uukuguy/speechless-mistral-7b-dare-0.85

Experiment for DARE(Drop and REscale), most of the delta parameters can be directly set to zeros without affecting the capabilities of SFT LMs and larger models can tolerate a higher proportion of discarded parameters.

Merged with below DARE models.

weight_mask_rate: 0.85 / use_weight_rescale: True / mask_stratery: random / scaling_coefficient: 1.0

Model	Average	ARC	HellaSwag	MMLU	TruthfulQA	Winogrande	GSM8K	DROP
Intel/neural-chat-7b-v3-1	59.06	66.21	83.64	62.37	59.65	78.14	19.56	43.84
migtissera/SynthIA-7B-v1.3	57.11	62.12	83.45	62.65	51.37	78.85	17.59	43.76
bhenrym14/mistral-7b-platypus-fp16	56.89	63.05	84.15	64.11	45.07	78.53	17.36	45.92
jondurbin/airoboros-m-7b-3.1.2	56.24	61.86	83.51	61.91	53.75	77.58	13.87	41.2
teknium/CollectiveCognition-v1.1-Mistral-7B	53.87	62.12	84.17	62.35	57.62	75.37	15.62	19.85
uukuguy/speechless-mistral-dolphin-orca-platypus-samantha-7b	53.34	64.33	84.4	63.72	52.52	78.37	21.38	8.66

2023.12.04

It seems that there are some issues with the calculation of the GSM8K and DROP metrics on the Open LLM Leaderboard. Currently, the DROP metric has been removed from the official website, while the calculation of GSM8K metric remains chaotic, with significant differences in values among various models. Therefore, I am temporarily using ARC, HellaSwag, MMLU, TruthfulQA, and Winogrande metrics to evaluate the performance of DARE.

Model	Average	ARC	HellaSwag	MMLU	TruthfulQA	Winogrande
CollectiveCognition-v1.1-Mistral-7B	68.326	62.12	84.17	62.35	57.62	75.37
CollectiveCognition-v1.1-Mistral-7B-dare-0.85	66.676	61.01	84.31	64.34	44.87	78.85
airoboros-m-7b-3.1.2	67.722	61.86	83.51	61.91	53.75	77.58
airoboros-m-7b-3.1.2-dare-0.85	66.144	61.09	83.57	64.05	43.64	78.37
SynthIA-7B-v1.3	67.688	62.12	83.45	62.65	51.37	78.85
SynthIA-7B-v1.3-dare-0.85	66.340	61.01	83.50	64.49	43.77	78.93
speechless-mistral-7b-dare-0.85 (merge 6 DARE models)	68.516	63.57	84.82	64.29	50.66	79.24

From the official website evaluation results, after deleting 85% of the incremental parameters, the overall indicators remain above 97.5% of the original performance indicators. Among them, ARC slightly decreases, TruthfulQA significantly decreases, MMLU significantly increases, and HellaSwagt and Winogrande slightly increase. The most significant impact is the significant decrease in TruthfulQA, while other indicators are relatively well maintained, with MMLU showing a noticeable increase.