results.csv

family,method,arch,pretrain,eval,image_size,lp,lp_gap,ep,ep_queries,ep_variant,model,loader,pretrain_tag,finetune,lp_variant,lp_log,ep_log,notes,verified,amp,waiting,gap_reason,knn,knn_k,knn_T,knn_log,extra_flags,ep_head
MIM,MAE,ViT-S/16,IN-1K,IN-1K,224,47.08,no,64.58,32,ep,vit_small_patch16,,,vit_small_patch16_224.mae/checkpoint-799.pth,cls,logs/mae_vits/lp.txt,logs/mae_vits/ep.txt,Checkpoint converted by the [Beyond [cls]](https://github.com/gmum/beyond_cls) authors; not redistributable.,yes,,,,18.02,50,0.07,logs/mae_vits/knn.txt,,final
MIM,MAE,ViT-B/16,IN-1K,IN-1K,224,67.78,no,75.46,32,ep,vit_base_patch16,,,vit_base_patch16_224.mae,cls,logs/mae_base/lp.txt,logs/mae_base/ep.txt,,yes,,,,35.57,20,0.07,logs/mae_base/knn.txt,,final
MIM,MAE,ViT-L/16,IN-1K,IN-1K,224,76.01,no,79.52,32,ep,vit_large_patch16,,,vit_large_patch16_224.mae,cls,logs/mae_vitl/lp.txt,logs/mae_vitl/ep.txt,"LP peaks ep84, EP ep86 -- this MIM encoder uses the whole schedule. GAP scores 73.58, below cls.",yes,,,,48.06,50,0.07,logs/mae_vitl/knn.txt,,final
MIM,BEiTv2,ViT-B/16,IN-1K,IN-1K,224,78.99,no,81.39,32,ep,vit_base_patch16,simmim,,vit_base_patch16_224.beitv2/checkpoint-799.pth,cls,logs/beitv2_vitb/lp.txt,logs/beitv2_vitb/ep.txt,Loads through the SimMIM architecture. Checkpoint converted by the [Beyond [cls]](https://github.com/gmum/beyond_cls) authors; not redistributable.,yes,,,,70.09,50,0.07,logs/beitv2_vitb/knn.txt,,final
MIM,SimMIM,ViT-B/16,IN-1K,IN-1K,224,47.15,no,64.86,32,ep,vit_base_patch16,simmim,,vit_base_patch16_224.simmim/checkpoint-799.pth,cls,logs/simmim_vitb/lp.txt,logs/simmim_vitb/ep.txt,"Both arms peak at ep84. LP is 4.35 below the paper number, EP 0.24 below. Checkpoint converted by the [Beyond [cls]](https://github.com/gmum/beyond_cls) authors; not redistributable.",yes,,,,8.59,50,0.1,logs/simmim_vitb/knn.txt,,final
MIM,CAPI,ViT-L/14,IN-1K,IN-1K,224,81.94,no,82.98,32,ep,capi_vitl14_in1k,capi,,,cls,logs/capi_vitl/lp.txt,logs/capi_vitl/ep.txt,Requires --amp float32: CAPI fails under bfloat16 with an index_add_ dtype mismatch.,yes,float32,,,72.94,50,0.07,logs/capi_vitl/knn.txt,,final
JEA,DINO,ViT-B/16,IN-1K,IN-1K,224,77.19,no,77.39,32,ep_all,vit_base_patch16,,,vit_base_patch16_224.dino,cls,logs/dino_vitb/lp.txt,logs/dino_vitb/ep.txt,"ep_all: DINO self-distills on the [CLS]; plain ep discards it and scores 75.73, 1.46 below LP.",yes,,,,71.64,10,0.07,logs/dino_vitb/knn.txt,,final
Hybrid,iBOT,ViT-B/16,IN-1K,IN-1K,224,78.74,no,78.98,32,ep_all,vit_base_patch16,,,vit_base_patch16_224.ibot/checkpoint-799.pth,cls,logs/ibot_vitb/lp.txt,logs/ibot_vitb/ep.txt,"ep_all: iBOT self-distills on the [CLS]; plain ep scores 78.41, below LP. Checkpoint converted by the [Beyond [cls]](https://github.com/gmum/beyond_cls) authors; not redistributable.",yes,,,,73.58,15,0.07,logs/ibot_vitb/knn.txt,,final
Hybrid,DINOv2,ViT-B/14,LVD-142M,IN-1K,224,83.55,no,84.01,32,ep,dinov2_vitb14,hub,,,cls,logs/dinov2_vitb/lp.txt,logs/dinov2_vitb/ep.txt,,yes,,,,79.34,15,0.07,logs/dinov2_vitb/knn.txt,,final
