135 lines
3.0 KiB
Markdown
135 lines
3.0 KiB
Markdown
---
|
||
language:
|
||
- ks
|
||
- en
|
||
license: apache-2.0
|
||
base_model:
|
||
- Omarrran/koshur-kouter-ks-en_v1
|
||
library_name: transformers
|
||
pipeline_tag: text-generation
|
||
tags:
|
||
- kashmiri
|
||
- gemma3
|
||
- continual-pretraining
|
||
- low-resource
|
||
- language-model
|
||
---
|
||
# KoshurAI_Tarjuma_v2 — Kashmiri Continual Pretraining Base
|
||
> ⚠️ **This is a base language model, not a translation model.**
|
||
> For Kashmiri ↔ English translation, use the fine-tuned adapter:
|
||
> [`Faizaniqbal/KoshurAI_Tarjuma_v3`](https://huggingface.co/Faizaniqbal/KoshurAI_Tarjuma_v3)
|
||
---
|
||
## What is this?
|
||
`KoshurAI_Tarjuma_v2` is a **5B-parameter Gemma 3** model that has been
|
||
continually pretrained on **2.8 million tokens of native Kashmiri text**,
|
||
giving it deep knowledge of the Kashmiri language (Perso-Arabic script).
|
||
It serves as the **Stage 1 base** in the KoshurAI two-stage pipeline:
|
||
```
|
||
Omarrran/koshur-kouter-ks-en_v1 ← fine-tuned on Kashmiri–English pairs
|
||
↓ continual pretraining on 2.8M Kashmiri tokens
|
||
Faizaniqbal/KoshurAI_Tarjuma_v2 ← this model (language knowledge)
|
||
↓ SFT on 16,637 EN↔KS pairs (LoRA adapter)
|
||
Faizaniqbal/KoshurAI_Tarjuma_v3 ← final translation model
|
||
```
|
||
---
|
||
## Model Details
|
||
|
|
||
|
|
||
|
|
||
|
|
||
---
|
||
|
|
||
---
|
||
|
|
||
|
|
||
**
|
||
Author
|
||
**
|
||
|
|
||
Faizan Iqbal (
|
||
[
|
||
@Faizaniqbal
|
||
](
|
||
https://huggingface.co/Faizaniqbal
|
||
)
|
||
)
|
||
|
|
||
|
|
||
**
|
||
Base model
|
||
**
|
||
|
|
||
`Omarrran/koshur-kouter-ks-en_v1`
|
||
|
|
||
|
|
||
**
|
||
Architecture
|
||
**
|
||
|
|
||
Gemma3ForCausalLM (5B parameters)
|
||
|
|
||
|
|
||
**
|
||
Tensor type
|
||
**
|
||
|
|
||
BF16
|
||
|
|
||
|
|
||
**
|
||
Pretraining data
|
||
**
|
||
|
|
||
2.8M tokens of Kashmiri text
|
||
|
|
||
|
|
||
**
|
||
Languages
|
||
**
|
||
|
|
||
Kashmiri (ks · kas_Arab), English (en)
|
||
|
|
||
|
|
||
**
|
||
License
|
||
**
|
||
|
|
||
Apache-2.0
|
||
|
|
||
---
|
||
## Pretraining Corpus
|
||
The 2.8M token corpus was assembled from two sources:
|
||
- **InPage documents** — professionally published Kashmiri literature,
|
||
journalism, academic scholarship, and religious texts spanning multiple
|
||
decades, converted to Unicode via a custom InPage converter.
|
||
- **Native speaker translations** — texts translated into Kashmiri by
|
||
native speakers, providing natural human-authored language coverage.
|
||
---
|
||
## Full Model Lineage
|
||
```
|
||
google/gemma-3-4b-it (Google)
|
||
└─ sarvamai/sarvam-translate (Sarvam AI)
|
||
└─ Omarrran/koshur-kouter-ks-en_v1 (Malik & Nissar, 2026)
|
||
└─ Faizaniqbal/KoshurAI_Tarjuma_v2 ← this model
|
||
└─ Faizaniqbal/KoshurAI_Tarjuma_v3 (translation adapter)
|
||
```
|
||
---
|
||
## Citation
|
||
```bibtex
|
||
@misc{iqbal2026koshurai,
|
||
title = {KoshurAI v3: A Fine-Tuned Neural Machine Translation System
|
||
for Kashmiri--English Bidirectional Translation},
|
||
author = {Iqbal, Faizan},
|
||
year = {2026},
|
||
howpublished = {\url{https://huggingface.co/Faizaniqbal/KoshurAI_Tarjuma_v3}}
|
||
}
|
||
```
|
||
Please also cite the base model:
|
||
```bibtex
|
||
@misc{malik2026koshurkouter,
|
||
title = {Koshur Kouter KS-EN v1: A Merged QLoRA Kashmiri--English Translation Model},
|
||
author = {Malik, Haq Nawaz and Nissar, Nahfid},
|
||
year = {2026},
|
||
howpublished = {\url{https://huggingface.co/Omarrran/koshur-kouter-ks-en_v1}}
|
||
}
|
||
``` |