Hugging Face's logo

Adv-FT-Interp
/
llava-1.5-7b-hf-s3-patch_loss-R1

PyTorch
Model card Files
xet
 Community
agarwali commited on
Commit
ccb182b
·
verified ·
1 Parent(s): 0e26e8c

Final final_layer_patch_loss model: llava-1.5-7b-hf-r1-s3-patch_loss-h200-20260821-050840

Browse files
Files changed (27) hide show
  1. .gitattributes +1 -0
  2. checkpoints/epoch_0001/faithful_training_metadata.json +8 -0
  3. checkpoints/epoch_0001/preprocessor_config.json +28 -0
  4. checkpoints/epoch_0001/pytorch_model.bin +3 -0
  5. checkpoints/epoch_0001/visual_projection.bin +3 -0
  6. checkpoints/epoch_0002/faithful_training_metadata.json +8 -0
  7. checkpoints/epoch_0002/preprocessor_config.json +28 -0
  8. checkpoints/epoch_0002/pytorch_model.bin +3 -0
  9. checkpoints/epoch_0002/visual_projection.bin +3 -0
  10. checkpoints/epoch_0003/faithful_training_metadata.json +8 -0
  11. checkpoints/epoch_0003/preprocessor_config.json +28 -0
  12. checkpoints/epoch_0003/pytorch_model.bin +3 -0
  13. checkpoints/epoch_0003/visual_projection.bin +3 -0
  14. faithful_training_metadata.json +8 -0
  15. preprocessor_config.json +28 -0
  16. pytorch_model.bin +3 -0
  17. training_history.json +193 -0
  18. visual_projection.bin +3 -0
  19. wandb/debug-internal.log +6 -0
  20. wandb/debug.log +22 -0
  21. wandb/run-20260821_051826-4qwwcf5b/files/output.log +41 -0
  22. wandb/run-20260821_051826-4qwwcf5b/files/requirements.txt +277 -0
  23. wandb/run-20260821_051826-4qwwcf5b/files/wandb-metadata.json +150 -0
  24. wandb/run-20260821_051826-4qwwcf5b/logs/debug-core.log +6 -0
  25. wandb/run-20260821_051826-4qwwcf5b/logs/debug-internal.log +6 -0
  26. wandb/run-20260821_051826-4qwwcf5b/logs/debug.log +22 -0
  27. wandb/run-20260821_051826-4qwwcf5b/run-4qwwcf5b.wandb +3 -0
.gitattributes CHANGED
@@ -33,3 +33,4 @@ saved_model/**/* filter=lfs diff=lfs merge=lfs -text
33
  *.zip filter=lfs diff=lfs merge=lfs -text
34
  *.zst filter=lfs diff=lfs merge=lfs -text
35
  *tfevents* filter=lfs diff=lfs merge=lfs -text
 
 
33
  *.zip filter=lfs diff=lfs merge=lfs -text
34
  *.zst filter=lfs diff=lfs merge=lfs -text
35
  *tfevents* filter=lfs diff=lfs merge=lfs -text
36
+ wandb/run-20260821_051826-4qwwcf5b/run-4qwwcf5b.wandb filter=lfs diff=lfs merge=lfs -text
checkpoints/epoch_0001/faithful_training_metadata.json ADDED
@@ -0,0 +1,8 @@
 
 
 
 
 
 
 
 
 
1
+ {
2
+ "format": "llava_vision_tower_plus_faithful_clip_projection",
3
+ "vision_checkpoint": "pytorch_model.bin",
4
+ "projection_checkpoint": "visual_projection.bin",
5
+ "clip_model_name": "openai/clip-vit-large-patch14-336",
6
+ "projection_is_used_by_llava_inference": false,
7
+ "vision_tower_is_loadable_by_existing_inference_loader": true
8
+ }
checkpoints/epoch_0001/preprocessor_config.json ADDED
@@ -0,0 +1,28 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ {
2
+ "crop_size": {
3
+ "height": 336,
4
+ "width": 336
5
+ },
6
+ "do_center_crop": true,
7
+ "do_convert_rgb": true,
8
+ "do_normalize": true,
9
+ "do_rescale": true,
10
+ "do_resize": true,
11
+ "image_mean": [
12
+ 0.48145466,
13
+ 0.4578275,
14
+ 0.40821073
15
+ ],
16
+ "image_processor_type": "CLIPImageProcessor",
17
+ "image_std": [
18
+ 0.26862954,
19
+ 0.26130258,
20
+ 0.27577711
21
+ ],
22
+ "processor_class": "LlavaProcessor",
23
+ "resample": 3,
24
+ "rescale_factor": 0.00392156862745098,
25
+ "size": {
26
+ "shortest_edge": 336
27
+ }
28
+ }
checkpoints/epoch_0001/pytorch_model.bin ADDED
@@ -0,0 +1,3 @@
 