Hybrid,DINOv2,ViT-L/14,LVD-142M,IN-1K,224,85.25,no,85.56,32,ep_all,dinov2_vitl14,hub,,,cls,logs/dinov2_vitl/lp.txt,logs/dinov2_vitl/ep.txt,"ep_all: DINOv2 self-distills on the [CLS]. Q=32; an earlier Q=64 run scored 85.55. Both arms re-run to their peak with best-epoch checkpointing under identical flags and schedule; the published logs and the HF-uploaded heads are from the recapture, whose peak matched the original to within 0.10 (most arms exactly). Original logs kept as *_prev.txt.",yes,,,,81.02,15,0.07,logs/dinov2_vitl/knn.txt,,peak
Hybrid,DINOv3,ViT-B/16,LVD-1689M,IN-1K,224,83.94,no,84.11,32,ep_all,dinov3_vitb16,dinov3,,,cls,logs/dinov3_vitb/lp.txt,logs/dinov3_vitb/ep.txt,,yes,,,,79.64,15,0.07,logs/dinov3_vitb/knn.txt,,final
Hybrid,DINOv3,ViT-L/16,LVD-1689M,IN-1K,224,86.62,no,87.05,32,ep_all,dinov3_vitl16,dinov3,,,cls,logs/dinov3_vitl/lp.txt,logs/dinov3_vitl/ep.txt,,yes,,,,83.39,200,0.07,logs/dinov3_vitl/knn.txt,,final
VLM,CLIP,ViT-L/14,WIT-400M,IN-1K,224,82.51,no,83.22,32,ep,ViT-L-14,openclip,openai,,cls,logs/clip_vitl/lp.txt,logs/clip_vitl/ep.txt,". Both arms re-run to their peak with best-epoch checkpointing under identical flags and schedule; the published logs and the HF-uploaded heads are from the recapture, whose peak matched the original to within 0.10 (most arms exactly). Original logs kept as *_prev.txt.",yes,,,,73.53,15,0.1,logs/clip_vitl/knn.txt,,peak
VLM,SigLIP,ViT-L/16,"WebLI (10B imgs, 109 langs)",IN-1K,256,84.17,yes,85.93,32,ep,ViT-L-16-SigLIP-256,openclip,webli,,pos,logs/siglip_vitl/lp.txt,logs/siglip_vitl/ep.txt,"LP peaks ep8, EP ep6; both decline over the remaining 80+ epochs. Both arms re-run to their peak with best-epoch checkpointing under identical flags and schedule; the published logs and the HF-uploaded heads are from the recapture, whose peak matched the original to within 0.10 (most arms exactly). Original logs kept as *_prev.txt.",yes,,,attn_pool,72.92,15,0.07,logs/siglip_vitl/knn.txt,,peak
VLM,SigLIP2,ViT-L/16,"WebLI (10B imgs, 109 langs)",IN-1K,256,85.31,yes,87.06,32,ep,ViT-L-16-SigLIP2-256,openclip,webli,,pos,logs/siglip2_vitl/lp.txt,logs/siglip2_vitl/ep.txt,"LP peaks ep10 and falls 1.7 points by ep89; EP peaks ep6 and falls 2.1. Both arms re-run to their peak with best-epoch checkpointing under identical flags and schedule; the published logs and the HF-uploaded heads are from the recapture, whose peak matched the original to within 0.10 (most arms exactly). Original logs kept as *_prev.txt.",yes,,,attn_pool,76.21,15,0.07,logs/siglip2_vitl/knn.txt,,peak
GEN,DiT,DiT-XL/2,IN-1K,IN-1K,256,32.70,yes,57.00,128,ep,DiT-XL/2,dit,,,pos,logs/dit_xl/lp.txt,logs/dit_xl/ep.txt,"Both arms manually stopped at epoch 87-89 of 90. The only row at Q=128 (head 2,627,560 = 1152^2 + 128*1152 + 1152*1000 + 1000). No [CLS], so GAP is forced. --dit_ckpt and --vae are omitted from the command: they default to DiT-XL-2-256x256.pt and sd-vae-ft-mse. The released EP head is this run's final-epoch checkpoint (ep86, 56.94, 0.06 below the ep85 peak), recovered from the original attention-mim tree and verified against the published log line by line",yes,,,no_cls,5.25,50,0.07,logs/dit_xl/knn.txt,,final
