Kokborok is a critically under-resourced Tibeto-Burman language spoken by over one million people primarily in Tripura, India, and is absent from modern massively multilingual machine translation (MT) systems. We describe the SCE-KIIT submission to the WMT 2026 Shared Task on Low-Resource Indic Language Translation (Category 2: English-Kokborok). We ask whether a large language model can be adapted to translate a language using no parallel supervision at all. Our system, KokLLaMA-3.2-3B-Instruct, fine-tunes Llama-3.2-3B-Instruct via QLoRA purely on Kokborok conversational instruction data, and recovers translation behaviour at inference time through structured prompting and a rule-based post-processing pipeline. On the official WMT 2026 test set it obtains 5.11 BLEU / 27.09 chrF++ (EN-TRP) and 3.58 BLEU / 26.14 chrF++ (TRP-EN), placing third of four and third of five submitted primary systems, and coming within 1.9 BLEU of the best WMT 2025 system for this pair despite using no parallel data. The approach is substantially more effective at generating Kokborok than at generating English, and we analyse why. We further show that our in-domain validation split under-estimated official performance by more than an order of magnitude, a cautionary result for participants who tune on held-out slices of provided training corpora, and we characterise a pathological Translation Edit Rate (TER) failure mode in which an untrained baseline scores well by silently under-generating.

Nagamese is a critically under-resourced Assamese-based creole language spoken in Nagaland, India, serving as the primary lingua franca for over two million speakers across 16 Naga tribal communities. Written in the Latin script, it is entirely absent from all modern massively multilingual machine translation (MT) systems. We describe the SCE-KIIT submission to the WMT 2026 Shared Task on Low-Resource Indic Language Translation (Category 2: English-Nagamese), for which we submitted a primary EN-NAG run. Our system builds on Facebook’s NLLB-200 distilled with 600M parameters using a two-phase surrogate-token strategy: we fine-tune with Assamese (asm_Beng) as the target-side language prefix, then force asm_Latn at inference time. This token is absent from NLLB’s vocabulary and resolves to the unknown token (UNK); we hypothesise that seeding generation with an out-of-vocabulary token leaves the decoder’s script preference unconstrained, allowing the Romanised output learned during fine-tuning to surface. On our 228-sentence held-out test split, the system achieves SacreBLEU 40.02, chrF 53.61, and COMET 0.7212 (wmt22-comet-da), against a zero-shot baseline of 0.40 BLEU and 0.4487 COMET: an absolute gain of 39.62 BLEU. On the official shared-task test set, which is out-of-domain news text, the same system scores 12.51 BLEU and 43.81 chrF++; we analyse this domain gap in detail. The model and corpus are publicly released at https://huggingface.co/agnivamaiti/nllb-200-en-nagamese.