 
 
 
1
+ version https://git-lfs.github.com/spec/v1
2
+ oid sha256:a3a9a51ae67cb21af2e95ad1f5c3800015125ce6dbd50f9ba4617bc69675ad2e
3
+ size 1214160127
checkpoints/epoch_0001/visual_projection.bin ADDED
@@ -0,0 +1,3 @@
 
 
 
 
1
+ version https://git-lfs.github.com/spec/v1
2
+ oid sha256:a145f624ef78eb86b012030649eeba34cc97f7607911b7cc7f12c68ffc987827
3
+ size 3147503
checkpoints/epoch_0002/faithful_training_metadata.json ADDED
@@ -0,0 +1,8 @@
 
 
 
 
 
 
 
 
 
1
+ {
2
+ "format": "llava_vision_tower_plus_faithful_clip_projection",
3
+ "vision_checkpoint": "pytorch_model.bin",
4
+ "projection_checkpoint": "visual_projection.bin",
5
+ "clip_model_name": "openai/clip-vit-large-patch14-336",
6
+ "projection_is_used_by_llava_inference": false,
7
+ "vision_tower_is_loadable_by_existing_inference_loader": true
8
+ }
checkpoints/epoch_0002/preprocessor_config.json ADDED
@@ -0,0 +1,28 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ {
2
+ "crop_size": {
3
+ "height": 336,
4
+ "width": 336
5
+ },
6
+ "do_center_crop": true,
7
+ "do_convert_rgb": true,
8
+ "do_normalize": true,
9
+ "do_rescale": true,
10
+ "do_resize": true,
11
+ "image_mean": [
12
+ 0.48145466,
13
+ 0.4578275,
14
+ 0.40821073
15
+ ],
16
+ "image_processor_type": "CLIPImageProcessor",
17
+ "image_std": [
18
+ 0.26862954,
19
+ 0.26130258,
20
+ 0.27577711
21
+ ],
22
+ "processor_class": "LlavaProcessor",
23
+ "resample": 3,
24
+ "rescale_factor": 0.00392156862745098,
25
+ "size": {
26
+ "shortest_edge": 336
27
+ }
28
+ }
checkpoints/epoch_0002/pytorch_model.bin ADDED
@@ -0,0 +1,3 @@
 
 
 
 
1
+ version https://git-lfs.github.com/spec/v1
2
+ oid sha256:2352d4058f139d29ed3e9f4f4026f905606a5f01d4af0375334c8210a0a52ffa
3
+ size 1214160127
checkpoints/epoch_0002/visual_projection.bin ADDED
@@ -0,0 +1,3 @@
 
 
 
 
1
+ version https://git-lfs.github.com/spec/v1
2
+ oid sha256:d7eb098f92d12ffd066834b47662f522ee9ae975a78469ca807acb92cf2fcb0f
3
+ size 3147503
checkpoints/epoch_0003/faithful_training_metadata.json ADDED
@@ -0,0 +1,8 @@
 
 
 
 
 
 
 
 
 
