Dictalm2.0

🚀 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. It has 7 billion parameters and is trained to specialize in Hebrew text. This project aims to adapt large - language models to Hebrew, offering enhanced vocabulary and instruction capabilities.

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.

🚀 Quick Start

The DictaLM - 2.0 is a powerful tool for Hebrew text generation. You can start using it right away with the provided code examples.

✨ Features

• Specialized in Hebrew text generation.
• Based on the Mistral - 7B - v0.1 model with extended tokenizer for Hebrew.
• Continued pretraining on a large corpus of naturally occurring text (190B tokens, 50% Hebrew and 50% English).

💻 Usage Examples

Basic Usage

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'}]

Advanced Usage

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:

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']

🔧 Technical Details

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 occurring text, 50% Hebrew and 50% English.

📄 License

This project is licensed under the Apache - 2.0 license.

📚 Documentation

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}, 
}