GEN,AIMv2,ViT-L/14,DFN-2B + COYO + HQITP,IN-1K,224,84.72,yes,85.94,32,ep,aimv2-large-patch14-224,aimv2,,,pos,logs/aimv2_vitl/lp.txt,logs/aimv2_vitl/ep.txt,"Pre-training (AIMv2 paper Table 2, arXiv:2411.14402): DFN-2B 1.90B alt-text + 3.80B synthetic-caption pairs (30% sampling each), COYO 560M (9%), HQITP 565M alt-text + 432M synthetic (28% / 3%); 12B samples seen. HQITP is Apple-internal, so the mixture is not reproducible from public data.",yes,,,no_cls,76.02,10,0.2,logs/aimv2_vitl/knn.txt,,final
VLM,SigLIP2,SO400M/14,"WebLI (10B imgs, 109 langs)",IN-1K,224,86.27,yes,87.68,32,ep,ViT-SO400M-14-SigLIP2,openclip,webli,,pos,logs/siglip2_so400m/lp.txt,logs/siglip2_so400m/ep.txt,"Early-stopped: LP 17 epochs, EP 30. ep_all scored 87.66, below the reported ep. Both arms re-run to their peak with best-epoch checkpointing under identical flags and schedule; the published logs and the HF head are from the recapture. Original logs kept as *_prev.txt.",yes,,,attn_pool,77.54,15,0.07,logs/siglip2_so400m/knn.txt,,peak
MIM,EVA02,ViT-L/14,IN-22K,IN-1K,224,82.76,no,83.58,32,ep,eva02_large_patch14_224.mim_in22k,timm,,,cls,logs/eva02_vitl_mim/lp.txt,logs/eva02_vitl_mim/ep.txt,"LP peaks ep84, EP ep15. ep_all also scores 83.58 -- this MIM [CLS] holds nothing the patch tokens do not, so plain ep is reported.",yes,,,,73.95,10,0.07,logs/eva02_vitl_mim/knn.txt,,final
Hybrid,RADIO,ViT-L/16,distilled (DFN5B/SigLIP/DINOv2/SAM),IN-1K,224,84.36,no,83.89,32,ep,radio_v2.5-l,radio,,,cls,logs/radio_v25_l/lp.txt,logs/radio_v25_l/ep.txt,"Multi-teacher distillation. Teachers read from the checkpoint's own args: clip = ViT-H-14-378-quickgelu (dfn5b), siglip = ViT-SO400M-14-SigLIP-384 (webli), dino_v2 = dinov2_vitg14_reg, sam = vit-h. cls_token_per_teacher=True with SAM use_summary=False, so the summary is 3xC: LP reads 3072-d while EP pools 1024-d patch tokens. Three of the four teachers optimise a global representation, so LP is strong here. The checkpoint records no training corpus; the RADIOv2.5 paper (arXiv:2412.07679) gives DataComp-1B.",yes,,,,78.21,15,0.07,logs/radio_v25_l/knn.txt,,final
MIM,Hiera,ViT-H/16,IN-1K,IN-1K,224,77.28,yes,79.89,32,ep,hiera_huge_224.mae,timm,,,pos,logs/hiera_huge/lp.txt,logs/hiera_huge/ep.txt,"Both arms peak late (LP ep85, EP ep81).",yes,,,no_cls,34.73,20,0.07,logs/hiera_huge/knn.txt,,final
MIM,Hiera,ViT-L/16,IN-1K,IN-1K,224,74.09,yes,78.55,32,ep,hiera_large_224.mae,timm,,,pos,logs/hiera_large/lp.txt,logs/hiera_large/ep.txt,,yes,,,no_cls,38.76,50,0.07,logs/hiera_large/knn.txt,,final
MIM,Hiera,ViT-B/16,IN-1K,IN-1K,224,69.25,yes,75.68,32,ep,hiera_base_224.mae,timm,,,pos,logs/hiera_base/lp.txt,logs/hiera_base/ep.txt,,yes,,,no_cls,39.84,50,0.07,logs/hiera_base/knn.txt,,final
Hybrid,RADIO,ViT-B/16,distilled (DFN5B/SigLIP/DINOv2/SAM),IN-1K,224,79.72,no,80.29,32,ep,radio_v2.5-b,radio,,,cls,logs/radio_v25_b/lp.txt,logs/radio_v25_b/ep.txt,"Multi-teacher distillation. Teachers read from the checkpoint's own args: clip = ViT-H-14-378-quickgelu (dfn5b), siglip = ViT-SO400M-14-SigLIP-384 (webli), dino_v2 = dinov2_vitg14_reg, sam = vit-h. cls_token_per_teacher=True with SAM use_summary=False, so the summary is 3xC: LP reads 2304-d while EP pools 768-d patch tokens. Three of the four teachers optimise a global representation, so LP is strong here. The checkpoint records no training corpus; the RADIOv2.5 paper (arXiv:2412.07679) gives DataComp-1B.",yes,,,,70.43,20,0.07,logs/radio_v25_b/knn.txt,,final
