|
--- |
|
license: apache-2.0 |
|
language: |
|
- nl |
|
library_name: transformers |
|
--- |
|
[Pieter Delobelle](https://pieter.ai), [François Remy](https://fremycompany.com), [Miryam de Lhoneux](https://people.cs.kuleuven.be/~miryam.delhoneux/), [Thomas Demeester](https://tdmeeste.github.io) |
|
|
|
<p align="center"> |
|
<img src="https://huggingface.co/DTAI-KULeuven/tweety-7b-dutch/resolve/main/tweety-7b-dutch.png?download=true" alt="Tweety-7b-dutch: A Dutch Large Language Model" width="20%"> |
|
</p> |
|
|
|
# Model Card for tweety-7b-dutch |
|
|
|
tweety-7b-dutch is a foundation model with a focus on the Dutch language, incorporating a [Dutch tokenizer](https://huggingface.co/yhavinga/gpt-neo-1.3B-dutch) for better understanding and generation of Dutch text. It's built on the mistral architecture, employing flash attention for efficient processing within a context window of 8192 tokens. Tweety-7b-dutch is trained on the [cleaned Dutch mC4 dataset](https://huggingface.co/datasets/yhavinga/mc4_nl_cleaned), without instruction finetuning. |
|
|
|
## Model Details |
|
|
|
### Model Description |
|
|
|
Our tweety-7b-dutch model has an Apache 2.0 license, encouraging applications in research, content creation, and language analysis. |
|
|
|
- **Tokenizer:** Dutch, 50k tokens ([yhavinga/gpt-neo-1.3B-dutch](https://huggingface.co/yhavinga/gpt-neo-1.3B-dutch)) |
|
- **Pre-training data:** Scraped Dutch ([yhavinga/mc4_nl_cleaned](https://huggingface.co/datasets/yhavinga/mc4_nl_cleaned)) |
|
- **Context window**: 8196 tokens |
|
- **Training data**: 8.5B tokens |
|
- **Developed by:** KU Leuven and UGent |
|
- **Funded by:** KU Leuven BOF, VSC (Flemish Supercomputer Center), [Vlaams AI-onderzoeksprogramma](https://www.flandersairesearch.be/nl) |
|
- **Model type:** Foundation model |
|
- **License:** Apache 2.0 |
|
|
|
## Uses |
|
|
|
As a base model, tweety-7b-dutch is primed for direct applications across text generation and understanding within the Dutch language. |
|
|
|
## Technical Specifications |
|
|
|
### Compute Infrastructure |
|
Training utilized Nvidia H100 and A100 GPUs. Inference is accessible on lower-end GPUs, basically any GPU capable of running mistral models. |
|
|
|
### Model Weights |
|
|
|
- This model was trained in bfloat16. |
|
- [GGUF weights](https://huggingface.co/BramVanroy/tweety-7b-dutch-v24a-GGUF) are released by Bram Vanroy. |
|
|
|
|
|
## Citation |
|
|
|
If you use this model, please cite our work as: |
|
|
|
``` |
|
@article{tweeties2024, |
|
title = {Trans-Tokenization and Cross-lingual Vocabulary Transfers: Language Adaptation of LLMs for Low-Resource NLP}, |
|
author = {François Remy and Pieter Delobelle and Hayastan Avetisyan and Alfiya Khabibullina and Miryam de Lhoneux and Thomas Demeester}, |
|
url = {https://arxiv.org/abs/2408.04303}, |
|
year = {2024}, |
|
note = {Accepted at COLM 2024} |
|
} |
|
``` |
