abc-folk:3

A 864,256-parameter Llama-style language model trained from scratch with slmkit for the abc_music project. It is a fine-tuned (SFT) model: it answers requests in the format it was fine-tuned on.

How to prompt it

Generation starts from the prompt and stops at <eos> (token ID 1). An example prompt, exactly as the evaluation used it:

% a reel in D major, 2/2 time

Serve it and ask over HTTP. A copy downloaded from the Hugging Face Hub goes in $SLM_HOME/models/abc-folk/3/ first; that is where slm serve looks for abc-folk:3.

uv run slm serve --model abc-folk:3
curl -s localhost:8000/generate -H 'content-type: application/json' \
  -d '{"prompt": "% a reel in D major, 2/2 time\n", "seed": 0}'

Or load it with Hugging Face transformers; the directory is a standard LlamaForCausalLM checkpoint:

from transformers import AutoModelForCausalLM, AutoTokenizer
path = "$SLM_HOME/models/abc-folk/3"  # or its Hugging Face Hub repo ID, if it was published there
tok = AutoTokenizer.from_pretrained(path)
model = AutoModelForCausalLM.from_pretrained(path)

Model

architecture	LlamaForCausalLM (RMSNorm, RoPE, SwiGLU, tied embeddings)
parameters	864,256 (853,120 non-embedding)
vocabulary	87 tokens
context	512 tokens
weights	model.safetensors, fp32, 3.5 MB

Training

run	run-9b890db6dced (abc-baseline-sft), checkpoint best at step 1,000
tokens seen (this stage)	16,384,000
validation loss	1.1092
GPU-hours (this stage)	0.006
tokenizer	tk-1030590a89ca
code	slmkit c34aed6

Full lineage back to the raw data: uv run slm lineage abc-folk:3.

Evaluation

From eval report ev-0ddc32e2a9: 3 sampling seeds × 200 samples, prompted with the project's sft prompts. Mean ± standard deviation across seeds.

metric	score
ended	0.982 ± 0.006
length	259.345 ± 4.707
plays	0.730 ± 0.048
bar_accuracy	0.851 ± 0.023
ends_on_tonic	0.340 ± 0.026
novelty	0.998 ± 0.001

Metric definitions: the project's README and docs/concepts/evaluation.md.

Verification

Checked before it was published, over 94 tokens: slmkit reloads the files with max |Δlogit| = 0.0e+00 (identical). transformers 5.17.0 loads them with max |Δlogit| = 0.0e+00 and identical token IDs.

Data and licence

Trained on 4,248 tunes in ABC notation from three public-domain tune books: Ryan's Mammoth Collection (1883), O'Neill's Music of Ireland (1903) and Aird's Selection of Airs (1778–1803), as transcribed by volunteers and shipped in the music21 package (BSD-3-Clause).

One open question, stated plainly. The printed books are public domain, but the transcription files carry no licence of their own. Faithful transcriptions of public-domain scores are generally not treated as new creative work, but that is an assumption, not a grant. If a transcriber or rights-holder objects, the collection is removed, the model retrained, and this version withdrawn. The Session's data was deliberately not used: its licence forbids training language models.

The weights are released under Apache-2.0, the same licence as slmkit.

More detail: projects/abc_music/README.md in the slmkit repository.

Limitations

• Context of 512 tokens. slmkit's sampler keeps only the last 512 tokens as context; positions beyond that were never trained.
• Fixed vocabulary of 87 tokens. Text the tokenizer has no token for becomes <unk>, which the model learned nothing about. slm serve reports how many prompt tokens that affected.
• Small and narrow. It knows only its training domain. The scores above are averages; individual samples can be wrong, so check them with the project's graders.