🦙 Fine-tune Llama 3 with ORPO

Releasing a Colab notebook to fine-tune the new Llama 3 8B with ORPO.

See you tomorrow for an article on this topic.

💻 Colab: colab.research.google.com/drive/1eHNWg9g…

One ORPO fine-tune, two frankenmerges, and three merges so far. 🫡

Llama 3's vocabulary size is much bigger (32000 => 128256). Also note 11008 => 14336.

Models dropped on Hugging Face!

huggingface.co/meta-llama/Met…
huggingface.co/meta-llama/Met…

Literally how these models are made!

How to make your own: huggingface.co/blog/mlabonne/…

Cool demo of a semantic autocompletion with a React component from Mihai Chirculescu

His repo deserves a lot more stars: github.com/Mihaiii/semant…

⚔️ Closed-source vs. Open-weight LLMs

The gap between the best-performing closed-source and open-source models is also narrowing in terms of Arena ELO (chat.lmsys.org/?leaderboard).

We now face a 6 to 10-month lag, rather than years as was the case when GPT-4 was released.

Let's

I was curious to see how the latest batch of open-weight models changes the dynamics with closed-source LLMs.

I need to gather more data, but it looks like the gap is slowly closing...