KashmirAI NLLB-1.3B
Kashmiri-to-English translation
2 min readupdated
A fine-tuned version of facebook/nllb-200-distilled-1.3B (Meta's base language model with distilled parameters), optimized for translating Kashmiri in Perso-Arabic script into English. It has been converted to the CTranslate2 int8 format to run efficiently on standard CPUs without powerful GPUs. Replacing the developer's earlier 600M model, this release powers the project's website and is well-suited for researchers, linguists, and developers building tools for low-resource languages.
Key features
- Built on Meta's NLLB-200 distilled 1.3B architecture.
- Exported to CTranslate2 int8 for fast CPU execution.
- Trained on a dataset of 128,028 carefully filtered sentence pairs.
- Delivers higher translation quality than the previous 600M version.
What you can do
- Integrate the model into a local translator or web service supporting Kashmiri.
- Run inference on a standard CPU at roughly 1–2.5 seconds per sentence.
- Use it for natural language processing (NLP) research focused on regional Indian languages.
Example
import ctranslate2
from tokenizers import Tokenizer
from huggingface_hub import snapshot_download
path = snapshot_download("Faizanayub/kashmirai-nllb13b-ks-en-ct2")
tr = ctranslate2.Translator(path, device="cpu", compute_type="int8")
tok = Tokenizer.from_file(f"{path}/tokenizer.json")
src = "سُہ چھُ آز اکھ کامیاب ویب ڈزاینَر۔"
tokens = ["kas_Arab"] + tok.encode(src, add_special_tokens=False).tokens + [""]
out = tr.translate_batch([tokens], target_prefix=[["eng_Latn"]], beam_size=4)[0].hypotheses[0]
ids = [tok.token_to_id(t) for t in out if t not in ("eng_Latn", "")]
print(tok.decode(ids, skip_special_tokens=True))
Read more
DetailsBenchmarks (KashEval v1): FLORES-200 devtest — 61.41 chrF++; Google SMOL — 49.36 chrF++; IN22-Gen — 61.26 chrF++.
Hardware requirements: optimized for CPU (single-threaded inference takes 1–2.5 seconds per sentence).
Training details: full fine-tune with frozen embeddings on a Kaggle T4, batch size 32, 2 epochs.
License: CC BY-NC 4.0.










Comments