KATHE / KatheBathe

English → Kashmiri Machine Translation

KATHE 2026/ KatheBathe is an English → Kashmiri machine translation model developed by Muqarab Farooq Vaid and Suhaib Fida for KATHE 2026.

The model is fine-tuned from sarvamai/sarvam-translate using QLoRA / LoRA with PEFT.

LINK : https://excalidraw.com/#json=fmEi9loC9z-9HUeORrQ4Y,cBAFupy9-eVpeILLJf1Yjg

Links
- Hugging Face
- GitHub
How to Run
- Option 1 - Run from Hugging Face (Preferred)
- Option 2 - Run from GitHub
Automatic Model Download
Model Information
Methodology
Training Data
Installation
Inference
Quick Inference Test
Single-Sentence Inference
Batch Inference

- Kaggle Usage: { Also preferred for easy check }

Input CSV
Manual CSV
Custom Text
Inference Flow
Generation Settings
- EOS Token - Required Setting
- Output Cleanup - Removing Trailing Lines
Technical Details
Repository Structure
Reproducibility
Intended Use
Limitations
Evaluation
Team
Citations
License
Acknowledgements
Model Card Authors

How to Run

There are two ways to run KatheBathe:

Option 1 — Hugging Face (Preferred)
Option 2 — GitHub

Recommended: We recommend Option 1 (Hugging Face) because it downloads the submitted inference.py directly from the Hugging Face model repository and uses the submitted model and adapter.

Option 1 - Run from Hugging Face (Preferred)

The preferred method is to download the submitted inference script directly from the KatheBathe Hugging Face repository.

Hugging Face Model:

https://huggingface.co/KatheBathe/Kathe-Bathe

1. Install Dependencies

Install the required packages:

pip install -r requirements.txt

If huggingface_hub is not already installed:

pip install -U huggingface_hub

2. Download `inference.py`

Use hf_hub_download to download the exact submitted inference script:

from huggingface_hub import hf_hub_download

hf_hub_download(
    repo_id="KatheBathe/Kathe-Bathe",
    filename="inference.py",
    local_dir="/kaggle/working",
    force_download=True,
)

print("inference.py downloaded")

This downloads:

KatheBathe/Kathe-Bathe
        ↓
   inference.py
        ↓
/kaggle/working/inference.py

3. Run the Inference Script

python /kaggle/working/inference.py

The script automatically loads:

Base model
KatheBathe adapter
Tokenizer
Model configuration

from Hugging Face.

You do not need to manually download the model weights.

Option 2 - Run from GitHub

The complete inference code is also available through GitHub.

GitHub Repository:

https://github.com/suhaibfida/Kathe-Bathe

1. Clone the Repository

git clone https://github.com/suhaibfida/Kathe-Bathe.git
cd Kathe-Bathe

2. Install Dependencies

pip install -r requirements.txt

3. Run the Inference Script

python inference.py

The GitHub inference.py loads the KatheBathe model and adapter from Hugging Face.

The model weights do not need to be manually copied into the GitHub repository.

Automatic Model Download

You do not need to manually download the model weights.

The provided inference.py automatically loads the required model components from Hugging Face.

The loading process is:

Run inference.py
       ↓
Load tokenizer
       ↓
Load BF16 base model
       ↓
Load KatheBathe QLoRA adapter
       ↓
Prepare model
       ↓
Generate translation

The base model is:

sarvamai/sarvam-translate

The submitted adapter is:

KatheBathe/Kathe-Bathe

The script automatically downloads the required files if they are not already available locally.

You do not need to manually download:

Adapter weights
Adapter configuration
Tokenizer files
Model configuration files

The first run may take longer because the model files need to be downloaded.

Subsequent runs can reuse the locally cached files.

Internet access is required when the required model files are not already cached locally.

A merged model is not required for the provided inference script.

Model Information

Configuration	Value
Task	English → Kashmiri
Base Model	`sarvamai/sarvam-translate`
Base Architecture	Gemma 3 4B IT
Fine-tuning	QLoRA / LoRA
Framework	PEFT
Inference dtype	BF16
Maximum input length	1024
Maximum new tokens	232
Beam size	6
Repetition penalty	1.15
No-repeat n-gram size	3
EOS token ID	`1` (`<eos>`)
Default batch size	16
License	GPL-3.0

Methodology

KatheBathe was developed using parameter-efficient fine-tuning with QLoRA / LoRA and PEFT.

Sarvam-Translate
       ↓
Gemma 3 4B IT
       ↓
QLoRA / LoRA Fine-tuning
       ↓
KatheBathe Adapter
       ↓
English → Kashmiri Translation

The base model used for fine-tuning is:

sarvamai/sarvam-translate

The trained adapter is loaded on top of the original base model during inference.

