Samdei Type Mini Copyright 2026 The Samdei Type Authors Licensed under the Apache License, Version 2.0 (see LICENSE). This covers every file in this repository: the ONNX graphs, onnx/vocab.json, rerank.json and the word language model. Word language model (lm/wordfull_nowiki_p25.json.gz) Word unigram and bigram counts of Khmer text, built without Wikipedia, from: - HPLT v2 (HPLT/HPLT2.0_cleaned, khm_Khmr): CC0 packaging; the Common Crawl Terms of Use apply to the text. - FineWeb-2 (HuggingFaceFW/fineweb-2, khm_Khmr): ODC-By 1.0; the Common Crawl Terms of Use apply to the text. - Asian Language Treebank (ALT), Khmer side: CC BY 4.0. - Samdei Type chat sentences (samdei/samdei-type-resources): CC BY 4.0. Training data The model was trained on samdei/samdei-type-pretraining-data and samdei/samdei-type-finetuning-data (CC BY-SA 4.0); their cards list every source. Part of the fine-tuning pairs were spelled with a table derived from the Chhunneng romanized Khmer dataset, MIT License, Copyright (c) 2025 Chrea Chanchhunneng.