MADLAD-400-3B-MT for Asturian, Aragonese and Aranese, CTranslate2 int8

CTranslate2 conversion (int8_float16 weights, 2.9 GB) of Ugiat/madlad400-3b-mt-ast-an-oc_aranes, the fine-tune of google/madlad400-3b-mt that translates into Asturian (<2ast>), Aragonese (<2an>) and Aranese (<2oc_aranes>) from Spanish, Catalan, English and French, and out of them with the base model's tags (<2es>, <2ca>, <2en>, <2fr>). This version is faster and lighter than the transformers version, for production translation and CPU serving. See the main model card for examples, the comparison with other models, evaluation and limitations.

This model was prepared by Ugiat Technologies. It is an independent work and is not produced, reviewed or endorsed by the TAN-IBE project, by Google or by the MADLAD-400 authors.

Usage

import ctranslate2
from huggingface_hub import snapshot_download
from transformers import AutoTokenizer

path = snapshot_download("Ugiat/madlad400-3b-mt-ast-an-oc_aranes-ct2")
tokenizer = AutoTokenizer.from_pretrained(path)
translator = ctranslate2.Translator(path, device="cuda", compute_type="int8_float16")   # or device="cpu", compute_type="int8"

def translate(sentences, tag):
    batch = [tokenizer.convert_ids_to_tokens(tokenizer.encode(f"{tag} {s}")) for s in sentences]
    return [tokenizer.decode(tokenizer.convert_tokens_to_ids(r.hypotheses[0]), skip_special_tokens=True)
            for r in translator.translate_batch(batch, beam_size=4)]

for tag in ("<2ast>", "<2an>", "<2oc_aranes>"):
    print(tag, translate(["The town council has approved the new budget for next year."], tag))
# <2ast> ['El conceyu aprobó\'l nuevu presupuestu pa l\'añu que viene.']
# <2an> ['O concello ha aprobau o nuevo presupuesto ta l\'anyo que viene.']
# <2oc_aranes> ['Er ajuntament a aprovat eth nau pressupòst entar an que ven.']

• Use this repository's tokenizer (<2oc_aranes> and <2ast> are pieces 256000 and 256001 of its SentencePiece model) with transformers 4.x.
• CTranslate2 4.x wheels are built for CUDA 12. If your environment only has CUDA 13 libraries (e.g. a recent PyTorch), install nvidia-cublas-cu12 and make libcublas.so.12 loadable before importing ctranslate2.
• Other target tags work as in the base model (<2fr>, <2de>, <2en>…).

Conversion

ct2-transformers-converter --model Ugiat/madlad400-3b-mt-ast-an-oc_aranes --quantization int8_float16 --output_dir ct2

(converted with CTranslate2 4.8.2 and transformers 4.57.6, which loads MADLAD-400 checkpoints correctly; 5.x does not).

License and attribution

Apache-2.0, like the base model. The training data and its licences, including the TAN-IBE corpora, are listed on the main model card. Please cite MADLAD-400 (Kudugunta et al., 2023, arXiv:2309.04662) and the main model.

Contact

Ugiat Technologies, ugiat@ugiat.com.