Adapting LLMs to Hebrew: Unveiling DictaLM 2.0 with Enhanced Vocabulary and Instruction Capabilities
The DictaLM-2.0 Large Language Model (LLM) is a pretrained generative text model with 7 billion parameters trained to specialize in Hebrew text.
For full details of this model please read our release blog post or the technical report.
This is the full-precision base model.
You can view and access the full collection of base/instruct unquantized/quantized versions of DictaLM-2.0
here.
Example Code
from transformers import pipeline
import torch
# This loads the model onto the GPU in bfloat16 precision
model = pipeline('text-generation', 'dicta-il/dictalm2.0', torch_dtype=torch.bfloat16, device_map='cuda')
# Sample few shot examples
prompt = """
עבר: הלכתי
עתיד: אלך
עבר: שמרתי
עתיד: אשמור
עבר: שמעתי
עתיד: אשמע
עבר: הבנתי
עתיד:
"""
print(model(prompt.strip(), do_sample=False, max_new_tokens=8, stop_sequence='\n'))
# [{'generated_text': 'עבר: הלכתי\nעתיד: אלך\n\nעבר: שמרתי\nעתיד: אשמור\n\nעבר: שמעתי\nעתיד: אשמע\n\nעבר: הבנתי\nעתיד: אבין\n\n'}]
Example Code - 4-Bit
There are already pre-quantized 4-bit models using the GPTQ
and AWQ
methods available for use: DictaLM-2.0-AWQ and DictaLM-2.0-GPTQ.
For dynamic quantization on the go, here is sample code which loads the model onto the GPU using the bitsandbytes
package, requiring :
from transformers import AutoModelForCausalLM, AutoTokenizer
import torch
model = AutoModelForCausalLM.from_pretrained('dicta-il/dictalm2.0', torch_dtype=torch.bfloat16, device_map='cuda', load_in_4bit=True)
tokenizer = AutoTokenizer.from_pretrained('dicta-il/dictalm2.0')
prompt = """
עבר: הלכתי
עתיד: אלך
עבר: שמרתי
עתיד: אשמור
עבר: שמעתי
עתיד: אשמע
עבר: הבנתי
עתיד:
"""
encoded = tokenizer(prompt.strip(), return_tensors='pt').to(model.device)
print(tokenizer.batch_decode(model.generate(**encoded, do_sample=False, max_new_tokens=4)))
# ['<s> עבר: הלכתי\nעתיד: אלך\n\nעבר: שמרתי\nעתיד: אשמור\n\nעבר: שמעתי\nעתיד: אשמע\n\nעבר: הבנתי\nעתיד: אבין\n\n']
Model Architecture
DictaLM-2.0 is based on the Mistral-7B-v0.1 model with the following changes:
- An extended tokenizer with 1,000 injected tokens specifically for Hebrew, increasing the compression rate from 5.78 tokens/word to 2.76 tokens/word.
- Continued pretraining on over 190B tokens of naturally occuring text, 50% Hebrew and 50% English.
Notice
DictaLM 2.0 is a pretrained base model and therefore does not have any moderation mechanisms.
Citation
If you use this model, please cite:
@misc{shmidman2024adaptingllmshebrewunveiling,
title={Adapting LLMs to Hebrew: Unveiling DictaLM 2.0 with Enhanced Vocabulary and Instruction Capabilities},
author={Shaltiel Shmidman and Avi Shmidman and Amir DN Cohen and Moshe Koppel},
year={2024},
eprint={2407.07080},
archivePrefix={arXiv},
primaryClass={cs.CL},
url={https://arxiv.org/abs/2407.07080},
}
- Downloads last month
- 8,651