Hyperion-3.0 has significantly improved performance over its predecessors.
"I found that having more code datasets than general purpose datasets ironically decreases performance in both coding and general tasks."
Data sources:
OpenOrca/SlimOrca
cognitivecomputations/dolphin (300k examples)
microsoft/orca-math-word-problems-200k (60k examples)
glaiveai/glaive-code-assistant
Vezora/Tested-22k-Python-Alpaca
Unnatural Instructions
BI55/MedText
LDJnr/Pure-Dove
Various domain-specific datasets by… See the full description on the dataset page:
https://huggingface.co/datasets/Locutusque/hyperion-v3.0.