1
+ {
2
+ "format": "llava_vision_tower_plus_faithful_clip_projection",
3
+ "vision_checkpoint": "pytorch_model.bin",
4
+ "projection_checkpoint": "visual_projection.bin",
5
+ "clip_model_name": "openai/clip-vit-large-patch14-336",
6
+ "projection_is_used_by_llava_inference": false,
7
+ "vision_tower_is_loadable_by_existing_inference_loader": true
8
+ }
checkpoints/epoch_0003/preprocessor_config.json ADDED
@@ -0,0 +1,28 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ {
2
+ "crop_size": {
3
+ "height": 336,
4
+ "width": 336
5
+ },
6
+ "do_center_crop": true,
7
+ "do_convert_rgb": true,
8
+ "do_normalize": true,
9
+ "do_rescale": true,
10
+ "do_resize": true,
11
+ "image_mean": [
12
+ 0.48145466,
13
+ 0.4578275,
14
+ 0.40821073
15
+ ],
16
+ "image_processor_type": "CLIPImageProcessor",
17
+ "image_std": [
18
+ 0.26862954,
19
+ 0.26130258,
20
+ 0.27577711
21
+ ],
22
+ "processor_class": "LlavaProcessor",
23
+ "resample": 3,
24
+ "rescale_factor": 0.00392156862745098,
25
+ "size": {
26
+ "shortest_edge": 336
27
+ }
28
+ }
checkpoints/epoch_0003/pytorch_model.bin ADDED
@@ -0,0 +1,3 @@
 
 
 
 
1
+ version https://git-lfs.github.com/spec/v1
2
+ oid sha256:3c6dfe6100c06dbc09b6aaf798970ca185c3ebf160634a7ac1eee1a3ff4fbefb
3
+ size 1214160127
checkpoints/epoch_0003/visual_projection.bin ADDED
@@ -0,0 +1,3 @@
 
 
 
 
1
+ version https://git-lfs.github.com/spec/v1
2
+ oid sha256:703e1f4d69e39315f97abc65f00a19f0680dd386122cddd63801c608918886ae
3
+ size 3147503
faithful_training_metadata.json ADDED
@@ -0,0 +1,8 @@
 
 
 
 
 
 
 
 
 
1
+ {
2
+ "format": "llava_vision_tower_plus_faithful_clip_projection",
3
+ "vision_checkpoint": "pytorch_model.bin",
4
+ "projection_checkpoint": "visual_projection.bin",
5
+ "clip_model_name": "openai/clip-vit-large-patch14-336",
6
+ "projection_is_used_by_llava_inference": false,
7
+ "vision_tower_is_loadable_by_existing_inference_loader": true
8
+ }
preprocessor_config.json ADDED
@@ -0,0 +1,28 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ {
2
+ "crop_size": {
3
+ "height": 336,
4
+ "width": 336
5
+ },
6
+ "do_center_crop": true,
7
+ "do_convert_rgb": true,
8
+ "do_normalize": true,
9
+ "do_rescale": true,
10
+ "do_resize": true,
11
+ "image_mean": [
12
+ 0.48145466,
13
+ 0.4578275,
14
+ 0.40821073
15
+ ],
16
+ "image_processor_type": "CLIPImageProcessor",
17
+ "image_std": [
18
+ 0.26862954,
19
+ 0.26130258,
20
+ 0.27577711
21
+ ],
22
+ "processor_class": "LlavaProcessor",
23
+ "resample": 3,
24
+ "rescale_factor": 0.00392156862745098,
25
+ "size": {
26
+ "shortest_edge": 336
27
+ }
28
+ }
pytorch_model.bin ADDED
@@ -0,0 +1,3 @@
 
 
 
