README.md

metadata

license: apache-2.0

GreenBit LLaMA

This is GreenBitAI's pretrained 2-bit TinyLLaMA model with extreme compression yet still strong performance.

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

Model Description

• Developed by: GreenBitAI
• Model type: Causal (Llama 2)
• Language(s) (NLP): English
• License: Apache 2.0, Llama 2 license agreement

Zero-Shot Evaluation

Task	Metric	TinyLLaMA 1.1B q2g32	TinyLLaMA 1.1B q2g8	LLaMA 3B q2g32	LLaMA 3B q2g16	LLaMA 3B q2g8	LLaMA-1 7B q2g32	LLaMA-2 7B q2g32	LLaMA-2 7B q2g8	LLaMA 1.1B FP16	LLaMA 3B FP16	LLaMA-1 7B FP16
Openbookqa	acc	0.152	0.192	0.196	0.238	0.242	0.224	0.246	0.296	0.208	0.27	0.29
	ac_norm	0.328	0.338	0.332	0.358	0.362	0.388	0.376	0.4	0.368	0.4	0.41
arc_challenge	acc	0.3268	0.2278	0.279	0.2978	0.3148	0.3422	0.3268	0.3618	0.243	0.34	0.39
	ac_norm	0.3387	0.273	0.2944	0.3319	0.3345	0.3387	0.3387	0.372	0.288	0.37	0.41
hellawswag	acc	0.34	0.3769	0.4238	0.444	0.462	0.4996	0.4961	0.5379	0.403	0.49	0.68
	ac_norm	0.4097	0.4711	0.5685	0.5988	0.6242	0.6447	0.6464	0.7014	0.503	0.67	0.73
piqa	acc	0.6518	0.6931	0.7024	0.716	0.7291	0.7476	0.7503	0.7715	0.71	0.75	0.78
	ac_norm	0.6393	0.6812	0.7116	0.7247	0.7312	0.7443	0.7421	0.7568	0.688	0.76	0.78
arc_easy	acc	0.4411	0.5109	0.5997	0.646	0.6528	0.6061	0.6174	0.6254	0.533	0.69	0.68
	ac_norm	0.3716	0.412	0.5417	0.58	0.5972	0.4566	0.4781	0.4958	0.43	0.65	0.52
Winogrande	acc	0.532	0.5249	0.5683	0.5888	0.6054	0.6283	0.6298	0.6582	0.558	0.62	0.68
boolq	acc	0.592	0.6174	0.6281	0.6636	0.6327	0.6425	0.7061	0.7242	0.583	0.68	0.75
truthfulqa_mc	mc1	0.2338	0.2277	0.2509	0.2118	0.2252	0.224	0.2313	0.2399	0.228	0.22	0.21
	mc2	0.4211	0.406	0.3962	0.3501	0.3625	0.3702	0.3854	0.3795	0.401	0.35	0.34
anli_r1	acc	0.363	0.336	0.337	0.334	0.344	0.331	0.333	0.363	0.354	0.33	0.35
anli_r2	acc	0.331	0.346	0.335	0.332	0.331	0.326	0.349	0.347	0.341	0.32	0.34
anli_r3	acc	0.3758	0.3633	0.3358	0.3383	0.3425	0.3417	0.36	0.3733	0.358	0.35	0.37
wic	acc	0.5	0.5	0.4984	0.5094	0.4969	0.4984	0.4953	0.489	0.5	0.48	0.5
rte	acc	0.4874	0.4874	0.5596	0.5993	0.5632	0.639	0.6065	0.6426	0.516	0.58	0.56
record	f1	0.7608	0.8023	0.8502	0.8625	0.8687	0.8859	0.8872	0.9037	0.82	0.88	0.91
	em	0.753	0.7934	0.8427	0.8545	0.8612	0.8781	0.8801	0.8959	0.818	0.89	0.91
Average		0.438	0.4498	0.4881	0.5037	0.5087	0.5122	0.5181	0.5391	0.469	0.528	0.5519
model size	GiB	0.5	0.6	1.2	1.3	1.5	2.2	2.2	2.9	4.4	6.8	12.5