Content
@
0 reply
0 recast
2 reactions
Josh Ellithorpe
@quest
How Unisloth fine tuned Phi-4 for better accuracy. https://unsloth.ai/blog/phi4
1 reply
0 recast
6 reactions
Mikko
@moo
This is interesting > We uploaded 4-bit bitsandbytes pre-quantized models for 4x faster downloading, however, Unsloth's Dynamic 4-bit quant shows we mustn't quantize all layers. This results in largely increased accuracy while only using 10% more VRAM.
0 reply
0 recast
1 reaction
