4 ms·
On a positive note, this is a nice look at the future and a direction that existing experiences seem likely to evolve towards. However I did find myself wonder
by WhitneyLand 2y ago
On a positive note, this is a nice look at the future and a direction that existing experiences seem likely to evolve towards.
However I did find myself wondering how crucial really were the model changes?
Imagine trying to implement these features as a wrapper around frontier apis and no 70B bespoke model.
Starting with a user query, we could ask for relevant svg diagrams, fit instruction steps or guides into layout templates, or even filter photos as the demo shows.
How much less would this simple approach leave us with?
- rushingcreek 2y agoAh, 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 https://www.phind.com/blog/phind-2-model-creation.
- Xmd5a 2y agoI'd love to read an in-depth explanation of how you improved your LLM abilities to lay out diagrams.
- rushingcreek 2y agoWe used a system of LLM critics to generate a high-quality dataset for this. More in the blog post linked in the answer above.
- anon373839 2y agoThat's a great blog post, really informative and interesting. Do you recommend any other helpful resources on creating and curating datasets for post-training?