File size: 3,797 Bytes
32eadf2 da6801b 32eadf2 da6801b 0897a01 da6801b f8ab320 da6801b 72fc812 da6801b f8ab320 da6801b ab18af0 da6801b 72fc812 da6801b f8ab320 da6801b |
1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31 32 33 34 35 36 37 38 39 40 41 42 43 44 45 46 47 48 49 50 51 52 53 54 55 56 57 58 59 60 61 62 63 64 65 66 67 68 69 70 71 72 73 74 75 76 77 78 79 80 81 82 83 84 85 86 87 88 89 90 91 92 93 94 95 96 97 98 99 100 101 102 103 104 105 106 107 108 109 110 |
---
license: apache-2.0
pipeline_tag: text-generation
language:
- en
- he
tags:
- pretrained
inference:
parameters:
temperature: 0.7
---
[<img src="https://i.ibb.co/5Lbwyr1/dicta-logo.jpg" width="300px"/>](https://dicta.org.il)
# 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](https://dicta.org.il/dicta-lm) or the [technical report](https://arxiv.org/abs/2407.07080).
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](https://huggingface.co/collections/dicta-il/dicta-lm-20-collection-661bbda397df671e4a430c27).
## Example Code
```python
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 = """
注讘专: 讛诇讻转讬
注转讬讚: 讗诇讱
注讘专: 砖诪专转讬
注转讬讚: 讗砖诪讜专
注讘专: 砖诪注转讬
注转讬讚: 讗砖诪注
注讘专: 讛讘谞转讬
注转讬讚:
"""
print(model(prompt.strip(), do_sample=False, max_new_tokens=8, stop_sequence='\n'))
# [{'generated_text': '注讘专: 讛诇讻转讬\n注转讬讚: 讗诇讱\n\n注讘专: 砖诪专转讬\n注转讬讚: 讗砖诪讜专\n\n注讘专: 砖诪注转讬\n注转讬讚: 讗砖诪注\n\n注讘专: 讛讘谞转讬\n注转讬讚: 讗讘讬谉\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](https://huggingface.co/dicta-il/dictalm2.0-AWQ) and [DictaLM-2.0-GPTQ](https://huggingface.co/dicta-il/dictalm2.0-GPTQ).
For dynamic quantization on the go, here is sample code which loads the model onto the GPU using the `bitsandbytes` package, requiring :
```python
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 = """
注讘专: 讛诇讻转讬
注转讬讚: 讗诇讱
注讘专: 砖诪专转讬
注转讬讚: 讗砖诪讜专
注讘专: 砖诪注转讬
注转讬讚: 讗砖诪注
注讘专: 讛讘谞转讬
注转讬讚:
"""
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> 注讘专: 讛诇讻转讬\n注转讬讚: 讗诇讱\n\n注讘专: 砖诪专转讬\n注转讬讚: 讗砖诪讜专\n\n注讘专: 砖诪注转讬\n注转讬讚: 讗砖诪注\n\n注讘专: 讛讘谞转讬\n注转讬讚: 讗讘讬谉\n\n']
```
## Model Architecture
DictaLM-2.0 is based on the [Mistral-7B-v0.1](https://huggingface.co/mistralai/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:
```bibtex
@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},
}
``` |