buddhist-nlp/qwen35-mitra-lexicon-annotation
The buddhist-nlp/qwen35-mitra-lexicon-annotation is a 9 billion parameter Qwen3.5-based word aligner developed by buddhist-nlp, specifically designed for creating attested Sanskrit to English, Tibetan, and Chinese lexica. It functions by taking a Sanskrit sentence and its translation, then annotating the translation with numbered tags corresponding to each Sanskrit word. This model excels at precise word-level alignment for Buddhist textual analysis, achieving high F1 scores (e.g., 88.9 Sanskrit to Tibetan) against a commercial LLM teacher.
Loading preview...
Overview
The buddhist-nlp/qwen35-mitra-lexicon-annotation is a 9 billion parameter Qwen3.5-based model developed by buddhist-nlp, specialized as a word aligner for the MITRA-lexicon project. Its primary function is to align Sanskrit words with their corresponding translations in Tibetan, Chinese, and English. The model takes a Sanskrit sentence (segmented and lemmatized) and its translation, then inserts numbered tags into the target translation to mark the words that render each Sanskrit source word. Words without a clear equivalent are left untagged.
Key Capabilities
- Precise Word Alignment: Accurately identifies and tags translated segments for individual Sanskrit words within target sentences.
- Multilingual Support: Handles alignments from Sanskrit to Tibetan, Chinese, and English, as well as Chinese to Tibetan and English to Tibetan.
- Specialized for Buddhist Texts: Fine-tuned on a corpus including Sanskrit, Tibetan, Buddhist Chinese, and Pāli, making it highly effective for religious and academic textual analysis.
Good for
- Lexicography: Generating attested lexica, particularly for ancient languages like Sanskrit.
- Linguistic Research: Studying word-level correspondences between Sanskrit and its translations.
- Digital Humanities: Creating annotated datasets for Buddhist texts and cross-lingual analysis.
- Machine Translation Post-editing: Assisting in the verification and refinement of machine-translated ancient texts by highlighting word-level alignments.