kobyadu1/degree-space-accompaniment

Degree-space piano accompaniment

Left-hand piano accompaniment patterns written entirely in scale degrees relative to each chord's root, paired across keys so that the same progression in two different keys carries the same answer.

What the model is meant to learn

Given a chord progression, a key, a meter, a bar count, a feel and a density, return a left-hand pattern in which every note is a scale degree relative to the current chord root, the roman numeral on each bar correctly names the chord given for that bar, and each bar's durations sum to the meter. The paired form of the same claim: the same progression presented in two different keys yields identical patterns.

A student who learns a pattern in degrees can play it from any starting note after learning it once. A student who learns specific piano keys has learned one twelfth of it.

How it was built

The patterns are synthesized, not distilled. An ablation over two frontier model families and three prompting strategies found that they return the same pattern for the same music in a second key once in 54 attempts, so copying their patterns would copy that failure. Each pattern is generated from the roman-numeral template, meter, bars, density and style, and never from the key, then rendered into several keys with the degree content untouched. Key-independence is therefore a property of the file rather than something training has to infer.

A frontier model (Claude Haiku 4.5) writes only the one-line gloss, which is the part it is genuinely good at.

Every example is checked by the same verifier used for evaluation before it is written. Nothing that would fail the spec enters the file.

Files

file	examples	progressions	bar lengths	pattern vocabularies	second constraint
train_v4.jsonl	12,000	2,365	4, 8, 16, 24, 32	115	yes
train_v3.jsonl	12,000	2,382	4, 8, 16	115	yes
train_v2.jsonl	12,000	2,384	4, 8, 16	38	no
train_v1.jsonl	2,315	14	4, 8, 16	38	no

Use train_v4.jsonl. It trained the model this dataset accompanies. The earlier versions are kept because the differences between them are the findings.

v3 adds a wider pattern library, so style selects the shape of the figure rather than reaching it through a hash, and passing tones let a bass line walk: 38 distinct degree vocabularies became 115. It also adds a second, competing constraint, requested with Hands: beginner-friendly in the prompt, capping a bar at three distinct degrees and no leap wider than a fifth. Roughly half the examples carry it so the model has to answer the request rather than learn one habit.

v4 is v3 with 24 and 32-bar requests added and nothing else changed. v3 was trained only on 4, 8 and 16 bars and handled a 32-bar request badly, stopping at 24 or running away entirely; adding the missing lengths produced a well-formed answer to every long request.

train_v1.jsonl is the counter-example. A model trained on v1 scored 90.3% on progressions it had seen and 0.0% on progressions it had not: with only 14 progressions repeated 165 times each, recalling which numerals belong to a progression was cheaper than deriving them from the chord symbols. Since a progression's numerals are the same in every key, rendering one progression into eleven keys produced eleven identical targets, which taught the paired behavior and simultaneously made the chord symbols irrelevant.

v2 keeps the pairing and removes the shortcut: one progression per request-group, four keys each. Trained on it, the same model and the same hyperparameters score 88.9% on progressions never seen, and robustness rose from 77.1% to 84.8%.

Format

Chat-style SFT records:

{"messages": [{"role": "system", "content": "..."},
              {"role": "user", "content": "I'm learning left-hand accompaniment..."},
              {"role": "assistant", "content": "{\"bars\": [...], \"gloss\": \"...\"}"}],
 "meta": {"progression": "...", "key": "...", "cof_distance": 2, "mode": "major",
          "meter": "4/4", "bars": 8, "density": "moderate", "style": "..."}}

Each bar carries a roman numeral and a list of {degree, beat, duration}. A degree is 1–7 with an optional b/# and an optional octave marker (_ down, ^ up).

The meta block is not part of the training text. It exists so results can be split by key distance around the circle of fifths, by density, and by whether the progression was held out.

Caveats

The patterns come from a bounded library of rhythms and contours, so pattern variety is narrower than a human arranger's. The measure that matters for that is diversity of answers across the eval set, reported alongside the behavioral numbers.

Glosses are model-generated and occasionally generic.