KrenTransV0-1

MWire Labs | WMT 2026 Submission | Team: KrenTransV0-1

Overview

NLLB-200-1.3B fine-tuned on 11 Northeast Indian languages paired with English. Extended vocabulary: 256,212 tokens (8 custom lang tags added).

This Checkpoint

• CP2 — retrain after critical tokenization bug fix (run 1 used batch["src_lang"][0] instead of per-sample src_lang)
• Trained: 60,000 steps | Effective batch: 96 | bf16
• mni_Mtei excluded (Meitei Mayek script unsupported by NLLB subword vocab)
• Best val loss: ~2.245 (plateaued at step ~26k)

Languages

Code	Language	Script	Category
asm_Beng	Assamese	Bengali	Cat 1
lus_Latn	Mizo	Latin	Cat 1
kha_Latn	Khasi	Latin	Cat 1
mni_Beng	Meitei	Bengali	Cat 1
njz_Latn	Nyishi	Latin	Cat 1
brx_Deva	Bodo	Devanagari	Cat 2
nag_Latn	Nagamese	Latin	Cat 2
trp_Latn	Kokborok	Latin	Cat 2
tgj_Latn	Tangsa	Latin	Cat 2
mjw_Latn	Karbi	Latin	Cat 2

Known Limitations

• Category 2 languages (low data) show repetition loops with greedy decode — use num_beams=4
• Karbi (1,745 train pairs) and Tangsa (4,875) severely data-starved

Training Data

• Train: 1,366,790 pairs | Dev: 74,964 pairs
• Sources: WMT 2026 official data + HuggingFace datasets