A merged model is not required for the provided inference script.

Training Data

The model was fine-tuned using the following datasets.

SMU Qamar - Kashmiri-English Parallel Corpus

Dataset:

https://huggingface.co/datasets/SMUQamar/Kashmiri-English-Parallel-Corpus

AI4Bharat - BPCC

Dataset:

https://huggingface.co/datasets/ai4bharat/BPCC

Installation

Requirements

Before running the inference script, you need:

Python
pip
An NVIDIA GPU with BF16 support for the submitted inference configuration
Internet access for downloading model files if they are not already cached

Install all Python dependencies:

pip install -r requirements.txt

The required packages are:

transformers==4.51.3
peft==0.15.2
accelerate
sentencepiece
safetensors
pandas
huggingface_hub

The provided inference setup:

Uses BF16 inference
Does not use 4-bit bitsandbytes quantization
Uses PEFT for loading the QLoRA adapter
Uses Hugging Face Hub for downloading the submitted inference script and model resources

Inference

The repository contains a single inference script:

inference.py

The script supports:

Model loading
Single-sentence inference
Multiple custom sentences
Batch CSV inference
Automatic Kaggle CSV detection
Manual CSV input
Output validation
EOS-controlled generation
Output cleanup

Quick Inference Test

The quickest way to verify that the model, tokenizer, adapter, and inference code are working is:

python inference.py --text "She was a true visionary."

The script will:

Load the tokenizer.
Load the base model.
Load the KatheBathe adapter.
Generate a Kashmiri translation.
Display the result.

Example:

Input:
She was a true visionary.

Output:
سۄ ٲس اکھ حقیقی بصیرت تھون واجیٚنۍ۔

If a translation is generated successfully, the inference setup is working.

Single-Sentence Inference

Translate a single English sentence:

python inference.py \
    --text "She was a true visionary."

You can also provide multiple sentences:

python inference.py \
    --text "She was a true visionary." \
    --text "The weather is beautiful today."

Batch Inference

For batch inference, the input CSV must contain:

ID,sentence

Example:

ID,sentence
1,She was a true visionary.
2,The weather is beautiful today.
3,I like learning new things.

Run:

python inference.py \
    --input /path/to/test.csv \
    --output predictions.csv

The output contains:

ID,kashmiri_text

Example:

ID,kashmiri_text
1,سۄ ٲس اکھ حقیقی بصیرت تھون واجیٚنۍ۔
2,...
3,...

The script validates:

Prediction count
Empty predictions
ID order
Output columns
Saved CSV row count
Saved CSV IDs
Empty translations

Kaggle Usage

1. Enable GPU

In Kaggle:

Notebook
→ Settings
→ Accelerator
→ GPU

Use an NVIDIA GPU with BF16 support.

2. Download `inference.py`

The preferred method is to download the exact submitted inference script directly from Hugging Face:

from huggingface_hub import hf_hub_download

hf_hub_download(
    repo_id="KatheBathe/Kathe-Bathe",
    filename="inference.py",
    local_dir="/kaggle/working",
    force_download=True,
)

print("inference.py downloaded")

This downloads:

KatheBathe/Kathe-Bathe
        ↓
   inference.py
        ↓
/kaggle/working/inference.py

3. Run

!python /kaggle/working/inference.py

The script automatically searches:

/kaggle/input/**/*.csv

for compatible CSV files.

The expected input columns are:

ID
sentence

If multiple compatible CSV files are found, specify the input manually:

!python /kaggle/working/inference.py \
    --input /kaggle/input/my-dataset/test.csv \
    --output /kaggle/working/predictions.csv

Input CSV

The required columns are:

ID
sentence

Example:

ID,sentence
1,She was a true visionary.
2,The weather is beautiful today.
3,I like learning new things.

Manual CSV

Specify the input and output paths:

!python inference.py \
    --input /path/to/test.csv \
    --output /path/to/predictions.csv

Custom Text

Single Sentence

!python inference.py \
    --text "She was a true visionary."

Multiple Sentences

!python inference.py \
    --text "She was a true visionary." \
    --text "The weather is beautiful today."

Inference Flow

Load tokenizer
      ↓
Load BF16 base model
      ↓
Load QLoRA adapter
      ↓
One-sentence diagnostic
      ↓
Translate input
      ↓
Validate predictions
      ↓
Save CSV
      ↓
Print first 10 results
      ↓
ALL CHECKS PASSED

The one-sentence inference is used as a diagnostic.

The first 10 results printed at the end are previews of predictions that have already been generated.

