Ancient Greek to Modern Greek Machine Translation: A Novel Benchmark and Fine-Tuning Experiments on LLMs and NMT Models

In three linesNew AG-MG parallel corpus with 132,481 sentence pairs for Ancient-to-Modern Greek translation. Creation pipeline combines web-scraping, VecAlign alignment with fine-tuned LaBSE embeddings, and Gemini 2.5 Flash LLM-based correction. Benchmark of NMT models (NLLB, M2M100) and Greek LLM (Llama-Krikri-8B): full fine-tuning achieves 13.16 BLEU, gains up to +10.3 points.

Summary generated by Claude — human-verified