 
1
+ version https://git-lfs.github.com/spec/v1
2
+ oid sha256:3c6dfe6100c06dbc09b6aaf798970ca185c3ebf160634a7ac1eee1a3ff4fbefb
3
+ size 1214160127
training_history.json ADDED
@@ -0,0 +1,193 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ {
2
+ "training_script": "finetuning_patch_loss.py",
3
+ "experiment_name": "final_layer_patch_loss",
4
+ "run_name": "llava-1.5-7b-hf-r1-s3-patch_loss-h200-20260821-050840",
5
+ "model_name": "llava-hf/llava-1.5-7b-hf",
6
+ "clip_model_name": "openai/clip-vit-large-patch14-336",
7
+ "device": "cuda",
8
+ "split_metadata_path": "/orcd/home/002/agarwali/projects/adv_ft_interp/scripts/finetuning_source_splits.json",
9
+ "num_train_pairs": 120000,
10
+ "num_val_pairs": 15000,
11
+ "num_test_pairs": 15000,
12
+ "projection_frozen": false,
13
+ "objective": {
14
+ "formula": "projected_global_l2 + lambda_patch * mean_normalized_final_patch_l2",
15
+ "lambda_patch": 0.1,
16
+ "patch_correspondence": "same_position",
17
+ "cls_excluded": true
18
+ },
19
+ "args": {
20
+ "objective_smoke_test": false,
21
+ "model_name": "llava-hf/llava-1.5-7b-hf",
22
+ "clip_model_name": "openai/clip-vit-large-patch14-336",
23
+ "original_path": "/home/agarwali/orcd/datasets/imagenet/images_complete/ilsvrc/val",
24
+ "perturbed_path": "/home/agarwali/orcd/pool/ImageNet-C",
25
+ "output_dir": "/orcd/home/002/agarwali/projects/adv_ft_interp/finetuning_runs",
26
+ "run_name": "llava-1.5-7b-hf-r1-s3-patch_loss-h200-20260821-050840",
27
+ "split_metadata_path": "/orcd/home/002/agarwali/projects/adv_ft_interp/scripts/finetuning_source_splits.json",
28
+ "corruptions": null,
29
+ "eval_corruptions": null,
30
+ "rotation": "R1",
31
+ "severities": "3",
32
+ "max_pairs": null,
33
+ "train_frac": 0.8,
34
+ "val_frac": 0.1,
35
+ "seed": 0,
36
+ "num_epochs": 3,
37
+ "batch_size": 64,
38
+ "skip_initial_metrics": false,
39
+ "expected_train_examples": 120000,
40
+ "expected_steps_per_epoch": 1875,
41
+ "num_workers": 8,
42
+ "lr": 1e-05,
43
+ "weight_decay": 0.01,
44
+ "lr_scheduler": "cosine",
45
+ "warmup_ratio": 0.1,
46
+ "max_grad_norm": 1.0,
47
+ "torch_dtype": "float16",
48
+ "device": "cuda",
49
+ "freeze_projection": false,
50
+ "wandb_project": "llava-1-5-7b-hf-patch_loss-ft",
51
+ "wandb_entity": "adv-ft-interp",
52
+ "wandb_mode": "online",
53
+ "wandb_run_name": "llava-1.5-7b-hf-r1-s3-patch_loss-h200-20260821-050840",
54
+ "wandb_log_every_steps": 125,
55
+ "push_to_hub": true,
56
+ "repo_id": "Adv-FT-Interp/llava-1.5-7b-hf-s3-patch_loss",
57
+ "lambda_patch": 0.1
58
+ },
59
+ "checkpoint": {
60
+ "format": "llava_vision_tower_plus_faithful_clip_projection",
61
+ "vision_checkpoint": "pytorch_model.bin",
62
+ "projection_checkpoint": "visual_projection.bin",
63
+ "clip_model_name": "openai/clip-vit-large-patch14-336",
64
+ "projection_is_used_by_llava_inference": false,
65
+ "vision_tower_is_loadable_by_existing_inference_loader": true
66
+ },
67
+ "history": {
68
+ "epoch": [
69
+ 0,
70
+ 1,
71
+ 2,
72
+ 3
73
+ ],
74
+ "train": [
75
+ {
76
+ "loss": 214.70432809244792,
77
+ "global_l2": 214.60581822102864,
78
+ "patch_l2": 0.9850996924400329,
79
+ "global_cosine_similarity": 0.7038396677970886,
80
+ "patch_cosine_similarity": 0.5074501883983612,
81
+ "learning_rate": 0.0
82
+ },
83
+ {
84
+ "loss": 55.9442484354655,
85
+ "global_l2": 55.84701837158203,
