GreenBitAI/Phi-3-mini-4k-instruct-layer-mix-bpw-2.5

GreenBit LLMs

This is GreenBitAI's pretrained low-bit LLMs with extreme compression yet still strong performance.

Please refer to our Github page for the code to run the model and more information.

zero-shot evaluation

Repository (Phi Family)	Avg Acc.	OpenBQ	ARC-E	Winogr.	HellaS.	ARC-C	PIQA	BoolQ	RACE	ANLI-R1	ANLI-R2	ANLI-R3	WiC
Phi-3-mini-128k-instruct-layer-mix-bpw-2.2	0.510	0.270	0.706	0.648	0.479	0.411	0.736	0.783	0.381	0.393	0.38	0.399	0.536
Phi-3-mini-128k-instruct-layer-mix-bpw-2.5	0.514	0.290	0.719	0.656	0.488	0.401	0.750	0.778	0.401	0.392	0.410	0.407	0.493
Phi-3-mini-128k-instruct-layer-mix-bpw-3.0	0.548	0.318	0.761	0.663	0.519	0.453	0.777	0.798	0.393	0.473	0.404	0.442	0.579
Phi-3-mini-128k-instruct-layer-mix-bpw-4.0	0.582	0.346	0.779	0.708	0.582	0.495	0.787	0.840	0.412	0.529	0.459	0.448	0.606
Phi-3-mini-128k-instruct	0.586	0.342	0.785	0.731	0.596	0.512	0.782	0.851	0.401	0.547	0.464	0.432	0.594

5-shot evaluation

Repository (Phi Family)	Avg Acc.	OpenBQ	ARC-E	Winogr.	HellaS.	ARC-C	PIQA	BoolQ	RACE	ANLI-R1	ANLI-R2	ANLI-R3	WiC
Phi-3-mini-128k-instruct-layer-mix-bpw-2.2	0.534	0.302	0.738	0.659	0.487/0.636	0.438	0.744	0.793	0.408	0.421	0.404	0.439	0.583
Phi-3-mini-128k-instruct-layer-mix-bpw-2.5	0.543	0.310	0.771	0.671	0.501/0.657	0.441	0.763	0.799	0.405	0.453	0.427	0.443	0.534
Phi-3-mini-128k-instruct-layer-mix-bpw-3.0	0.563	0.346	0.796	0.687	0.528/0.694	0.500	0.782	0.809	0.410	0.473	0.394	0.474	0.565
Phi-3-mini-128k-instruct-layer-mix-bpw-4.0	0.602	0.374	0.817	0.725	0.598/0.768	0.542	0.766	0.864	0.428	0.523	0.456	0.497	0.658
Phi-3-mini-128k-instruct	0.608	0.408	0.825	0.725	0.608/0.781	0.534	0.768	0.866		0.538	0.483	0.515	0.627