Modelpublic

AvoCahDoe/llava-1.5-13b-rlmpq-balanced

sourceHugging Faceotherupdated 4mo agoView on Hugging Face

0likes20downloads

Model Card

RL-MPQ Balanced — llava-hf/llava-1.5-13b-hf

RL-MPQ fake-quantized VLM evaluation artifacts.


Repo	`AvoCahDoe/llava-1.5-13b-rlmpq-balanced`
Scenario	Balanced
Avg bits	4.4
Collection	RL-MPQ VLM — LLaVA-1.5-13B
Full results	dataset

Model	avg_bits	group	MMMU	MMBench	ScienceQA	Avg	ΔMMMU	ΔMMBench	ΔScienceQA	ΔAvg
FP16 (baseline)	16.0	llama13b_aggressive	35.33	63.78	71.24	56.78	0.0	0.0	0.0	0.0
INT4 (bnb NF4)	4.0	llama13b_aggressive	35.56	61.76	71.15	56.16	0.23	-2.02	-0.09	-0.62
RL-MPQ Aggressive	3.75	llama13b_aggressive	34.56	63.0	71.1	56.22	-0.77	-0.78	-0.14	-0.56