86
+ "patch_l2": 0.9723009362856547,
87
+ "global_cosine_similarity": 0.9134025128046671,
88
+ "patch_cosine_similarity": 0.5138495662530264,
89
+ "learning_rate": 8.430372292240985e-06
90
+ },
91
+ {
92
+ "loss": 33.018047685750325,
93
+ "global_l2": 32.92380768737793,
94
+ "patch_l2": 0.9423996650695801,
95
+ "global_cosine_similarity": 0.9492187916755677,
96
+ "patch_cosine_similarity": 0.5288002019246419,
97
+ "learning_rate": 3.019073633524573e-06
98
+ },
99
+ {
100
+ "loss": 24.365502514648437,
101
+ "global_l2": 24.273061657714845,
102
+ "patch_l2": 0.9244090539932251,
103
+ "global_cosine_similarity": 0.9627302259445191,
104
+ "patch_cosine_similarity": 0.5377955075263977,
105
+ "learning_rate": 0.0
106
+ }
107
+ ],
108
+ "val": [
109
+ {
110
+ "loss": 215.55668030598957,
111
+ "global_l2": 215.4582931640625,
112
+ "patch_l2": 0.9838897080739339,
113
+ "global_cosine_similarity": 0.7020610805511475,
114
+ "patch_cosine_similarity": 0.5080551821390787
115
+ },
116
+ {
117
+ "loss": 53.70806598307291,
118
+ "global_l2": 53.610738907877604,
119
+ "patch_l2": 0.973270923614502,
120
+ "global_cosine_similarity": 0.9158644541422526,
121
+ "patch_cosine_similarity": 0.5133645721435547
122
+ },
123
+ {
124
+ "loss": 49.97278981119792,
125
+ "global_l2": 49.87842252604167,
126
+ "patch_l2": 0.9436731877644857,
127
+ "global_cosine_similarity": 0.9222555021921793,
128
+ "patch_cosine_similarity": 0.5281634449005127
129
+ },
130
+ {
131
+ "loss": 49.36027342936198,
132
+ "global_l2": 49.26644939778646,
133
+ "patch_l2": 0.9382394207000733,
134
+ "global_cosine_similarity": 0.9234211781819661,
135
+ "patch_cosine_similarity": 0.5308803233464559
136
+ }
137
+ ],
138
+ "classification": [
139
+ {
140
+ "loss": 0.5977288237571716,
141
+ "embedding_l2": 0.5977288237571716,
142
+ "embedding_cosine_similarity": 0.7011356225967407,
143
+ "original_accuracy": 0.779,
144
+ "perturbed_accuracy": 0.38945,
145
+ "perturbed_accuracy_by_corruption": {
146
+ "defocus_blur": 0.4162,
147
+ "glass_blur": 0.2478,
148
+ "motion_blur": 0.503,
149
+ "zoom_blur": 0.3908
150
+ }
151
+ },
152
+ {
153
+ "loss": 0.19579188580513002,
154
+ "embedding_l2": 0.19579188580513002,
155
+ "embedding_cosine_similarity": 0.9021040996551514,
156
+ "original_accuracy": 0.7524,
157
+ "perturbed_accuracy": 0.6436,
158
+ "perturbed_accuracy_by_corruption": {
159
+ "defocus_blur": 0.6716,
160
+ "glass_blur": 0.6608,
161
+ "motion_blur": 0.7098,
162
+ "zoom_blur": 0.5322
163
+ }
164
+ },
165
+ {
166
+ "loss": 0.1871997651576996,
167
+ "embedding_l2": 0.1871997651576996,
168
+ "embedding_cosine_similarity": 0.9064001627922058,
169
+ "original_accuracy": 0.752,
170
+ "perturbed_accuracy": 0.6605,
171
+ "perturbed_accuracy_by_corruption": {
172
+ "defocus_blur": 0.6882,
173
+ "glass_blur": 0.6884,
174
+ "motion_blur": 0.7268,
175
+ "zoom_blur": 0.5386
176
+ }
177
+ },
178
+ {
179
+ "loss": 0.18541483886241913,
180
+ "embedding_l2": 0.18541483886241913,
181
+ "embedding_cosine_similarity": 0.9072926286697388,
182
+ "original_accuracy": 0.7568,
183
+ "perturbed_accuracy": 0.66665,
184
+ "perturbed_accuracy_by_corruption": {
185
+ "defocus_blur": 0.695,
186
+ "glass_blur": 0.6914,
187
+ "motion_blur": 0.7296,
188
+ "zoom_blur": 0.5506
189
+ }
190
+ }
191
+ ]
192
+ }
193
+ }
visual_projection.bin ADDED
@@ -0,0 +1,3 @@
 