VLM,CLIP,ViT-B/16,WIT-400M,IN-1K,224,77.86,no,77.85,32,ep_all,ViT-B-16,openclip,openai,,cls,logs/clip_vitb16/lp.txt,logs/clip_vitb16/ep.txt,"ep_all: CLIP trains contrastively on the projected [CLS] and runs in that 512-d space, where plain ep on raw 768-d patch tokens is a different space. LP is cls (77.86); GAP scores 75.17. Both arms re-run to their peak with best-epoch checkpointing under identical flags and schedule; the published logs and the HF-uploaded heads are from the recapture, whose peak matched the original to within 0.10 (most arms exactly). Original logs kept as *_prev.txt.",yes,,,,65.66,15,0.1,logs/clip_vitb16/knn.txt,,peak
VLM,PE-Core,L-14/336,MetaCLIP-curated 5.4B,IN-1K,336,85.44,yes,87.25,32,ep,PE-Core-L-14-336,openclip,meta,,pos,logs/pe_core_l336/lp.txt,logs/pe_core_l336/ep.txt,"LP is GAP: PE-Core pools with its own learned attention head, which scores 87.76 -- above EP -- and is excluded on principle. EP is plain ep; ep_all 87.24. Pre-training (Perception Encoder, arXiv:2504.13181): 5.4B public image alt-text pairs curated with the MetaCLIP text-only pipeline, 58B samples seen at B/L scale. The released L/14-336 is the checkpoint after image+video finetuning, not the image-only model. Both arms re-run to their peak with best-epoch checkpointing under identical flags and schedule; the published logs and the HF-uploaded heads are from the recapture, whose peak matched the original to within 0.10 (most arms exactly). Original logs kept as *_prev.txt.",yes,,,attn_pool,49.07,10,0.07,logs/pe_core_l336/knn.txt,,peak
VLM,EVA02-CLIP,E-14,LAION-2B (4B seen),IN-1K,224,87.44,no,87.70,32,ep,EVA02-E-14,openclip,laion2b_s4b_b115k,,cls,logs/eva02_e14/lp.txt,logs/eva02_e14/ep.txt,"Both arms reported at their epoch-6 peak; logs trimmed at epoch 14 where the early-stopping rule is met, untrimmed runs under outputs/. LP ran to epoch 58 and EP to 28 with no gain. EP wins by 0.26, the thinnest margin at the top of the table. Both arms re-run to their peak with best-epoch checkpointing under identical flags and schedule; the published logs and the HF-uploaded heads are from the recapture, whose peak matched the original to within 0.10 (most arms exactly). Original logs kept as *_prev.txt.",yes,,,,82.21,20,0.1,logs/eva02_e14/knn.txt,,peak
VLM,MetaCLIP2,ViT-bigG/14,MetaCLIP2-worldwide (300+ langs),IN-1K,224,87.07,no,87.11,32,ep,ViT-bigG-14-worldwide,openclip,metaclip2_worldwide,,cls,logs/metaclip2_bigg/lp.txt,logs/metaclip2_bigg/ep.txt,"EP is plain ep; ep_all peaked at 86.95 over 47 epochs, below it. LP is cls (GAP 86.94), 49 epochs, peak ep6. LP reads open_clip's projected 1280-d embedding while EP pools the unprojected 1664-d patch tokens. Data (Meta CLIP 2, arXiv:2507.22062): worldwide alt-text, metadata for 300+ languages, ~44%% English, ~29B pairs seen. The paper's largest model is ViT-H/14; these bigG weights are a later open_clip release (timm/vit_gigantic_patch14_clip_224.metaclip2_worldwide). Both arms re-run to their peak with best-epoch checkpointing under identical flags and schedule; the published logs and the HF-uploaded heads are from the recapture, whose peak matched the original to within 0.10 (most arms exactly). Original logs kept as *_prev.txt.",yes,,,,79.58,15,0.1,logs/metaclip2_bigg/knn.txt,,peak
