring, no LLM judge: the synthetic gold is exact by construction, so correctness is computed directly (exact match, per-position pitch accuracy, interval accuracy, edit distance, and a 0–2 melodic composite). Gold-vs-gold scores a perfect 100%, confirming the scorer measures real correctness.
Correct-by-construction data: every training/eval pair is generated from a vetted interval grammar and re-derived by an independent verifier (0 content + 0 reversibility errors).

Headline finding

The same class of model scored ~10% positional accuracy on the real corpus and 96% exact on correct-by-construction data. That contrast isolates the earlier failure as a data-labeling problem, not a model-capability problem: when the neume↔pitch pairs actually correspond, a small fine-tuned model learns the grammar almost perfectly.

Usage

python

from transformers import AutoModelForCausalLM, AutoTokenizer
from peft import PeftModel

base = "unsloth/Qwen2.5-Coder-7B-bnb-4bit"
tok = AutoTokenizer.from_pretrained(base)
model = AutoModelForCausalLM.from_pretrained(base, device_map="auto")
model = PeftModel.from_pretrained(model, "YOUR_USERNAME/byzantine-synthetic-grammar-lora")

prompt = ("Transcribe this Byzantine neume sequence (6 neumes) to Western staff pitches:\n"
          "Mode 1\nIson: D4\noligon_kentema petaste elaphron petaste ison oligon_hypsili")
msgs = [{"role": "system", "content": "You are a Byzantine chant notation assistant."},
        {"role": "user", "content": prompt}]
ids = tok.apply_chat_template(msgs, add_generation_prompt=True, return_tensors="pt").to(model.device)
out = model.generate(ids, max_new_tokens=96,
                     eos_token_id=tok.convert_tokens_to_ids("<|im_end|>"))
print(tok.decode(out[0, ids.shape[1]:], skip_special_tokens=True))
# -> Mode 1 / Ison: D4 / G4 A4 F4 G4 G4 D5   (example shape)

Input format. Line 1: an instruction naming the token count. Line 2: Mode X. Line 3: Ison: <pitch> (the starting-pitch anchor — pitches are relative, so this is required). Line 4: the space-separated neume (or, for west→neume, pitch) tokens.

Vocabulary (diatonic, degree-shift)

ison 0 · oligon/petaste +1 · apostrophos −1 · elaphron −2 · elaphron_apostrophos −3 · chamile −4 · oligon_kentema +3 · oligon_hypsili +4 · ypsili_left_oligon +5 · ypsili_kentima_oligon +6 · ypsili_over_kentima_oligon +7 (octave). Duration signs: apli (2 beats) · dipli (3) · tetrapli (5). Breath/barline signs are no-ops on pitch (neume→west only).

Training

Base: unsloth/Qwen2.5-Coder-7B-bnb-4bit (4-bit), LoRA via PEFT, response-only loss.
Data: 23,942 rows of verifier-checked synthetic grammar (both directions, 4 diatonic modes, leaps + durations); 4,794-row disjoint held-out slice (0 id/content leakage).
Objective: exact bidirectional transcription of the interval grammar; the task is small, finite-vocabulary interval arithmetic anchored to a stated ison.

Intended use & limitations

Educational / research demonstration that supervised fine-tuning teaches a small model a notation grammar that prompting a frontier model did not reliably perform. Not a production transcription tool for real manuscripts, and not a source of musicological ground truth for microtonal or chromatic repertoire.

License

Apache-2.0 (adapter). The base model retains its own license.

FableMogger9000/byzantine-synthetic-grammar-lora

Byzantine Synthetic-Grammar LoRA — Neume ↔ Western Pitch Transcription