Generation Settings

MAX_INPUT_LENGTH = 1024
MAX_NEW_TOKENS = 232

NUM_BEAMS = 6
REPETITION_PENALTY = 1.15
NO_REPEAT_NGRAM_SIZE = 3

# Explicit EOS token - required
EOS_TOKEN_ID = 1

Generation uses deterministic decoding:

do_sample=False

Default batch size:

If GPU memory is insufficient:

python inference.py --batch-size 8

EOS Token - Required Setting

eos_token_id must be set to 1, the tokenizer's <eos> token.

Do not change it to 106, which corresponds to <end_of_turn>.

The inference configuration uses:

eos_token_id = 1

The two settings behave differently:

Setting	Behavior
`eos_token_id = 1` (`<eos>`)	Correct setting used for the submitted model
`eos_token_id = 106` (`<end_of_turn>`)	Do not use for the submitted configuration

In testing, using <end_of_turn> caused generation to stop earlier, but resulted in a lower reported score.

Using:

eos_token_id = 1

allows beam search to continue evaluating candidate sequences until the actual EOS token is reached.

The trade-off is:

Longer generation
Higher compute usage
More tokens may be generated
Some raw outputs may contain trailing content

However, this setting produced the better reported score for the submitted configuration.

Therefore:

Do not change eos_token_id = 1 to 106 if you want to reproduce the submitted configuration.

If runtime is a concern, adjust:

--batch-size
GPU hardware

rather than changing the EOS token.

Output Cleanup - Removing Trailing Lines

Because generation uses the true <eos> token (1) rather than <end_of_turn> (106), the raw decoded output can sometimes contain extra trailing content.

This may include:

Blank lines
Stray special tokens
<unused...> tokens
Repeated content
Additional generated lines

The inference script handles this through post-processing.

After decoding, the script keeps the first valid non-empty translation line and removes trailing content.

This ensures that:

Beam search can use the full generation.
The EOS configuration remains unchanged.
The final translation is clean.
The output CSV contains a single translation line.

If you re-implement or modify inference.py, keep this cleanup step.

The cleanup step should not be replaced by changing the EOS token to 106.

Technical Details

Architecture

Gemma 3 4B IT
      ↓
Sarvam-Translate
      ↓
QLoRA Fine-tuning
      ↓
KatheBathe Adapter

Inference Hardware

BF16 inference is designed for an NVIDIA CUDA GPU with BF16 support.

Software

Transformers 4.51.3
PEFT 0.15.2
Accelerate
SentencePiece
Safetensors
PyTorch
Hugging Face Hub

Repository Structure

KATHE-KatheBathe/
│
├── inference.py
├── requirements.txt
└── README.md

`inference.py`

The main inference script responsible for:

Loading the tokenizer
Loading the base model
Loading the KatheBathe adapter
Single-sentence inference
Batch inference
Prediction validation
Saving predictions
Using the required EOS configuration
Cleaning trailing generated output

`requirements.txt`

Contains the Python packages required to run the inference script.

`README.md`

Contains:

Model information
Methodology
Installation instructions
Inference instructions
Dataset information
Reproducibility information
EOS configuration
Generation settings

Reproducibility

The GitHub inference code and Hugging Face model are designed to work together.

GitHub
   │
   └── inference.py
          │
          ▼
Hugging Face
   │
   ├── KatheBathe Adapter
   └── Tokenizer / Configuration
          │
          ▼
sarvamai/sarvam-translate
          │
          ▼
English → Kashmiri

The adapter is loaded together with the original base model.

A merged model is not required for the included inference script.

The submitted code can therefore be tested directly against the submitted Hugging Face model weights.

Important Generation Configuration

To reproduce the submitted inference configuration:

eos_token_id = 1

Do not replace it with:

eos_token_id = 106

The generation settings should otherwise remain unchanged:

MAX_INPUT_LENGTH = 1024
MAX_NEW_TOKENS = 232
NUM_BEAMS = 6
REPETITION_PENALTY = 1.15
NO_REPEAT_NGRAM_SIZE = 3
do_sample = False

Intended Use

KatheBathe is intended for:

English → Kashmiri translation
Translation applications
Websites
APIs
Batch translation pipelines
Research projects
Evaluation systems
Kaggle inference workflows

The model should not be treated as:

A general-purpose factual knowledge model
An authoritative source of information
A replacement for human review in high-stakes translation
A guaranteed dialect or domain specialist

Limitations

Machine translation can produce:

Incorrect word choices
Grammar errors
Contextual mistakes
Literal translations
Errors with names and uncommon terminology
Dialect or spelling variation

For important translations, human review is recommended.