Hybrid,DINOv3,ViT-7B/16,LVD-1689M,IN-1K,224,87.95,no,88.36,32,ep_all,dinov3_vit7b16,dinov3,,,cls,logs/dinov3_vit7b/lp.txt,logs/dinov3_vit7b/ep.txt,"ep_all at Q=32 (plain ep 88.05). Short arms: LP early-stopped at 17 epochs, EP stopped at 18 on the 48h wall; both peak at ep6-8 then decline, and the early-stopping rule was independently met at epoch 14 on the EP arm. At ~2.4 h/epoch a full 90 would be ~9 node-days per arm. Both arms re-run to their peak with best-epoch checkpointing under identical flags and schedule; the published logs and the HF-uploaded heads are from the recapture, whose peak matched the original to within 0.10 (most arms exactly). Original logs kept as *_prev.txt.",yes,,,,84.20,200,0.1,logs/dinov3_vit7b/knn.txt,,peak
MIM,MaskFeat,ViT-B/16,IN-1K,IN-1K,224,62.24,no,71.76,32,ep,vit_base_patch16,,,vit_base_patch16_224.maskfeat/checkpoint-799.pth,cls,logs/maskfeat_vitb/lp.txt,logs/maskfeat_vitb/ep.txt,"The released checkpoint carries NaN in norm.weight/norm.bias; models_vit never applies that final LayerNorm, so nothing here touches it. GAP scores 56.94, below cls. Checkpoint converted by the [Beyond [cls]](https://github.com/gmum/beyond_cls) authors; not redistributable.",yes,,,,15.94,50,0.1,logs/maskfeat_vitb/knn.txt,,final
JEA,I-JEPA,ViT-H/14,IN-1K,IN-1K,224,78.08,yes,79.00,32,ep,vit_huge_patch14,,,vit_huge_patch14_224.ijepa/checkpoint-799.pth,pos,logs/ijepa_vith/lp.txt,logs/ijepa_vith/ep.txt,"Built with --no_cls_token: I-JEPA has no [CLS] at all, so GAP is the only linear probe available. Checkpoint converted by the [Beyond [cls]](https://github.com/gmum/beyond_cls) authors; not redistributable.",yes,,,no_cls,68.09,50,0.07,logs/ijepa_vith/knn.txt,--no_cls_token,final
JEA,MoCov3,ViT-B/16,IN-1K,IN-1K,224,75.70,no,76.53,32,ep_all,vit_base_patch16,,,vit_base_patch16_224.mocov3/checkpoint-799.pth,cls,logs/mocov3_vitb/lp.txt,logs/mocov3_vitb/ep.txt,"ep_all: MoCov3 is contrastive on the [CLS]; plain ep 75.90, ep_all 76.53. Checkpoint converted by the [Beyond [cls]](https://github.com/gmum/beyond_cls) authors; not redistributable.",yes,,,,66.76,20,0.07,logs/mocov3_vitb/knn.txt,,final
Hybrid,iBOT,ViT-L/16,IN-1K,IN-1K,224,80.54,no,79.97,32,ep_all,vit_large_patch16,,,vit_large_patch16_224.ibot/checkpoint-799.pth,cls,logs/ibot_vitl/lp.txt,logs/ibot_vitl/ep.txt,EP loses to the [CLS] by 0.57. Checkpoint converted by the [Beyond [cls]](https://github.com/gmum/beyond_cls) authors; not redistributable.,yes,,,,73.92,10,0.07,logs/ibot_vitl/knn.txt,,final
