hailaykidu/MoVoC_MT

Downstream MT evaluation of MoVoC_Tok: from-scratch MarianMT, English<->Amharic/Tigrinya, zero-shot Tigre # MoVoC_MT: Downstream Machine Translation Evaluation of MoVoC_Tok Real downstream validation of the MoVoC_Tok tokenizer (built for the MoVoC project, arXiv:2509.08812): a **from-scratch MarianMT** model, trained bidirectionally on English↔Amharic and English↔Tigrinya, then evaluated **zero-shot on English↔Tigre** (a third Ge'ez-script language never seen during training) to test whether MoVoC_Tok's shared vocabulary gives any real cross-lingual transfer. This is the kind of downstream MT validation the MoVoC paper's own Table 3 describes, but which the MoVoC project itself never had until now. Model: Hailay/movoc-mt-en-am-ti on the Hugging Face Hub. ## Architecture Matches the exact architecture of the paper's own original MarianMT run (`Paralleldata/results/checkpoint-524316/config.json`), confirmed field-for-field: 6 encoder + 6 decoder layers, 8 attention heads, `d_model=512`, feedforward dimension 2048, Swish activation, shared encoder/decoder embeddings, static (sinusoidal) position embeddings. **106,104,832 parameters.** The one deliberate difference: vocabulary. The original run used a bespoke 63,050-token vocabulary; this run uses **MoVoC_Tok's 120,000-token shared Ge'ez-script + English SentencePiece Unigram vocabulary** instead (120,004 after adding 3 direction tags and 1 dedicated ` ` token) — still comfortably above the ≥63,050 requirement, and the actual point of this project: testing MoVoC_Tok downstream, not reproducing the original run byte-for-byte. This is a genuinely new model trained from random initialization, not a fine-tune of any pretrained checkpoint — no pretrained MarianMT model anywhere has a MoVoC_Tok-compatible vocabulary, and the goal is to reproduce the paper's own from-scratch training methodology, not run a transfer-learning experiment. ## Data | Pair | Source | Train | Dev | |---|---|---|---| | English–Amharic | Raw NLLB (mined), cleaned with the same pipeline EnTiMT built for Tigrinya (NFC normalization, length/length-rat …