Hebrew-Gemma-11B-Instruct

Hebrew-Gemma-11B-Instruct

Base Models:

• 07.03.2024: Hebrew-Gemma-11B
• 16.03.2024: Hebrew-Gemma-11B-V2

Instruct Models:

• 07.03.2024: Hebrew-Gemma-11B-Instruct

The Hebrew-Gemma-11B-Instruct Large Language Model (LLM) is a instruct fine-tuned version of the Hebrew-Gemma-11B generative text model using a variety of conversation datasets.

It is continued pretrain of gemma-7b, extended to a larger scale and trained on 3B additional tokens of both English and Hebrew text data.

Instruction format

This format must be strictly respected, otherwise the model will generate sub-optimal outputs.

<bos><start_of_turn>user
Write a hello world program<end_of_turn>
<start_of_turn>model
Here is a simple hellow world program<end_of_turn><eos>

• The conversation starts with <bos>.
• Each turn is preceded by a <start_of_turn> delimiter and then the role of the entity (user or model).
• Turns finish with the <end_of_turn> token.
• Conversation finish with the <eos> token.

You can follow this format to build the prompt manually, if you need to do it without the tokenizer's chat template.

A simple example using the tokenizer's chat template:

from transformers import AutoTokenizer, AutoModelForCausalLM

model_id = "Hebrew-Gemma-11B-Instruct"

tokenizer = AutoTokenizer.from_pretrained(model_id)
model = AutoModelForCausalLM.from_pretrained(model_id, device_map="cuda")

chat = [
    { "role": "user", "content": "כתוב קוד פשוט בפייתון שמדפיס למסך את התאריך של היום" },
]
prompt = tokenizer.apply_chat_template(chat, tokenize=False, add_generation_prompt=True)

Terms of Use

As an extention of Gemma-7B, this model is subject to the original license and terms of use by Google.

Benchmark Results

• Coming Soon!

Notice

Hebrew-Gemma-11B is a pretrained base model and therefore does not have any moderation mechanisms.

Authors

• Trained by Yam Peleg.
• In collaboration with Jonathan Rouach and Arjeo, inc.