mohamadpk/nllb-200-3.3B-ct2-int8_float16

NLLB-200 3.3B CTranslate2 int8_float16 Bundle

This folder contains an optimized CTranslate2 build of facebook/nllb-200-3.3B plus helper scripts. It is focused on fast, high-quality English to Persian (en->fa) and Persian to English (fa->en) translation.

Model link

• Hugging Face model: https://huggingface.co/mohamadpk/nllb-200-3.3B-ct2-int8_float16
• GitHub repo: https://github.com/mongo-driver/nllb-200-3.3B-ct2-int8_float16
• GitHub version of this repo excludes model.bin to keep repository size manageable.

Keywords: english to persian translation, persian to english translation, en fa, fa en, farsi translation, NLLB, CTranslate2

File details

File	Required	Purpose
model.bin	Yes	Quantized CTranslate2 weights (int8_float16). This is the main model payload used at inference time.
config.json	Yes	CTranslate2 runtime configuration (model architecture and decoding metadata expected by CT2).
shared_vocabulary.json	Yes	Vocabulary mapping used by CTranslate2 during token generation.
sentencepiece.bpe.model	Yes	SentencePiece tokenizer model. Encodes input text to subword pieces and decodes output pieces back to text.
tokenizer.json	Recommended	Full tokenizer definition used by Hugging Face tooling and some integrations.
tokenizer_config.json	Recommended	Tokenizer runtime options (special token behavior, defaults).
special_tokens_map.json	Recommended	Mapping for special tokens (language tags, BOS/EOS style tokens).
translate.py	Optional helper	Example EN<->FA inference script using CTranslate2 + SentencePiece.
convert_model.py	Optional helper	One-time conversion script from HF model to CT2 int8_float16.
check_gpu.py	Optional helper	Simple CUDA availability check in local PyTorch environment.
requirements.txt	Optional helper	Python dependencies for helper scripts (ctranslate2, sentencepiece, etc.).

Quick usage

Run commands from the project root:

venv\Scripts\pip.exe install -r nllb-200-3.3B-ct2-int8_float16\requirements.txt
venv\Scripts\python.exe nllb-200-3.3B-ct2-int8_float16\check_gpu.py
venv\Scripts\python.exe nllb-200-3.3B-ct2-int8_float16\translate.py

Benchmark results (CUDA)

Benchmark mode: HF float16 GPU vs CTranslate2 int8_float16 GPU

#	Direction	HF time	CT2 time	Speedup
1	en->fa	36.41s	5.25s	6.94x
2	en->fa	15.74s	5.99s	2.63x
3	en->fa	47.74s	3.74s	12.76x
4	fa->en	32.98s	2.98s	11.07x
Total	mixed	132.87s	17.96s	7.40x

Benchmark environment:

• GPU: NVIDIA GeForce RTX 2060 SUPER (8GB VRAM)
• Runtime mode: CUDA
• Comparison: HF float16 GPU vs CT2 int8_float16 GPU

Quality snapshots:

1. en->fa
   Input (EN):
   In the Golang programming language, we can use defer to execute code at the end of a function.
   HF (FA):
   در زبان برنامه نویسی گولانگ گولانگ، ما می توانیم از defer استفاده کنیم تا کد را در پایان تابع اجرا کنیم.
   CT2 (FA):
   در زبان برنامه نویسی Golang، ما می توانیم از defer برای اجرای کد در پایان یک تابع استفاده کنیم.

---

2. en->fa
   Input (EN):
   Artificial intelligence is transforming the world.
   HF (FA):
   هوش مصنوعی هوش مصنوعی در حال تغییر جهان است
   CT2 (FA):
   هوش مصنوعی در حال تغییر جهان است.

---

3. en->fa
   Input (EN):
   Machine learning models require large amounts of data to train effectively.
   HF (FA):
   مدل های یادگیری ماشینی مدل های یادگیری ماشین ماشین مدل های یادگیری یادگیری ماشین ماشین نیاز به مقدار زیادی از داده برای آموزش به طور موثر نیاز به داده های بزرگ برای آموزش موثر. .
   CT2 (FA):
   مدل های یادگیری ماشین به مقادیر زیادی از داده ها برای آموزش موثر نیاز دارند.

---

4. fa->en
   Input (FA):
   پردازش زبان طبیعی یکی از شاخه های مهم هوش مصنوعی است.
   HF (EN):
   Natural Language Processing (NLP) is one of the most important branches of AI.
   CT2 (EN):
   Natural language processing is one of the important branches of artificial intelligence.

Attribution

• Base model: facebook/nllb-200-3.3B
• Original model authors: Meta AI / NLLB team
• This package is a converted and quantized derivative (CTranslate2 int8_float16)

Notes

• translate.py is self-contained and loads model/tokenizer files from this folder.
• convert_model.py is for local one-time conversion from facebook/nllb-200-3.3B.

License

• Base/derived model artifacts follow: CC-BY-NC-4.0
• Non-commercial use only, unless you obtain separate permission from rights holders.
• Keep attribution to the base model when redistributing this package.