eo-mt-v7-mini
30M-parameter Marian MT model for English ↔ Esperanto, trained on
jensjepsen/esperanto-mt-parallel (5.1M sentence pairs).
Offline FLORES devtest, en→eo, sacrebleu lowercase=True (full 1012 pairs):
| Metric | Score |
|---|---|
| BLEU | 30.55 |
| chrF | 61.08 |
| chrF++ | 58.20 |
Roughly matches eo-mt-v6 (BLEU 31.00, 60M params) at half the parameter count. Above NLLB-200-distilled-600M (BLEU 29.28).
Tokenizer is SentencePiece, case-folding: outputs are always lowercase. Use
--lowercase for apples-to-apples sacrebleu.
Training: fp16 on a 1080 Ti, 3 epochs, 24h wall time, batch 64 × grad-accum 2. Architecture: d_model 384, 4+4 enc/dec layers, 6 heads, ffn 1536.
- Downloads last month
- 38