Introduce model set list

[KartikP]

To do

• Use runnable status instead of registered status by loading the model.

Summary

Here we categorize the benchmarks_model table into distinct groups:

• reference
• top10_2024
• base
• NULL

Using these groups, future benchmark submissions will only be evaluated on reference + top10_2024 + base. Supports expanding to other groups.

Reference:

• hmax
• alexnet
• CORnet-S
• resnet-50-robust
• voneresnet-50-non_stochastic
• grcnn
• custom_model_cv_18_dagger_408
• ViT_L_32_imagenet1k
• mobilenet_v2_1_4_224
• pixels

Top10_2024:

• cvt_cvt-w24-384-in22k_finetuned-in1k_4
• resnext101_32x8d_wsl
• effnetb1_cutmixpatch_SAM_
• effnetb1_cutmixpatch_augmix_robust32_avge4e7_manylayers_324x288
• resnext101_32x32d_wsl
• effnetb1_272x240
• resnext101_32x48d_wsl
• pnasnet_large_halfPrecision
• resnet-152_v2
• focalnet_tiny_lrf_in1k

Base:

Represents ~95 models that provide diverse coverage neural region performance, behavioral performance, and non-overlapping top 42 global scores

Selection Criteria:

1. Registered and public (only considers models that are registered and public - goes from >600 registered models to 266 registered and public).
2. Exclude models already part of reference and top10_2024 sets
3. Selection strategy (in order):
   • Top 5 models per neural area (V1, V2, V4, IT) with overlap handling (~20 models)
   • Top 20 behavioral performers (behavior_vision) not already selected
   • Fill remaining with next top-ranked models until reaching 100 models