kathe2026-kashmiri-lora

KATHE 2026 — English to Kashmiri Machine Translation

[!WARNING] DEVELOPER WARNING — READ THIS BEFORE TOUCHING THE CODE

Dear programmer:

When I wrote this code, only God and I knew how it worked.

Now, only God knows it.

Therefore, if you are trying to optimize this routine and it fails (most surely), please increase this counter as a warning for the next person:

total_hours_wasted_here = 254

If you understand this code immediately, please don't tell me. I don't want to know.

Proceed at your own risk.

Model Details

Model Description

• Developed by: Mehru
• Shared by: Mehru
• Model type: LoRA / PEFT adapter for neural machine translation
• Task: English → Kashmiri machine translation
• Source language: English (en)
• Target language: Kashmiri (ks)
• Target script: Perso-Arabic (kas_Arab)
• License: MIT
• Fine-tuned from: ai4bharat/indictrans2-en-indic-dist-200M
• Training framework: PEFT
• Precision: FP16
• Hardware: NVIDIA T4 GPU

This model is a LoRA fine-tuned adapter designed to adapt the AI4Bharat IndicTrans2 English-to-Indic model for English → Kashmiri machine translation.

Model Sources

• Repository: KATHE 2026 LoRA Model
• Base Model: ai4bharat/indictrans2-en-indic-dist-200M
• Demo: Not provided
• Paper: Not applicable to this fine-tuned adapter

Uses

Direct Use

The adapter can be used with the AI4Bharat IndicTrans2 English-to-Indic base model to perform English → Kashmiri translation.

It is intended for:

• English → Kashmiri translation
• Kashmiri NLP research
• Machine translation experiments
• Educational purposes
• Research prototypes
• Development of Kashmiri language technology

The adapter must be loaded together with the original base model.

Downstream Use

The model can be integrated into applications such as:

• Translation tools
• Language-learning applications
• Kashmiri NLP systems
• Research projects
• Text-processing pipelines
• Translation prototypes

Any downstream application should independently evaluate translation quality for its intended domain.

Out-of-Scope Use

This model should not be relied upon as the sole source of translation for:

• Legal documents
• Medical information
• Financial documents
• Safety-critical communication
• Official documents
• High-stakes decisions

Human review is recommended when translation accuracy is important.

Bias, Risks, and Limitations

Machine translation systems can produce incorrect, incomplete, or contextually inappropriate translations.

Potential limitations include:

• Vocabulary coverage
• Sentence complexity
• Domain-specific terminology
• Proper nouns
• Ambiguous English expressions
• Kashmiri linguistic variation
• Differences between formal and conversational language
• Possible grammatical or semantic errors

The model was fine-tuned using a specific English–Kashmiri training corpus and may perform differently on domains or writing styles that are not well represented in the training data.

Recommendations

Users should evaluate translations before using them in important applications.

For high-stakes or official content, human verification is recommended.

How to Get Started with the Model

The model is a LoRA adapter and should be loaded on top of the original IndicTrans2 base model.

from transformers import AutoModelForSeq2SeqLM, AutoTokenizer
from peft import PeftModel

base_model = "ai4bharat/indictrans2-en-indic-dist-200M"
adapter_model = "codewithmehru/kathe2026-kashmiri-lora"

tokenizer = AutoTokenizer.from_pretrained(base_model)

model = AutoModelForSeq2SeqLM.from_pretrained(base_model)
model = PeftModel.from_pretrained(model, adapter_model)

For the complete preprocessing and inference workflow, use the project repository associated with the KATHE 2026 submission.

Training Details

Training Data

The model was fine-tuned using:

• Dataset: BPCC — Bharat Parallel Corpus Collection
• Subset: bpcc-seed-latest
• Language pair: English–Kashmiri
• Target script: kas_Arab
• Sentence pairs: 98,929
• Dataset provider: AI4Bharat, IIT Madras

No additional Kashmiri-English dataset is claimed as training data for this model.

Training Procedure

The base model was fine-tuned using Low-Rank Adaptation (LoRA).

The base model parameters were kept frozen while approximately 6.5M trainable LoRA parameters were optimized for the English → Kashmiri translation task.

Preprocessing

The training data consisted of parallel English–Kashmiri sentence pairs.

The target language uses the Kashmiri Perso-Arabic script (kas_Arab).

Preprocessing was performed using IndicTransToolkit utilities.

Training Hyperparameters

• Fine-tuning method: LoRA
• LoRA rank: 16
• LoRA alpha: 32
• LoRA dropout: 0.05
• Target modules: q_proj, k_proj, v_proj, out_proj, fc1, fc2
• Learning rate: 2e-4
• Per-device batch size: 8
• Gradient accumulation: 4
• Effective batch size: 32
• Precision: FP16
• GPU: NVIDIA T4
• Epochs: Not reported

Approximately 6.5M parameters were trained, representing approximately 2.97% of the approximately 218M-parameter base model.

