medical-expert/sft_v3_step2000

Dataset

Task	Num
Grounding	20973
Classification	30000
total	50973

Note: the difference of sft_v3 and sft_v2 is the num of training data, i.e., sft_v3 has near 10 times as much training data as sft_v2.

Train

Global Batchsize: 32
LoRA Rank: 32
1 epoch ~= 1600 step

Eval

	Chexpert (macro F1)	MS-CXR	ChestXray	Vindr-cxr	Overall
chexfound	51.1	\	\	\
eva-x	41.2	\	\	\
unichest	60.7	\	\	\
sft-v3-step1000	60.8	38.2	20.7	27.9	36.90
sft-v3-step1250	55.6	46.0	19.2	30.9	37.92
sft-v3-step1500	61.4	52.3	22.6	31.8	42.02
sft-v3-step1750	59.5	49.9	23.2	30.5	40.77
sft-v3-step2000	62.7	52.5	23.2	32.2	42.65
sft-v3-step2250	62.0	54.7	22.4	32.9	43.00
sft-v3-step2500	56.6	56.2	23.6	32.6	42.25