Hebrew-Mixtral-8x22B

Hebrew-Mixtral-8x22B is an open-source Large Language Model (LLM) pretrained in hebrew and english pretrained with 141 billion parameters, based on Mixtral-8x22B from Mistral.

It is continuesly pretrained from Mixtral-8x22B on tokens in both English and Hebrew.

The resulting model is a powerful general-purpose language model suitable for a wide range of natural language processing tasks, with a focus on Hebrew language understanding and generation.

Usage

Below are some code snippets on how to get quickly started with running the model.

First make sure to pip install -U transformers, then copy the snippet from the section that is relevant for your usecase.

Running on CPU

from transformers import AutoTokenizer, AutoModelForCausalLM

tokenizer = AutoTokenizer.from_pretrained("yam-peleg/Hebrew-Mixtral-8x22B")
model = AutoModelForCausalLM.from_pretrained("yam-peleg/Hebrew-Mixtral-8x22B")

input_text = "שלום! מה שלומך היום?"
input_ids = tokenizer(input_text, return_tensors="pt")

outputs = model.generate(**input_ids)
print(tokenizer.decode(outputs[0]))

Running on GPU

from transformers import AutoTokenizer, AutoModelForCausalLM

tokenizer = AutoTokenizer.from_pretrained("yam-peleg/Hebrew-Mixtral-8x22B")
model = AutoModelForCausalLM.from_pretrained("yam-peleg/Hebrew-Mixtral-8x22B", device_map="auto")

input_text = "שלום! מה שלומך היום?"
input_ids = tokenizer(input_text, return_tensors="pt").to("cuda")

outputs = model.generate(**input_ids)
print(tokenizer.decode(outputs[0]))

Running with 4-Bit precision

from transformers import AutoTokenizer, AutoModelForCausalLM, BitsAndBytesConfig

tokenizer = AutoTokenizer.from_pretrained("yam-peleg/Hebrew-Mixtral-8x22B")
model = AutoModelForCausalLM.from_pretrained("yam-peleg/Hebrew-Mixtral-8x22B", quantization_config = BitsAndBytesConfig(load_in_4bit=True))

input_text = "שלום! מה שלומך היום?"
input_ids = tokenizer(input_text, return_tensors="pt").to("cuda")

outputs = model.generate(**input_ids)
print(tokenizer.decode(outputs[0])

Notice

Hebrew-Mixtral-8x22B is a pretrained base model and therefore does not have any moderation mechanisms.