vngrs/VBART-XLarge-Paraphrasing
119
1---2language:3- tr4inference:5 parameters:6 max_new_tokens: 1287arXiv: 2403.013088library_name: transformers9pipeline_tag: text2text-generation10license: cc-by-nc-sa-4.011datasets:12- vngrs-ai/vngrs-web-corpus13---14# VBART Model Card15 16## Model Description 17 18VBART 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. 19The model is capable of conditional text generation tasks such as text summarization, paraphrasing, and title generation when fine-tuned.20It outperforms its multilingual counterparts, albeit being much smaller than other implementations.21 22VBART-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.23VBART-XLarge improves the results compared to VBART-Large albeit in small margins.24 25This repository contains fine-tuned TensorFlow and Safetensors weights of VBART for sentence-level text paraphrasing task.26 27- **Developed by:** [VNGRS-AI](https://vngrs.com/ai/)28- **Model type:** Transformer encoder-decoder based on mBART architecture29- **Language(s) (NLP):** Turkish30- **License:** CC BY-NC-SA 4.031- **Finetuned from:** VBART-XLarge32- **Paper:** [arXiv](https://arxiv.org/abs/2403.01308)33## How to Get Started with the Model 34```python35from transformers import AutoTokenizer, AutoModelForSeq2SeqLM36 37tokenizer = AutoTokenizer.from_pretrained("vngrs-ai/VBART-XLarge-Paraphrasing",38 model_input_names=['input_ids', 'attention_mask'])39# Uncomment the device_map kwarg and delete the closing bracket to use model for inference on GPU40model = AutoModelForSeq2SeqLM.from_pretrained("vngrs-ai/VBART-XLarge-Paraphrasing")#, device_map="auto")41 42input_text="..."43 44token_input = tokenizer(input_text, return_tensors="pt")#.to('cuda')45outputs = model.generate(**token_input)46print(tokenizer.decode(outputs[0]))47```48 49## Training Details 50### Training Data 51The base model is pre-trained on [vngrs-web-corpus](https://huggingface.co/datasets/vngrs-ai/vngrs-web-corpus). It is curated by cleaning and filtering Turkish parts of [OSCAR-2201](https://huggingface.co/datasets/oscar-corpus/OSCAR-2201) and [mC4](https://huggingface.co/datasets/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](https://arxiv.org/abs/2403.01308).52 53The fine-tuning dataset is a mixture of [OpenSubtitles](https://huggingface.co/datasets/open_subtitles), [TED Talks (2013)](https://wit3.fbk.eu/home) and [Tatoeba](https://tatoeba.org/en/) datasets.54 55### Limitations56This model is fine-tuned for paraphrasing tasks and finetuned in sentence level only. It is not intended to be used in any other case and can not be fine-tuned to any other task with full performance of the base model. It is also not guaranteed that this model will work without specified prompts.57 58### Training Procedure 59Pre-trained for 8 days and for a total of 84B tokens. Finetuned for 25 epoch.60#### Hardware61- **GPUs**: 8 x Nvidia A100-80 GB62#### Software63- TensorFlow64#### Hyperparameters 65##### Pretraining66- **Training regime:** fp16 mixed precision67- **Training objective**: Sentence permutation and span masking (using mask lengths sampled from Poisson distribution λ=3.5, masking 30% of tokens)68- **Optimizer** : Adam optimizer (β1 = 0.9, β2 = 0.98, Ɛ = 1e-6)69- **Scheduler**: Custom scheduler from the original Transformers paper (20,000 warm-up steps)70- **Weight Initialization**: Model Enlargement from VBART-Large. See the related section in the [paper](https://arxiv.org/abs/2403.01308) for the details. 71- **Dropout**: 0.1 (dropped to 0.05 and then to 0 in the last 80K and 80k steps, respectively)72- **Initial Learning rate**: 5e-673- **Training tokens**: 84B74 75##### Fine-tuning76- **Training regime:** fp16 mixed precision77- **Optimizer** : Adam optimizer (β1 = 0.9, β2 = 0.98, Ɛ = 1e-6)78- **Scheduler**: Linear decay scheduler79- **Dropout**: 0.1 80- **Learning rate**: 5e-681- **Fine-tune epochs**: 5582 83#### Metrics8485 86## Citation 87```88@article{turker2024vbart,89 title={VBART: The Turkish LLM},90 author={Turker, Meliksah and Ari, Erdi and Han, Aydin},91 journal={arXiv preprint arXiv:2403.01308},92 year={2024}93}94```