EvaCun: ORACC Akkadian Parallel Corpus

EvaCun: ORACC Akkadian Parallel Corpus for Machine Translation (train/validation), v0.1

DESCRIPTION

Overview:This dataset provides aligned Akkadian (transliteration), Akkadian (Unicode cuneiform), and English segments prepared for machine translation experiments in the EvaCun workflow. Data are sourced from ORACC project exports and organized for reproducible training, tokenization, and evaluation with the EvaCun/Akkademia pipeline.

Contents:Six plain-text files (UTF-8), one segment per line, aligned by row index across languages within each split.

akkadian_train.txt

transcription_train.txt

english_train.txt

akkadian_validation.txt

transcription_validation.txt

english_validation.txt

Note on terms:“Transcription” denotes standardized Romanized Akkadian used by downstream tools. “Akkadian (Unicode cuneiform)” follows the EvaCun conversion stage described in the workflow.

Source and preprocessing:The corpus originates from ORACC project data (“corpusjson” exports). Preprocessing in the EvaCun pipeline includes:

cleaning of spurious markers (e.g., replacing stray NaN / Uncertain);

normalization of gap tokens (X / ...);

per-text segmentation; and

alignment of transliteration, cuneiform, and English lines under a shared id_text.The split strategy follows an 80/10/10 design (train/validation/test). The test portion is withheld for separate evaluation. This deposit includes train and validation only.

Intended use:Neural MT experiments and reproducible benchmarking with the EvaCun/Akkademia stack. The accompanying notebooks demonstrate tokenization with SentencePiece and model training/translation with fairseq (train.py, generate.py), in line with the documented workflow.

Format:

Encoding: UTF-8 (Unix line endings)

Alignment: strict line-wise alignment within each split (row n in each file corresponds to the same segment)

Headers: none; plain text only

Limitations and notes:Coverage reflects the subset of ORACC Akkadian projects processed to date; genre distribution may be uneven. Alignments are sentence/line-level per the workflow heuristics and may require task-specific re-segmentation. Users should review domain/genre needs before training.

Versioning:v0.1 corresponds to the dataset prepared for the EvaCun Colab Notebook release described in the workflow notes. Future versions may expand coverage or revise segmentation.

Related resources:EvaCun Colab notebooks and setup guides (sister repositories) that consume this dataset and reproduce the pipeline stages (tokenization, training, translation, detokenization):https://github.com/ancient-world-citation-analysis/EvaCun-Colab-Notebook/tree/main

Keywords:Akkadian; cuneiform; ORACC; machine translation; parallel corpus; SentencePiece; fairseq; EvaCun

License:CC BY 1.0 (N.B redistribution is permitted for the specific ORACC derivatives we prepared).

SHORT DATASET CARD

EvaCun: ORACC Akkadian Parallel Corpus (v0.1). A line-aligned, three-column parallel corpus (Akkadian transliteration / Akkadian Unicode cuneiform / English) prepared from ORACC sources for MT. Preprocessing and the training workflow (tokenization with SentencePiece; training/translation with fairseq) follow the EvaCun/Akkademia notes. Provided files: six UTF-8 .txt for train/validation splits; one segment per line; aligned by index. Intended for MT training and reproducible benchmarking with the EvaCun Colab notebooks. Limitations: coverage and genre balance reflect available ORACC data; alignment and normalization decisions may affect downstream scores. Cite this dataset’s Version DOI; see the GitHub repository above for notebooks and pipeline details.

HOW TO CITE (DATASET, VERSION-SPECIFIC — RECOMMENDED)

APA (human-readable):Anderson, A. (2025). EvaCun: ORACC Akkadian Parallel Corpus for Machine Translation (train/validation), v0.1 [Dataset]. Zenodo. https://doi.org/10.5281/zenodo.17220688

BibTeX:@dataset{evacun_oracc_parallel_v01_2025,title = {EvaCun: ORACC Akkadian Parallel Corpus for Machine Translation (train/validation), v0.1},author = {Anderson, Adam},year = {2025},publisher = {Zenodo},version = {v0.1},doi = {10.5281/zenodo.17220688},url = {https://doi.org/10.5281/zenodo.17220688},note = {Contributors: Christian Karren (Project manager), Olivia McCauley (Project manager), Melanie Her (Data collector), Mackenzie Moffit (Project member), Lirui Harrison Huang (Data curator), Jenna Kudaimi (Data collector), Neha Ramakrishnan (Data collector), Aurelia Widjaja (Project member), Mohammad Shahnawaz (Project member), Emily Xu (Data collector), Indu Abhilash (Data collector).}}

CREATORS

Anderson, Adam (University of California, Berkeley)

CONTRIBUTORS

Karren, Christian (University of California, Berkeley) — Project manager

McCauley, Olivia (University of California, Berkeley) — Project manager

Her, Melanie (University of California, Berkeley) — Data collector

Moffit, Mackenzie (University of California, Berkeley) — Project member

Huang, Lirui Harrison (University of California, Berkeley) — Data curator

Kudaimi, Jenna (University of California, Berkeley) — Data collector

Ramakrishnan, Neha (University of California, Berkeley) — Data collector

Widjaja, Aurelia (University of California, Berkeley) — Project member

Shahnawaz, Mohammad (University of California, Berkeley) — Project member

Xu, Emily (University of California, Berkeley) — Data collector

Abhilash, Indu (University of California, Berkeley) — Data collector