Llama-3.2-8b used in the course and lab -- does it exist within meta-llama repo?

Course Q&A Fine-tuning & RL for LLMs: Intro to Post-training
,
1

Hello team,

I’m looking into utils.py to understand the code. I noticed that llama-3.2-8b is referenced, but I wasn’t able to find a matching model in the meta-llama repo.

The model was called here within the lab notebook:

Load error analysis model from utils

ERROR_CATEGORIES = [
“calculation_error”, # Math computation mistakes
“reasoning_error”, # Logical reasoning issues
“incomplete_solution”, # Partial or unfinished answers
“format_error”, # Wrong answer format
“other” # Other errors
]

tokenizer, model, hf_model_available = load_error_analysis_model()

I then looked into load_error_analysis_model() in utils.py:

def load_error_analysis_model():
“”"
Load the error analysis model from local disk.
Returns:
tuple: (tokenizer, model, hf_model_available) where hf_model_available is bool
“”"
print(“Loading local model for error analysis…”)
try:

local_model_path = ‘/app/models/llama-3.2-8b’

    # Verify the model directory exists
    if not os.path.exists(local_model_path):
        raise FileNotFoundError(f"Model directory not found at: {local_model_path}")
    
    if not os.path.isdir(local_model_path):
        raise NotADirectoryError(f"Path exists but is not a directory: {local_model_path}")
    
    print(f"📦 Loading model from: {local_model_path}")
    
    # Load tokenizer from local path only
    print("Loading tokenizer...")
    tokenizer = AutoTokenizer.from_pretrained(
        local_model_path,
        local_files_only=True,
        trust_remote_code=True
    )

It appears llama-3.2-8b is the model being loaded. However, after checking Meta-llama repo on huggingface ( meta-llama (Meta Llama) ), I wasn’t able to locate a matching model – I may be looking in the wrong place. There is a mradermacher/Llama-3.2-8B-Instruct-GGUF · Hugging Face , but it looks different from the model used in this course and lab.

Questions for the instructor / community

  1. Could llama-3.2-8b be a typo, or am I missing something about where this model is hosted? Or, should it be llama-3.1-8b or llama-3-8b?

  2. What checkpoint was the course originally validated against? (Could you share the exact HuggingFace model ID or expected directory contents — e.g., does it contain tokenizer.json, tokenizer.model, or both?)

Thanks in advances!

C.L.

2

I don’t know where the model is coming from and this is not likely to be an answer to your question, but we are loading the model from a lab directory. I have verified that the requested model exists in the path.

3

Hello my friend, thank you for replying! Yes, the model is working ok. Maybe my question is more about where the model is from as I can’t find it from hugging face. Could you help share a little hints? Thank you!

4

Nice observation! You’re correct — there is no official “Llama 3.2 8B” model from Meta. The Llama 3.2 family only includes 1B and 3B text models (plus 11B and 90B vision models).

The 8B parameter size exists in the official Meta repos:

  • Llama 3 → meta-llama/Meta-Llama-3-8B

  • Llama 3.1 → meta-llama/Llama-3.1-8B

There’s also the community model you found:

  • mradermacher → mradermacher/Llama-3.2-8B-Instruct-GGUF

Any of these three could be what’s pre-loaded at /app/models/llama-3.2-8b in the lab environment. The code uses local_files_only=True, so it loads whatever weights are in that directory regardless of the folder name — which is why it works fine even though the directory name doesn’t match any official HuggingFace model ID.

Only the course developers can confirm which exact checkpoint was used. Tagging @Community-Team to clarify the exact model ID and whether the directory name is intentional or a typo.

5

Better to post a message in the Mentor area for this course.

6

Better to post a reply on the Mentor topic for this course. In theory, the course staff monitor that topic.

7

Hi, this lab was developed by our partners from AMD, and we downloaded and stored the model, so that the lab doesnt have to communicate with the internet (and you dont need to download the model every time you run the lab). But I dont know the details about where it was downloaded from. I already reached out to them to ask about it. I will let you know when I get a reply