opus-mt-en-it-ct2

CTranslate2 int8 quantized conversion of Helsinki-NLP/opus-mt-en-it, packaged for offline use in Playto.

Purpose

Lightweight en→it (Italian) NMT for the Playto desktop game-language-learning tool, used as a low-spec / no-GPU fallback when local LLM inference is not available.

Source model

• Upstream: Helsinki-NLP/opus-mt-en-it — MarianMT, transformer architecture
• License: CC-BY 4.0
• No fine-tuning, no weight modification — purely a format conversion + int8 quantization of the upstream weights

Conversion

pip install ctranslate2 transformers sentencepiece
ct2-transformers-converter \
  --model Helsinki-NLP/opus-mt-en-it \
  --output_dir opus-mt-en-it \
  --quantization int8 \
  --copy_files source.spm target.spm
tar czf opus-mt-en-it-ct2.tar.gz opus-mt-en-it-ct2

File layout (inside opus-mt-en-it-ct2.tar.gz)

File	Approx Size	Purpose
model.bin	~62MB	CTranslate2 int8 quantized weights
shared_vocabulary.json	~1.5MB	CTranslate2 vocab
source.spm	~800 KB	SentencePiece source tokenizer
target.spm	~800 KB	SentencePiece target tokenizer
config.json	~250 B	CTranslate2 config

Usage

With ctranslate2 (Python)

import ctranslate2
import sentencepiece

translator = ctranslate2.Translator("opus-mt-en-it", device="cpu", compute_type="int8")
sp_source = sentencepiece.SentencePieceProcessor("opus-mt-en-it/source.spm")
sp_target = sentencepiece.SentencePieceProcessor("opus-mt-en-it/target.spm")

source_tokens = sp_source.encode("Hello, how are you?", out_type=str) + ["</s>"]
results = translator.translate_batch([source_tokens])
print(sp_target.decode(results[0].hypotheses[0]))
# → "Ciao, come stai?"

With ct2rs (Rust)

use ct2rs::{Translator, Tokenizer};

let tokenizer = Tokenizer::new("opus-mt-en-it")?;
let translator = Translator::with_tokenizer("opus-mt-en-it", tokenizer, /* config */)?;
let result = translator.translate_batch(&["Hello, how are you?".to_string()], /* options */)?;

Important: MarianMT models require </s> appended to source token sequences. The ct2rs::Tokenizer wrapper handles this automatically; raw SentencePiece calls must add it manually.

Quality

~54% Good rate on Playto fixture corpus. Base opus-mt comparable to tc-big variant in quality, base selected for smaller size (67MB vs 206MB).

For higher-quality translation, Playto's LLM-based translation path is recommended.

Attribution

This is a derivative work of Helsinki-NLP/opus-mt-en-it. License is CC-BY 4.0 inherited from upstream.

Helsinki-NLP. OPUS-MT — Open Machine Translation Models.
https://github.com/Helsinki-NLP/Opus-MT

Disclaimer

NMT output quality is significantly lower than modern LLM-based translation. This model is intended as a lightweight fallback for environments where LLM inference is not viable (= low VRAM, mobile, slow connection). For higher-quality translation, Playto's LLM-based translation path is recommended.