 
 
 
1
+ version https://git-lfs.github.com/spec/v1
2
+ oid sha256:703e1f4d69e39315f97abc65f00a19f0680dd386122cddd63801c608918886ae
3
+ size 3147503
wandb/debug-internal.log ADDED
@@ -0,0 +1,6 @@
 
 
 
 
 
 
 
1
+ {"time":"2026-08-21T05:18:27.602575633-04:00","level":"INFO","msg":"stream: starting","core version":"0.23.1"}
2
+ {"time":"2026-08-21T05:18:28.098338792-04:00","level":"INFO","msg":"stream: created new stream","id":"4qwwcf5b"}
3
+ {"time":"2026-08-21T05:18:28.098401348-04:00","level":"INFO","msg":"handler: started","stream_id":"4qwwcf5b"}
4
+ {"time":"2026-08-21T05:18:28.099180044-04:00","level":"INFO","msg":"stream: started","id":"4qwwcf5b"}
5
+ {"time":"2026-08-21T05:18:28.099187832-04:00","level":"INFO","msg":"writer: started","stream_id":"4qwwcf5b"}
6
+ {"time":"2026-08-21T05:18:28.099198132-04:00","level":"INFO","msg":"sender: started","stream_id":"4qwwcf5b"}
wandb/debug.log ADDED
@@ -0,0 +1,22 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ 2026-08-21 05:18:26,314 INFO MainThread:769791 [wandb_setup.py:_flush():80] Current SDK version is 0.23.1
2
+ 2026-08-21 05:18:26,314 INFO MainThread:769791 [wandb_setup.py:_flush():80] Configure stats pid to 769791
3
+ 2026-08-21 05:18:26,314 INFO MainThread:769791 [wandb_setup.py:_flush():80] Loading settings from /home/agarwali/.config/wandb/settings
4
+ 2026-08-21 05:18:26,314 INFO MainThread:769791 [wandb_setup.py:_flush():80] Loading settings from /orcd/home/002/agarwali/projects/adv_ft_interp/wandb/settings
5
+ 2026-08-21 05:18:26,314 INFO MainThread:769791 [wandb_setup.py:_flush():80] Loading settings from environment variables
6
+ 2026-08-21 05:18:26,314 INFO MainThread:769791 [wandb_init.py:setup_run_log_directory():714] Logging user logs to /orcd/home/002/agarwali/projects/adv_ft_interp/finetuning_runs/llava-1.5-7b-hf-r1-s3-patch_loss-h200-20260821-050840/wandb/run-20260821_051826-4qwwcf5b/logs/debug.log
7
+ 2026-08-21 05:18:26,314 INFO MainThread:769791 [wandb_init.py:setup_run_log_directory():715] Logging internal logs to /orcd/home/002/agarwali/projects/adv_ft_interp/finetuning_runs/llava-1.5-7b-hf-r1-s3-patch_loss-h200-20260821-050840/wandb/run-20260821_051826-4qwwcf5b/logs/debug-internal.log
8
+ 2026-08-21 05:18:26,314 INFO MainThread:769791 [wandb_init.py:init():841] calling init triggers
9
+ 2026-08-21 05:18:26,314 INFO MainThread:769791 [wandb_init.py:init():846] wandb.init called with sweep_config: {}
10
+ config: {'objective_smoke_test': False, 'model_name': 'llava-hf/llava-1.5-7b-hf', 'clip_model_name': 'openai/clip-vit-large-patch14-336', 'original_path': '/home/agarwali/orcd/datasets/imagenet/images_complete/ilsvrc/val', 'perturbed_path': '/home/agarwali/orcd/pool/ImageNet-C', 'output_dir': '/orcd/home/002/agarwali/projects/adv_ft_interp/finetuning_runs', 'run_name': 'llava-1.5-7b-hf-r1-s3-patch_loss-h200-20260821-050840', 'split_metadata_path': '/orcd/home/002/agarwali/projects/adv_ft_interp/scripts/finetuning_source_splits.json', 'corruptions': None, 'eval_corruptions': None, 'rotation': 'R1', 'severities': '3', 'max_pairs': None, 'train_frac': 0.8, 'val_frac': 0.1, 'seed': 0, 'num_epochs': 3, 'batch_size': 64, 'skip_initial_metrics': False, 