Model Card for Bapynshngain/opus-mt-kha-en-v1

Model Details

Model Description

Bapynshngain/opus-mt-kha-en-v1 is a Transformer-based neural machine translation (NMT) model for translating Khasi to English. The model is fine-tuned from the OPUS-MT multilingual family, adapted for Khasi.

This model is part of an effort to improve digital accessibility and NLP tooling for Khasi through data-centric and transfer learning approaches.

• Developed by: Bapynshngainlang Nongkynrih
• Funded by: Independent / self-initiated research
• Shared by: Bapynshngainlang Nongkynrih
• Model type: Transformer-based Neural Machine Translation (Seq2Seq)
• Language(s): Khasi → English
• License: CC BY-NC-SA 4.0
• Finetuned from model: Helsinki/opus-mt-en-vi
• Demo: N/A

---

Uses

Direct Use

• Khasi → English translation for:
  • Research in low-resource NLP
  • Digital content accessibility
  • Linguistic analysis
  • Dataset creation / augmentation

Downstream Use

• Integration into:
  • Translation pipelines
  • Cross-lingual information retrieval systems
  • Multilingual chatbots
  • Preprocessing for English-centric NLP models

Out-of-Scope Use

• High-stakes domains (legal, medical, financial) without human validation
• Highly domain-specific translation (technical/legal/formal jargon)
• Long-form document translation without segmentation
• English → Khasi translation (not supported)

---

Bias, Risks, and Limitations

• Low-resource constraints: Despite scaling to ~60,000 parallel pairs, coverage remains limited compared to high-resource languages
• Domain bias: Training data may not fully represent all domains or dialectal variation in Khasi
• Structural inconsistencies: Complex sentence structures may degrade translation quality
• Hallucination risk: The model may occasionally generate plausible but unsupported content, particularly on noisy or unfamiliar inputs
• Named entity handling: Proper nouns and rare entities may be mistranslated or dropped

---

How to Get Started with the Model

from transformers import MarianMTModel, MarianTokenizer

model_name = "Bapynshngain/opus-mt-kha-en-v1"

tokenizer = MarianTokenizer.from_pretrained(model_name)
model = MarianMTModel.from_pretrained(model_name)

text = "Phi long kumno?"
inputs = tokenizer(text, return_tensors="pt", padding=True)

translated = model.generate(**inputs)
output = tokenizer.decode(translated[0], skip_special_tokens=True)

print(output)