DictaBERT: A State-of-the-Art BERT Suite for Modern Hebrew
Paper
โข 2308.16687 โข Published
Google's mT5 multilingual Seq2Seq model, finetuned on HeQ for the Hebrew Question-Answering task.
This is the model that was reported in the DictaBERT release here.
Sample usage:
import torch
from transformers import AutoModelForSeq2SeqLM, AutoTokenizer
tokenizer = AutoTokenizer.from_pretrained('dicta-il/mt5-xl-heq')
model = AutoModelForSeq2SeqLM.from_pretrained('dicta-il/mt5-xl-heq')
model.eval()
question='ืืืฆื ืืืืื ืืืืืข ืฉื ืืชื ืืืฉืื ืืืืฆืขืืช ืืขืืืืืช?'
context='ืื ืืืช ืคืจืืคืืืื ืฉื ืืฉืชืืฉืื ื ืืฉืืช ืขื ืืื ืจืืื ืืืืื ืคืืื ืฆืืืื ืขื ืืคืจืืืืช. ืืกืืื ืื ืืืืืื ืืืง ืืืืืื ืืช ืืืืฆืขืืช ืืงืืงื ืืช ืืืืืข ืฉื ืืชื ืืืฉืื ืืืืฆืขืืช ืขืืืืืช ืืืช ืืืคื ืืฉืืืืฉ ืืขืืืืืช. ืืจืฆืืช ืืืจืืช, ืืืฉื, ืงืืขื ืืืงืื ื ืืงืฉืื ืืื ืื ืืืข ืืืฆืืจืช ืขืืืืืช ืืืฉืืช. ืืืงืื ืืื, ืืฉืจ ื ืงืืขื ืืฉื ืช 2000, ื ืงืืขื ืืืืจ ืฉื ืืฉืฃ ืื ืืืฉืจื ืืืืฉืื ืืืืื ืืืช ืฉื ืืืืฉื ืืืืจืืงืื ื ืื ืืฉืืืืฉ ืืกืืื (ONDCP) ืืืืช ืืืื ืืฉืชืืฉ ืืขืืืืืช ืืื ืืขืงืื ืืืจื ืืฉืชืืฉืื ืฉืฆืคื ืืคืจืกืืืืช ื ืื ืืฉืืืืฉ ืืกืืื ืืืืจื ืืืืืง ืืื ืืฉืชืืฉืื ืืื ื ืื ืกื ืืืชืจืื ืืชืืืืื ืืฉืืืืฉ ืืกืืื. ืื ืืื ืืจืื ื, ืคืขืื ืืืืื ืืคืจืืืืช ืืืฉืชืืฉืื ืืืื ืืจื ื, ืืฉืฃ ืื ื-CIA ืฉืื ืขืืืืืช ืงืืืขืืช ืืืืฉืื ืืืจืืื ืืืฉื ืขืฉืจ ืฉื ืื. ื-25 ืืืฆืืืจ 2005 ืืืื ืืจืื ื ืื ืืกืืื ืืช ืืืืืืื ืืืืื (ื-NSA) ืืฉืืืจื ืฉืชื ืขืืืืืช ืงืืืขืืช ืืืืฉืื ืืืงืจืื ืืืื ืฉืืจืื ืชืืื ื. ืืืืจ ืฉืื ืืฉื ืคืืจืกื, ืื ืืืืื ืืื ืืช ืืฉืืืืฉ ืืื.'
with torch.inference_mode():
prompt = 'question: %s context: %s ' % (question, context)
kwargs = dict(
inputs=tokenizer(prompt, return_tensors='pt').input_ids.to(model.device),
do_sample=True,
top_k=50,
top_p=0.95,
temperature=0.75,
max_length=100,
min_new_tokens=2
)
print(tokenizer.batch_decode(model.generate(**kwargs), skip_special_tokens=True))
Output:
["ืืืืฆืขืืช ืืงืืงื"]
If you use mt5-xl-heq in your research, please cite DictaBERT: A State-of-the-Art BERT Suite for Modern Hebrew
BibTeX:
@misc{shmidman2023dictabert,
title={DictaBERT: A State-of-the-Art BERT Suite for Modern Hebrew},
author={Shaltiel Shmidman and Avi Shmidman and Moshe Koppel},
year={2023},
eprint={2308.16687},
archivePrefix={arXiv},
primaryClass={cs.CL}
}
This work is licensed under a Creative Commons Attribution 4.0 International License.