Using HuggingFace Datasets

To use a HuggingFace dataset in Unity, you need to configure the HuggingFace environment variable. If you need a HuggingFace model, check its availability at Unity AI Datasets. If the model is unavailable, please contact the Unity team.

Example

To load the Llama-3.2-3B model, which is stored at /datasets/ai/llama3, set the HuggingFace environment variable accordingly
There are 3 supported options for huggingface library to recognize the model

Option 1: Set environment variable directly on your terminal before running the script
Note that notebook cell doesn’t support Option 1. Try this on your terminal

export HF_HOME=/datasets/ai/llama3

# export HF_HOME=/datasets/ai/llama3 # <- Set environment path in the terminal
from transformers import AutoTokenizer, AutoModelForCausalLM

tokenizer = AutoTokenizer.from_pretrained("meta-llama/Llama-3.2-3B")
model = AutoModelForCausalLM.from_pretrained("meta-llama/Llama-3.2-3B")

Option 2: Use Absolute Model Path directly
Huggingface stores model parameters/tokenizer/config files in a specific directory structure as follows

from transformers import AutoTokenizer, AutoModelForCausalLM

model_path = '/datasets/ai/llama3/hub/models--meta-llama--Llama-3.2-3B/snapshots/13afe5124825b4f3751f836b40dafda64c1ed062'
tokenizer = AutoTokenizer.from_pretrained(model_path)
model = AutoModelForCausalLM.from_pretrained(model_path)

Option 3: Set HF_HOME in the script
os.environ["HF_HOME"] = "/datasets/ai/llama3" must be placed above importing transformer library
Note that notebook cell doesn’t support Option 3. Try this on your terminal

import os

os.environ["HF_HOME"] = "/datasets/ai/llama3" # <- This line is placed above "from transformers.."
from transformers import AutoTokenizer, AutoModelForCausalLM

tokenizer = AutoTokenizer.from_pretrained("meta-llama/Llama-3.2-3B")
model = AutoModelForCausalLM.from_pretrained("meta-llama/Llama-3.2-3B")

See HuggingFace’s cache management documentation for more information.

For usage details, click the “Use this model” button on the HuggingFace model page.