Evaluation

KatheBathe is intended for English → Kashmiri translation.

No numerical evaluation results are claimed in this repository because an official evaluation table and test-set results are not provided.

The final competition evaluation may use a private test set.

Team

KATHE 2026

Muqarab Farooq Vaid
Suhaib Fida

Year: 2026

Citations

If you use KatheBathe or its training resources in research, projects, or tools, please acknowledge the model creators, base model, and datasets.

KatheBathe

@misc{kathebathe2026,
    title={KATHE / KatheBathe: English-to-Kashmiri Translation Model},
    author={Muqarab Farooq Vaid and Suhaib Fida},
    year={2026},
    publisher={Hugging Face}
}

Sarvam-Translate

Base model:

https://huggingface.co/sarvamai/sarvam-translate

Gemma 3

@article{gemma_2025,
    title={Gemma 3},
    url={https://arxiv.org/abs/2503.19786},
    publisher={Google DeepMind},
    author={Gemma Team},
    year={2025}
}

Paper:

https://arxiv.org/abs/2503.19786

Kashmiri-English Parallel Corpus

Dataset:

https://huggingface.co/datasets/SMUQamar/Kashmiri-English-Parallel-Corpus

Please cite:

Qumar, S.M.U., Azim, M. & Quadri, S.M.K.
Addressing the data gap: building a parallel corpus for Kashmiri language.
Int. J. Inf. Tecnol. (2024).
https://doi.org/10.1007/s41870-024-01979-8

BPCC / IndicTrans2

Dataset:

https://huggingface.co/datasets/ai4bharat/BPCC

Please cite:

@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 K Khapra and Raj Dabre and Anoop Kunchukuttan},
    journal={Transactions on Machine Learning Research},
    issn={2835-8856},
    year={2023},
    url={https://openreview.net/forum?id=vfT4YuzAYA}
}

Paper:

https://openreview.net/forum?id=vfT4YuzAYA

License

This model is released under:

GPL-3.0

Acknowledgements

We acknowledge the creators of:

sarvamai/sarvam-translate
SMUQamar/Kashmiri-English-Parallel-Corpus
ai4bharat/BPCC
Gemma 3

These resources were used in developing KatheBathe.

Model Card Authors

KATHE / KatheBathe

Authors:

Muqarab Farooq Vaid
Suhaib Fida

For questions or issues, use the model repository discussion/issues mechanism on Hugging Face.

Downloads last month: -; Downloads are not tracked for this model. How to track

Inference Providers NEW

This model isn't deployed by any Inference Provider. 🙋 Ask for provider support

Model tree for KatheBathe/Kathe-Bathe

Base model

google/gemma-3-4b-pt

Finetuned

google/gemma-3-4b-it

Finetuned

sarvamai/sarvam-translate

Finetuned

(16)

this model

Datasets used to train KatheBathe/Kathe-Bathe

Paper for KatheBathe/Kathe-Bathe

Gemma 3 Technical Report

Paper • 2503.19786 • Published Mar 25, 2025 • 59

KATHE / KatheBathe

English → Kashmiri Machine Translation

Table of Contents

- Kaggle Usage: { Also preferred for easy check }

Links

Hugging Face

GitHub

How to Run

Option 1 - Run from Hugging Face (Preferred)

1. Install Dependencies

2. Download inference.py

3. Run the Inference Script

Option 2 - Run from GitHub

1. Clone the Repository

2. Install Dependencies

3. Run the Inference Script

Automatic Model Download

Model Information

Methodology

Training Data

SMU Qamar - Kashmiri-English Parallel Corpus

AI4Bharat - BPCC

Installation

Requirements

Inference

Quick Inference Test

Single-Sentence Inference

Batch Inference

Kaggle Usage

1. Enable GPU

2. Download inference.py

3. Run

Input CSV

Manual CSV

Custom Text

Single Sentence

Multiple Sentences

Inference Flow

Generation Settings

EOS Token - Required Setting

Output Cleanup - Removing Trailing Lines

Technical Details

Architecture

Inference Hardware

Software

Repository Structure

inference.py

requirements.txt

README.md

Reproducibility

Important Generation Configuration

Intended Use

Limitations

Evaluation

Team

Citations

KatheBathe

Sarvam-Translate

Gemma 3

Kashmiri-English Parallel Corpus

BPCC / IndicTrans2

License

Acknowledgements

Model Card Authors

Model tree for KatheBathe/Kathe-Bathe

Datasets used to train KatheBathe/Kathe-Bathe

Paper for KatheBathe/Kathe-Bathe

2. Download `inference.py`

2. Download `inference.py`

`inference.py`

`requirements.txt`

`README.md`