Speeds, Sizes, Times

• Base model size: approximately 218M parameters
• Trainable parameters: approximately 6.5M
• Trainable percentage: approximately 2.97%
• Training GPU: NVIDIA T4
• Training time: Not reported

Evaluation

Testing Data, Factors & Metrics

Testing Data

A separate evaluation dataset or held-out test set was not specified in the provided training information.

Therefore, official test-set results are not reported in this model card.

Factors

Important factors for evaluating this model include:

• English sentence complexity
• Kashmiri vocabulary
• Grammatical structure
• Proper nouns
• Domain-specific terminology
• Formal versus conversational language
• Translation fluency
• Semantic preservation

Metrics

No official evaluation metrics were provided for this fine-tuned adapter.

• BLEU: Not reported
• chrF: Not reported
• chrF++: Not reported
• COMET: Not reported

Results

Official quantitative evaluation results for this fine-tuned adapter have not been reported.

Summary

The model is a LoRA adaptation of IndicTrans2 specifically targeting English → Kashmiri translation.

The primary objective of the fine-tuning process was to adapt the multilingual IndicTrans2 model to Kashmiri translation using the BPCC English–Kashmiri parallel corpus.

Model Examination

This model is an adapter rather than a complete standalone base model.

It requires the original:

ai4bharat/indictrans2-en-indic-dist-200M

base model for inference.

The adapter contains the learned LoRA parameters produced during fine-tuning.

Environmental Impact

Training was performed using an NVIDIA T4 GPU.

The following information was not recorded or reported:

• Hardware Type: NVIDIA T4
• Hours used: Not reported
• Cloud Provider: Not reported
• Compute Region: Not reported
• Carbon Emitted: Not reported

Carbon emissions can be estimated using the Machine Learning Impact calculator if the required training-time and infrastructure information becomes available.

Technical Specifications

Model Architecture and Objective

The model is based on the AI4Bharat IndicTrans2 English-to-Indic distilled 200M model.

The fine-tuned component is a LoRA adapter trained for English → Kashmiri machine translation.

The objective is to adapt the base multilingual translation model to Kashmiri translation using parameter-efficient fine-tuning.

Compute Infrastructure

Hardware

• GPU: NVIDIA T4
• Precision: FP16

Software

• Python
• PyTorch
• Transformers
• PEFT
• IndicTransToolkit

Acknowledgments

• AI4Bharat — for the IndicTrans2 base model and BPCC dataset
• IndicTransToolkit — for preprocessing utilities
• Gaash Lab, NIT Srinagar — for organizing the KATHE 2026 challenge

Citation

If you use this model, please credit the KATHE 2026 fine-tuned adapter and its developer.

KATHE 2026 Model

Mehru. KATHE 2026 — English to Kashmiri Machine Translation.
LoRA fine-tuned adapter based on AI4Bharat IndicTrans2.

Base Model Citation

@article{gala2023indictrans,
  title={IndicTrans2: Towards High-Quality and Accessible Machine Translation Models for all 22 Scheduled Indian Languages},
  author={Jay Gala and Pranjal A Chitale and A K Raghavan and Varun Gumma and Sumanth Doddapaneni and Aswanth Kumar M and Janki Atul Nawale and Anupama Sujatha and Ratish Puduppully and Vivek Raghavan and Pratyush Kumar and Mitesh M Khapra and Raj Dabre and Anoop Kunchukuttan},
  journal={Transactions on Machine Learning Research},
  year={2023},
  url={https://openreview.net/forum?id=vfT4YuzAYA}
}

Glossary

• LoRA: Low-Rank Adaptation, a parameter-efficient fine-tuning method.
• PEFT: Parameter-Efficient Fine-Tuning.
• IndicTrans2: AI4Bharat's multilingual machine translation model for Indian languages.
• BPCC: Bharat Parallel Corpus Collection.
• kas_Arab: Kashmiri language represented using the Perso-Arabic script.
• FP16: 16-bit floating-point numerical precision.

More Information

This model was developed as part of the KATHE 2026 English → Kashmiri Machine Translation challenge.

Developer

Mehru

Developer and Machine Learning enthusiast focused on AI, machine translation, and language technology.

• GitHub
• LinkedIn
• Portfolio

Model Repository

codewithmehru/kathe2026-kashmiri-lora

Model Card Authors

Mehru

Model Card Contact

For questions, issues, or collaboration regarding this model, please use the model repository's community/discussion section:

Hugging Face Discussions

Framework Versions

• PEFT: 0.13.2
• Precision: FP16
• Hardware: NVIDIA T4

License

MIT License.