cuneiform-stoicheia

Sign-level model for cuneiform text. It restores the signs lost in a break and, when given a library of tablets, retrieves parallel passages and reports the tablet and line each restoration draws on. Full documentation to follow.

import sys
from huggingface_hub import snapshot_download, hf_hub_download
path = snapshot_download("Ericu950/cuneiform-stoicheia", allow_patterns=["*.json", "model.safetensors", "cuneiform_stoicheia/*", "examples/*"])
sys.path.insert(0, path)
from cuneiform_stoicheia import load_model, Library, restore

model = load_model(path)
lib = Library.from_parquet(hf_hub_download("Ericu950/cuneiform-stoicheia-library", "library.parquet", repo_type="dataset"), vocab=model.vocab)
res = restore(model, "1(disz) udu [x x] lugal-e\nki ur-{d}en-lil2-la2-ta\n[...] i3-dab5", lang="sux", library=lib)
print(res.translit)

Files: model.safetensors (one trained model), model_a/b/c.safetensors (three seeds, for load_model(path, ensemble=True)), config.json, vocab.json, the Python package cuneiform_stoicheia/, examples/. The first release is kept under the tag v1.

Licence: weights, vocabulary and examples CC BY-NC-SA 4.0 (the training data include editions of the electronic Babylonian Library, which are non-commercial). Bundled Python package: MIT.

Training data: Cuneiform Digital Library Initiative (CDLI); Open Richly Annotated Cuneiform Corpus (Oracc) and its projects; electronic Babylonian Library (eBL); Thesaurus Linguarum Hethaeorum digitalis (TLHdig).

Author: Eric Cullhed.

Downloads last month
21
Safetensors
Model size
0.2B params
Tensor type
F32
ยท
Inference Providers NEW
This model isn't deployed by any Inference Provider. ๐Ÿ™‹ Ask for provider support

Dataset used to train Ericu950/cuneiform-stoicheia