Llama-3.1-Minitron-4B-Chat

This is an instruction-tuned version of nvidia/Llama-3.1-Minitron-4B-Width-Base that has underwent supervised fine-tuning with 64k instruction-response pairs from the teknium/OpenHermes-2.5 dataset on a single A100 40GB GPU.

How to use

Chat Format

Given the nature of the training data, the Llama-3.1-Minitron-4B chat model is best suited for prompts using the chat format as follows. You can provide the prompt as a question with a generic template as follows:

<|im_start|>system
You are a helpful assistant.<|im_end|>
<|im_start|>user
Question?<|im_end|>
<|im_start|>assistant

For example:

<|im_start|>system
You are a helpful assistant.<|im_end|>
<|im_start|>user
How to explain Internet for a medieval knight?<|im_end|>
<|im_start|>assistant

where the model generates the text after <|im_start|>assistant .

Sample inference code

Support for this model will be added in the upcoming transformers release. In the meantime, please install the library from source:

pip install git+https://github.com/huggingface/transformers

This code snippets show how to get quickly started with running the model on a GPU:

import torch 
from transformers import AutoModelForCausalLM, AutoTokenizer, pipeline 

torch.random.manual_seed(0) 

model_id = "rasyosef/Llama-3.1-Minitron-4B-Chat"
model = AutoModelForCausalLM.from_pretrained( 
    model_id,  
    device_map="auto",  
    torch_dtype=torch.bfloat16 
) 

tokenizer = AutoTokenizer.from_pretrained(model_id) 

messages = [ 
    {"role": "system", "content": "You are a helpful AI assistant."}, 
    {"role": "user", "content": "Can you provide ways to eat combinations of bananas and dragonfruits?"}, 
    {"role": "assistant", "content": "Sure! Here are some ways to eat bananas and dragonfruits together: 1. Banana and dragonfruit smoothie: Blend bananas and dragonfruits together with some milk and honey. 2. Banana and dragonfruit salad: Mix sliced bananas and dragonfruits together with some lemon juice and honey."}, 
    {"role": "user", "content": "What about solving an 2x + 3 = 7 equation?"}, 
] 

pipe = pipeline( 
    "text-generation", 
    model=model, 
    tokenizer=tokenizer, 
) 

generation_args = { 
    "max_new_tokens": 256, 
    "return_full_text": False, 
    "temperature": 0.0, 
    "do_sample": False, 
} 

output = pipe(messages, **generation_args) 
print(output[0]['generated_text'])  

Note: If you want to use flash attention, call AutoModelForCausalLM.from_pretrained() with attn_implementation="flash_attention_2"