Confused about the "hf" meaning.

[pokerc]

So, could any one tell me the "hf" mean in Llama-2-70b-hf? What's the difference between Llama-2-70b-hf and Llama-2-70b in hugging face?
"hf" means fp16? or hugging-face-format?

[eyuansu62]

hugging-face-format

[brianjking]

Isn't it human feedback? The 70b-hf is the chat tuned one with RLHF, the 70b base, isn't RLHF'ed and not trained for conversational data.

[MatousAc]

Isn't it human feedback? The 70b-hf is the chat tuned one with RLHF, the 70b base, isn't RLHF'ed and not trained for conversational data.

That would make sense to me if there wasn't already a Llama-2-chat and a Llama-2-chat-hf
Do you mean something different by human feedback than the Chat?

[Kwanky-Millet]

Do you mean something different by human feedback than the Chat?

AFAIK, the chat model is trained for conversations, but they went further and refined the model to be more helpful using RLHF.

Here

[MatousAc]

Do you mean something different by human feedback than the Chat?

AFAIK, the chat model is trained for conversations, but they went further and refined the model to be more helpful using RLHF.

Here

Gotcha. That makes a bit more sense. Thanks.

[lmarcon]

Seems like this issue has been answered, thanks. Closing

[armyofda12mnkeys]

so the highest rated answer here says "hugging-face-format" but another answer says the HF is cause its RLHF (Reinforced Learned Human Feedback). Which is it?

[bjascob]

I can't say for sure but, I can tell that .from_pretrained("meta-llama/Llama-2-7b") fails because it doesn't have a config.json file. However, the meta-llama/Llama-2-7b-hf model does have config.json and you can download it via the transformers .from_pretrained() command.

ie... it looks like -hf identifies a version that is converted to the Huggingface / transformers format.

[vtecftwy]

The model card on HF is clear on this:

meta-llama/Llama-2-7b:
"Llama 2 is a collection of pretrained and fine-tuned generative text models ranging in scale from 7 billion to 70 billion parameters. This is the repository for the 7B pretrained model. Links to other models can be found in the index at the bottom."

meta-llama/Llama-2-7b-hf:
"Llama 2 is a collection of pretrained and fine-tuned generative text models ranging in scale from 7 billion to 70 billion parameters. This is the repository for the 7B pretrained model, converted for the Hugging Face Transformers format. Links to other models can be found in the index at the bottom."

[wjfu99]

The model card on HF is clear on this:

meta-llama/Llama-2-7b: "Llama 2 is a collection of pretrained and fine-tuned generative text models ranging in scale from 7 billion to 70 billion parameters. This is the repository for the 7B pretrained model. Links to other models can be found in the index at the bottom."

meta-llama/Llama-2-7b-hf: "Llama 2 is a collection of pretrained and fine-tuned generative text models ranging in scale from 7 billion to 70 billion parameters. This is the repository for the 7B pretrained model, converted for the Hugging Face Transformers format. Links to other models can be found in the index at the bottom."

Right answer, i guess you are the first one who notices the difference in model cards.