Modelpublic

Urdatorn/sphragis-alm-olmo1b-metre-callimachus

sourceHugging Faceotherupdated 29d agoView on Hugging Face

0likes21downloads

Model Card

Authorial language model: Callimachus (Sphragis-Metre verse)

One of 17 authorial language models (ALMs) for the Sphragis-Metre Ancient Greek authorship-attribution benchmark, following Huang, Murakami and Grieve (2025), "Attributing authorship via the perplexity of authorial language models", PLoS ONE 20(7): e0327081. Each model is a full further-pretraining of `allenai/OLMo-1B-hf` (revision aee7752d9c08ee4775e9b0091426d8410e8f6a89) on one author's training rows only; a row is attributed to whichever of the 17 models finds it least surprising.

This model was trained on Callimachus: 1,100 rows, 34,230 scored tokens from the verse_1 training split. Its base model and the number of epochs were both chosen by coordinate ascent on validation attribution, so a model here may start from the vanilla OLMo-1B or from the Greek-adapted one.

Training


Objective	causal LM on `<\	endoftext\	> sentence <\	endoftext\	>`, one sentence per sequence
Selection	validation attribution macro-F1 over all 17 models
Epochs	2
Learning rate	5e-05, constant after 25 warmup steps
Effective batch	16 sentences
Precision	fp32 master weights, bf16 compute, FSDP full shard on 2x GH200
Weights	bf16

Unlike Huang and colleagues' fixed 100 epochs, training length here is chosen by held-out evidence -- and on attribution itself rather than on this author's own perplexity, because what attribution needs is not how well a model fits its author but how much better it fits than the others.

Scoring

Score a sentence exactly as it was trained, and compare the per-token negative log-likelihood against the other sixteen models. Together the seventeen models reach 56.81 test macro-F1 on verse1, 76.15 on verse5, 80.99 on verse10 and 72.88 on verse50.

Training, scoring, and attribution code: `Urdatorn/sphragis_models`.

Licence and provenance

The base model is Apache-2.0. The training text comes from the benchmark, whose sources carry mixed licences including CC BY-NC-SA material, so this derived model is released as other: see the dataset's LICENSES.md before reuse.