microsoft/Phi-3.5-mini-instruct · Evaluation of Phi-3.5 on long-context BABILong bench

Hi! I want to share some promising results of the new Phi-3.5 models on the BABILong benchmark, which evaluates models' long-context reasoning over multiple distributed facts.

The new Phi-3.5 models show much better results on the BABILong benchmark compared to previous versions and larger competitor models. The Phi-3.5-mini-instruct model notably improves on context lengths up to 32K, outperforming the Phi-3-mini-128k-instruct. The Phi-3.5-MoE-instruct achieves similar performance to the Phi-3-medium-128k-instruct but with much fewer active parameters (6.6B vs 14B), and is very close to LLama-3.1-8B-Instruct.

Here are the results for models that support 128k contexts. Full BABILong leaderboard is here.

Models (128k)	params	0K	1K	2K	4K	8K	16K	32K	64K	128K	avg
Phi-3-mini-128k-instruct	3.8B	64	57	55	51	50	46	42	37	7	45,4
ai21labs/Jamba-v0.1	12B (52B)	65	53	50	48	46	45	41	40	34	46,9
Phi-3.5-mini-instruct	3.8B	70	70	62	59	58	53	43	38	10	51,4
c4ai-command-r-v01	35B	64	64	63	61	59	52	51	46	38	55,4
Phi-3.5-MoE-instruct	6.6B (16x3.8B)	77	71	65	61	59	52	50	43	37	57,2
Phi-3-medium-128k-instruct	14B	72	70	67	62	60	57	53	45	30	57,5
Meta-Llama-3.1-8B-Instruct	8B	67	68	66	66	62	60	56	49	39	59,2
gpt-4o-mini-2024-07-18	-	74	72	71	65	62	60	54	45	43	60,7
GPT-4 (gpt-4-0125-preview)	-	87	81	77	74	71	64	53	43	36	65,1
Meta-Llama-3.1-70B-Instruct	70B	85	81	78	74	70	65	59	53	45	67,8