Hacker News Reader

Ah, great question! We tried using off-the-shelf models and found that they are incredibly bad at generating these visual components reliably, let alone with the nuance (such as color, placement, details, etc.) that would be expected in a great answer.

We write more about this in the technical model blog post: https://www.phind.com/blog/phind-2-model-creation.

View on HN

That's a great blog post, really informative and interesting. Do you recommend any other helpful resources on creating and curating datasets for post-training?

I'd love to read an in-depth explanation of how you improved your LLM abilities to lay out diagrams.

We used a system of LLM critics to generate a high-quality dataset for this. More in the blog post linked in the answer above.

Reply on news.ycombinator.com