Congrats! 🎉
#2
by
MaziyarPanahi
- opened
Just wanted to say congratulations and thank you for giving credit! 🎉
Next step training on all 40k samples from this dataset?? 👀
I suggest diving into the metrics, looking at what has increased slightly and which metrics have dropped. Then mixing a diverse ORPO datasets to balance the improvements and cover the losses.
This way you will have a nice pipeline to prepare a good mix for each model
@MaziyarPanahi Maybe this could be useful.
Unity Subjects tab.
This is very interesting! Thanks for sharing it! 🙌
@MaziyarPanahMate there any other benchmarks or models that would be useful to you?
BTW we're looking for specialized models for specific knowledge domains, but so far as the data shows, they don't deliver much value (Saul lm for example)