HexaBee-4B: Six-Language Translation with Multilingual Answering and Bidirectional Supervision
HexaBee-4B reports stronger six-language translation than larger Gemma baselines across every tested direction.
The paper adapts Gemma 3 4B IT through continued Wikipedia pretraining, multilingual answering, answer-text translation, and bidirectional translation specialization. Its training mix includes 19.46 million answering records and 10.70 million answer-text translations, followed by 658,274 synthetic forward and reverse translation examples. On 36,572 synthetic tests across Arabic, Chinese, English, Hebrew, Persian, and Spanish, it reaches 73.75 direction-macro chrF++, above Gemma 3 4B IT at 57.11 and Gemma 3 12B IT at 66.83. The authors say BLEU, TER, COMET, and a matched model-judged subset support the same ranking. Source: HF Daily Papers' note.
The paper adapts Gemma 3 4B IT through continued Wikipedia pretraining, multilingual answering, answer-text translation, and bidirectional translation specialization. Its training mix includes 19.46 million answering records and 10.70 million answer-text translations, followed by 658,274 synthetic forward and reverse translation examples. On 36,572 synthetic tests across Arabic, Chinese, English, Hebrew, Persian, and Spanish, it reaches 73.75 direction-macro chrF++, above Gemma 3 4B IT at 57.11 and Gemma 3 12B IT at 66.83. The authors say BLEU, TER, COMET, and a matched model-judged subset support the same ranking. Source: HF Daily Papers' note.
score 4