VBART is the first sequence-to-sequence LLM pre-trained on Turkish corpora from scratch on a large scale. It was pre-trained by VNGRS in February 2023.
The model is capable of conditional text generation tasks such as text summarization, paraphrasing, and title generation when fine-tuned.
It outperforms its multilingual counterparts, albeit being much smaller than other implementations.
It comes in two sizes:
VBART-Large: 387M parameters
VBART-XLarge: 740M parameters
VBART-XLarge is created by adding extra Transformer layers between the layers of VBART-Large. Hence it was able to transfer learned weights from the smaller model while doublings its number of layers.
VBART-XLarge improves the results compared to VBART-Large albeit in small margins.
The base model is pre-trained on vngrs-web-corpus. It is curated by cleaning and filtering Turkish parts of OSCAR-2201 and mC4 datasets. These datasets consist of documents of unstructured web crawl data. More information about the dataset can be found on their respective pages. Data is filtered using a set of heuristics and certain rules, explained in the appendix of our paper.
Software
TensorFlow
Pre-training Setting
Duration: Pre-trained for 30 days.
GPUs: 8 x Nvidia A100-80 GB
Training tokens: 708B
Context Length: 1024 for both encoder and decoder
Training regime: fp16 mixed precision
Training objective: Sentence permutation and span masking (using mask lengths sampled from Poisson distribution λ=3.5, masking 30% of tokens)
Scheduler: Custom scheduler from the original Transformers paper (20,000 warm-up steps)
Dropout: 0.1 (dropped to 0.05 and then to 0 in the last 165K and 205k steps, respectively)
Initial Learning rate: 5e-6
Citation
@article{turker2024vbart,
title={VBART: The Turkish LLM},
author={Turker, Meliksah and Ari, Erdi and Han, Aydin},
journal={arXiv preprint arXiv:2403.01308},
year={2024}
}