VLM,MetaCLIP2,ViT-bigG/14-378,MetaCLIP2-worldwide (300+ langs),IN-1K,378,88.09,no,88.12,32,ep,ViT-bigG-14-worldwide-378,openclip,metaclip2_worldwide,,cls,logs/metaclip2_bigg378/lp.txt,logs/metaclip2_bigg378/ep.txt,"Same checkpoint as the 224 row at 378px. Both arms manually stopped and reported at their epoch-6 peak: LP after 20 epochs, EP after 10, at ~4.5 h/epoch against ~17 node-days per arm for a full schedule. Every VLM here peaks at epoch 6-12, and the 224 run of this checkpoint peaks at ep6 across 49 LP and 30 EP epochs. EP is plain ep; ep_all 88.08. Data (Meta CLIP 2, arXiv:2507.22062): worldwide alt-text, metadata for 300+ languages, ~44%% English, ~29B pairs seen. The paper's largest model is ViT-H/14; these bigG weights are a later open_clip release (timm/vit_gigantic_patch14_clip_224.metaclip2_worldwide). Both arms re-run to their peak with best-epoch checkpointing under identical flags and schedule; the published logs and the HF-uploaded heads are from the recapture, whose peak matched the original to within 0.10 (most arms exactly). Original logs kept as *_prev.txt.",yes,,,,80.59,15,0.07,logs/metaclip2_bigg378/knn.txt,,peak
MIM,MaskFeat,ViT-L/16,IN-1K,IN-1K,224,40.85,no,69.63,32,ep,vit_large_patch16,,,vit_large_patch16_224.maskfeat/checkpoint-799.pth,cls,logs/maskfeat_large/lp.txt,logs/maskfeat_large/ep.txt,"This checkpoint's 1000-way head is entirely NaN; main_linprobe drops head.weight/head.bias and refuses NaN tensors, without which training dies at epoch 0 with ""Loss is nan"". LP peaks ep88, EP ep84. k-NN reported as raw (5.72) like every other row; final_norm gives 5.75. Inverted scaling: ViT-L trails ViT-B on all three measures -- k-NN 5.7 vs 15.9, LP 40.9 vs 62.2, EP 69.6 vs 71.8.",yes,,,,5.72,50,0.07,logs/maskfeat_large/knn.txt,,final
VLM,EVA02-CLIP,E-14-plus,LAION-2B (9B seen),IN-1K,224,87.82,no,87.98,32,ep,EVA02-E-14-plus,openclip,laion2b_s9b_b144k,,cls,logs/eva02_e14_plus/lp.txt,logs/eva02_e14_plus/ep.txt,"Larger-batch variant of E-14 (laion2b_s9b_b144k). Both arms reported at their epoch-6 peak; logs trimmed at epoch 14, untrimmed runs under outputs/. LP ran to epoch 25 with no gain. Both arms re-run to their peak with best-epoch checkpointing under identical flags and schedule; the published logs and the HF-uploaded heads are from the recapture, whose peak matched the original to within 0.10 (most arms exactly). Original logs kept as *_prev.txt.",yes,,,,82.92,15,0.1,logs/eva02_e14_plus/knn.txt,,peak
Hybrid,Franca,ViT-L/14,LAION-600M,IN-1K,224,83.63,no,84.28,32,ep_all,franca_vitl14,franca,,,cls,logs/franca_laion/lp.txt,logs/franca_laion/ep_all.txt,"Run WITHOUT --use_rasa_head: upstream builds RASAHead(n_pos_layers=9) while both published RASA checkpoints carry only 8 pre_pos_layers, so the flag cannot load against valeoai/Franca v1.0.0. The backbone franca_vitl14_Laion600M.pth loads cleanly and is what is probed. EP is ep_all (84.28) over plain ep (83.92). LP peaks ep82, EP ep14. Replaces the IN-21k row: release v1.0.0 never uploaded franca_vitl14_In21K.pth. Both arms re-run to their peak with best-epoch checkpointing under identical flags and schedule; the published logs and the HF-uploaded heads are from the recapture, whose peak matched the original to within 0.10 (most arms exactly). Original logs kept as *_prev.txt.",yes,,,,77.98,10,0.1,logs/franca_laion/knn.txt,--franca_weights LAION,peak