Egonyu/ateso-english-nllb

Ateso ↔ English (NLLB-200 fine-tune) — v1 · SUPERSEDED

## ⚠️ Use `Egonyu/ateso-english-nllb-v3` instead English → Ateso on this model does not produce Ateso. It produces Luganda, an unrelated Bantu language — fluently and confidently, which makes the failure easy to miss if you do not read Ateso. `` "The children are going to school." -> "Abaana bagenda ku ssomero." # Luganda "I am hungry." -> "Nnina enjala." # Luganda ``

Ateso → English works on this model and is genuinely Ateso-aware. Only the reverse direction is affected.

Why it happened

The teo_Latn token does not exist in base NLLB-200, so it was added and its embedding initialised by copying the lug_Latn (Luganda) row. The model was then wrapped in LoRA targeting only q_proj, v_proj, k_proj, out_proj, fc1, fc2, with no modules_to_save.

That froze the embedding table. The Luganda seed was made untrainable and never moved across the whole training run:

torch.equal(emb[teo_id], emb[lug_id])    # True
(emb[teo_id] - emb[lug_id]).abs().max()  # 0.000e+00
torch.equal(lm_head[teo_id], lm_head[lug_id])  # True (weights are tied)

Since NLLB ties lm_head to the input embedding, forced_bos_token_id=teo_Latn was mathematically identical to forced_bos_token_id=lug_Latn. Ateso → English worked because the encoder learned Ateso; the decoder side was never trained.

The general lesson: if you add a new language token to a multilingual model and then fine-tune with LoRA, the new embedding is frozen unless you explicitly include it — via modules_to_save, or by unfreezing the embedding table. Training will complete, the loss will fall, and your new language will silently remain a copy of whatever you seeded it from.

Test for it directly. BLEU will not catch this — fluent output in the wrong language can score plausibly against nothing:

a = generate(text, forced_bos_token_id=new_lang_id)
b = generate(text, forced_bos_token_id=seed_lang_id)
assert a != b, "the new language token never separated from its seed"

What v3 changed

Kept for the record

This model remains published because the failure is instructive and because Ateso → English still works. For anything practical, use v3.

Licence

cc-by-nc-4.0, inherited from facebook/nllb-200-distilled-600M. Non-commercial use only.

Citation

@misc{egonyu2026ateso,
  title  = {Ateso--English Neural Machine Translation},
  author = {Egonyu, Daniel},
  year   = {2026},
  howpublished = {\url{https://huggingface.co/Egonyu/ateso-english-nllb-v3}},
  note   = {Ateker Language Initiative}
}

Built by Daniel Egonyu · Ateker Language Initiative