LLaMA-Factory VS HALOs

Depends what model you want to train, and how well you want your computer to keep working while you're doing it.

If you're interested in large language models there's a table of vram requirements for fine-tuning at [1] which says you could do the most basic type of fine-tuning on a 7B parameter model with 8GB VRAM.

You'll find that training takes quite a long time, and as a lot of the GPU power is going on training, your computer's responsiveness will suffer - even basic things like scrolling in your web browser or changing tabs uses the GPU, after all.

Spend a bit more and you'll probably have a better time.

[1] https://github.com/hiyouga/LLaMA-Factory?tab=readme-ov-file#...

If you are using no-code solutions, increasing an "idea" in a dataset will make that idea more likely to appear.

If you are fine-tuning your own LLM, there are other ways to get your idea to appear. In the literature this is sometimes called RLHF or preference optimization, and here are a few approaches:

Direct Preference Optimization

This uses Elo-scores to learn pairwise preferences. Elo is used in chess and basketball to rank individuals who compete in pairs.

@argilla_io on X.com has been doing some work in evaluating DPO.

Here is a decent thread on this: https://x.com/argilla_io/status/1745057571696693689?s=20

Identity Preference Optimization

IPO is research from Google DeepMind. It removes the reliance of Elo scores to address overfitting issues in DPO.

Paper: https://x.com/kylemarieb/status/1728281581306233036?s=20

Kahneman-Tversky Optimization

KTO is an approach that uses mono preference data. For example, it asks if a response is "good or not." This is helpful for a lot of real word situations (e.g. "Is the restaurant well liked?").

Here is a brief discussion on it:

https://x.com/ralphbrooks/status/1744840033872330938?s=20

Here is more on KTO:

* Paper: https://github.com/ContextualAI/HALOs/blob/main/assets/repor...

* Code: https://github.com/ContextualAI/HALOs