vngrs-ai/VBART-XLarge-QAQG

VBART Model Card

Model Description

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.

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.

This repository contains fine-tuned TensorFlow and Safetensors weights of VBART for question-answering and generation tasks described in the paper.

• Developed by: VNGRS-AI
• Model type: Transformer encoder-decoder based on mBART architecture
• Language(s) (NLP): Turkish
• License: CC BY-NC-SA 4.0
• Finetuned from: VBART-XLarge
• Paper: arXiv

How to Get Started with the Model

from transformers import AutoTokenizer, AutoModelForSeq2SeqLM

tokenizer = AutoTokenizer.from_pretrained("vngrs-ai/VBART-XLarge-QAQG",
                            model_input_names=['input_ids', 'attention_mask'])
# Uncomment the device_map kwarg and delete the closing bracket to use model for inference on GPU
model = AutoModelForSeq2SeqLM.from_pretrained("vngrs-ai/VBART-XLarge-QAQG")#, device_map="auto")

context="..."
question="..."
highlighted_context="..."

# Prompt for question generation
qg_prompt = f'Soru yarat: cevap: {context}'
# Prompt for question answering
qa_prompt = f'Soru cevapla: {question} kaynak: {context}'
# Prompt for answer extraction
ae_prompt = f'yanıtları çıkar: {highlighted_context}'


token_input = tokenizer(ae_prompt, return_tensors="pt")#.to('cuda')
outputs = model.generate(**token_input)
print(tokenizer.decode(outputs[0]))

Training Details

Fine-tuning prompt

This model is fine-tuned on three tasks:

• question answering: Answer a question in a given context. Prompted with Soru cevapla: <question> kaynak: <context>
• question generation: Generate a question from a given context. Will accept a highlight token (<hl>, without spaces) to specify the answer to the question generated. Prompted with Soru yarat: <context>
• answer extraction: Will extract possible answers from a highlighted range (using the same highlight token). Prompted with yanıtları çıkar: <context with highlighted parts>

Training Data

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.

The fine-tuning dataset is TQuAD, which has two versions. We have concatenated them and dropped duplicate samples. More information about this process can be found in Appendix B of our paper.

Limitations

This model is fine-tuned for question-answering and question-generation tasks with specific prompts. 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.

Training Procedure

Pre-trained for 8 days and for a total of 84B tokens. Finally, finetuned for 55 epochs.

Hardware

• GPUs: 8 x Nvidia A100-80 GB

Software

• TensorFlow

Hyperparameters

Pretraining

• 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)
• Optimizer : Adam optimizer (β1 = 0.9, β2 = 0.98, Ɛ = 1e-6)
• Scheduler: Custom scheduler from the original Transformers paper (20,000 warm-up steps)
• Weight Initialization: Model Enlargement from VBART-Large. See the related section in the paper for the details.
• Dropout: 0.1 (dropped to 0.05 and then to 0 in the last 80K and 80k steps, respectively)
• Initial Learning rate: 5e-6
• Training tokens: 84B

Fine-tuning

• Training regime: fp16 mixed precision
• Optimizer : Adam optimizer (β1 = 0.9, β2 = 0.98, Ɛ = 1e-6)
• Scheduler: Linear decay scheduler
• Dropout: 0.1
• Learning rate: 5e-6
• Fine-tune epochs: 55

Metrics

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}
}