nllb-en-kas-lora

NLLB-200 English → Kashmiri LoRA

A LoRA fine-tuned adapter for English-to-Kashmiri machine translation, based on facebook/nllb-200-distilled-600M.

This model was developed for the KATHE 2026: AI Challenge for Kashmiri Language Translation.

Important: This repository contains the trained LoRA adapter, not the full NLLB-200 base model. The base model is loaded separately from Hugging Face.

---

Model Details

Model Description

• Model type: NLLB-200 Transformer with LoRA adapter
• Base model: facebook/nllb-200-distilled-600M
• Task: English → Kashmiri translation
• Source language: English (eng_Latn)
• Target language: Kashmiri Arabic script (kas_Arab)
• Fine-tuning method: LoRA
• PEFT version: 0.19.1
• Training precision: FP16
• Training epochs: 1

LoRA Configuration

---

Intended Use

This adapter is intended for:

• English-to-Kashmiri machine translation
• Research and experimentation in low-resource language translation
• Kashmiri NLP research
• Evaluation in the KATHE 2026 competition

Out-of-Scope Use

This model should not be treated as a general-purpose multilingual model or as a source of authoritative translations.

Translations should be reviewed by a fluent Kashmiri speaker when accuracy is important, particularly for legal, medical, financial, or other high-stakes content.

---

Training Data

The model was fine-tuned using English-Kashmiri parallel sentence data derived from the Bharat Parallel Corpus Collection (BPCC).

The training data consists of English source sentences paired with Kashmiri target sentences written in the Arabic script.

The project used separate training, validation, and test splits.

---

Training Procedure

The model was fine-tuned using Parameter-Efficient Fine-Tuning (PEFT) with LoRA.

Instead of updating the complete NLLB-200 model, LoRA adapters were attached to selected attention projection layers.

Training Configuration

Base model:
facebook/nllb-200-distilled-600M

Fine-tuning:
LoRA

Epochs:
1

Learning rate:
2e-4

LoRA rank:
16

LoRA alpha:
32

LoRA dropout:
0.1

Target modules:
q_proj, v_proj

Effective batch size:
16

FP16:
Enabled