Skip to main content

Token translation for language models.

Project description

transtokenizers

pypi python

Token translation for language models

Features

  • Translate a model from one language to another.
  • Support for most scripts beyond Latin.

Installation

pip install trans-tokenizers

Usage

You do need an installation of fast_align to align the tokens. You can install from the following repo: https://github.com/FremyCompany/fast_align.

To convert a Llama model from English to Dutch, you can use the following code. This might

from transtokenizers import create_aligned_corpus, align, map_tokens, smooth_mapping, remap_model
from transformers import AutoTokenizer, AutoModelForCausalLM
import os

source_model = "meta-llama/Meta-Llama-3-8B"

target_tokenizer = "yhavinga/gpt-neo-1.3B-dutch"
export_dir = "en-nl-llama3-8b"

corpus = create_aligned_corpus(
    source_language="en",
    target_language="nl",
    source_tokenizer=source_model,
    target_tokenizer=target_tokenizer,
)

mapped_tokens_file = align(corpus, fast_align_path="fast_align")

tokenized_possible_translations, untokenized_possible_translations = map_tokens(mapped_tokens_file, source_model, target_tokenizer)

smoothed_mapping = smooth_mapping(target_tokenizer, tokenized_possible_translations)

model = remap_model(source_model, target_tokenizer, smoothed_mapping, source_model)
os.makedirs(export_dir, exist_ok=False)
new_tokenizer = AutoTokenizer.from_pretrained(target_tokenizer)
model.save_pretrained(export_dir)
new_tokenizer.save_pretrained(export_dir)

Credits

If this repo was useful to you, please cite the following paper

@inproceedings{remy-delobelle2024transtokenization,
    title={Trans-Tokenization and Cross-lingual Vocabulary Transfers: Language Adaptation of {LLM}s for Low-Resource {NLP}},
    author={Remy, Fran{\c{c}}ois and Delobelle, Pieter and Avetisyan, Hayastan and Khabibullina, Alfiya and de Lhoneux, Miryam and Demeester, Thomas},
    booktitle={First Conference on Language Modeling},
    year={2024},
    url={https://openreview.net/forum?id=sBxvoDhvao}
}

Project details


Download files

Download the file for your platform. If you're not sure which to choose, learn more about installing packages.

Source Distribution

trans_tokenizers-0.1.3.tar.gz (12.2 kB view hashes)

Uploaded Source

Built Distribution

trans_tokenizers-0.1.3-py3-none-any.whl (12.3 kB view hashes)

Uploaded Python 3

Supported by

AWS AWS Cloud computing and Security Sponsor Datadog Datadog Monitoring Fastly Fastly CDN Google Google Download Analytics Microsoft Microsoft PSF Sponsor Pingdom Pingdom Monitoring Sentry Sentry Error logging StatusPage StatusPage Status page