'expected_train_examples': 120000, 'expected_steps_per_epoch': 1875, 'num_workers': 8, 'lr': 1e-05, 'weight_decay': 0.01, 'lr_scheduler': 'cosine', 'warmup_ratio': 0.1, 'max_grad_norm': 1.0, 'torch_dtype': 'float16', 'device': 'cuda', 'freeze_projection': False, 'wandb_project': 'llava-1-5-7b-hf-patch_loss-ft', 'wandb_entity': 'adv-ft-interp', 'wandb_mode': 'online', 'wandb_run_name': 'llava-1.5-7b-hf-r1-s3-patch_loss-h200-20260821-050840', 'wandb_log_every_steps': 125, 'push_to_hub': True, 'repo_id': 'Adv-FT-Interp/llava-1.5-7b-hf-s3-patch_loss', 'lambda_patch': 0.1, '_wandb': {}}
11
+ 2026-08-21 05:18:26,314 INFO MainThread:769791 [wandb_init.py:init():889] starting backend
12
+ 2026-08-21 05:18:27,536 INFO MainThread:769791 [wandb_init.py:init():892] sending inform_init request
13
+ 2026-08-21 05:18:27,591 INFO MainThread:769791 [wandb_init.py:init():900] backend started and connected
14
+ 2026-08-21 05:18:27,595 INFO MainThread:769791 [wandb_init.py:init():970] updated telemetry
15
+ 2026-08-21 05:18:27,675 INFO MainThread:769791 [wandb_init.py:init():994] communicating run to backend with 90.0 second timeout
16
+ 2026-08-21 05:18:28,457 INFO MainThread:769791 [wandb_init.py:init():1041] starting run threads in backend
17
+ 2026-08-21 05:18:29,373 INFO MainThread:769791 [wandb_run.py:_console_start():2521] atexit reg
18
+ 2026-08-21 05:18:29,373 INFO MainThread:769791 [wandb_run.py:_redirect():2369] redirect: wrap_raw
19
+ 2026-08-21 05:18:29,373 INFO MainThread:769791 [wandb_run.py:_redirect():2438] Wrapping output streams.
20
+ 2026-08-21 05:18:29,373 INFO MainThread:769791 [wandb_run.py:_redirect():2461] Redirects installed.
21
+ 2026-08-21 05:18:29,383 INFO MainThread:769791 [wandb_init.py:init():1081] run started, returning control to user process
22
+ 2026-08-21 05:18:29,383 INFO MainThread:769791 [wandb_run.py:_config_callback():1396] config_cb None None {'training_script': 'finetuning_patch_loss.py', 'experiment_name': 'final_layer_patch_loss', 'objective': {'formula': 'projected_global_l2 + lambda_patch * mean_normalized_final_patch_l2', 'lambda_patch': 0.1, 'patch_correspondence': 'same_position', 'cls_excluded': True}, 'projection_frozen': False}
wandb/run-20260821_051826-4qwwcf5b/files/output.log ADDED
@@ -0,0 +1,41 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ building class prototypes: 100%|██████████| 1875/1875 [19:34<00:00, 1.60it/s]
2
+ optimizer=AdamW lr=1e-05 weight_decay=0.01 scheduler=cosine warmup_steps=562
3
+ epoch 0 train metrics: 100%|██████████| 1875/1875 [36:27<00:00, 1.17s/it]
4
+ epoch 0 validation: 100%|██████████| 235/235 [04:38<00:00, 1.19s/it]
5
+ epoch 0 classification metrics: 100%|██████████| 313/313 [09:03<00:00, 1.74s/it]
6
+ epoch=0 classification clean=0.7790 perturbed=0.3895 defocus_blur=0.4162 glass_blur=0.2478 motion_blur=0.5030 zoom_blur=0.3908
7
+ epoch 1/3: 100%|██████████| 1875/1875 [1:12:27<00:00, 2.32s/it, loss=55.9, lr=8.43e-6]
8
+ epoch 1 validation: 100%|██████████| 235/235 [04:37<00:00, 1.18s/it]
9
+ epoch 1 classification metrics: 100%|██████████| 313/313 [09:06<00:00, 1.75s/it]
10
+ epoch=1 classification clean=0.7524 perturbed=0.6436 defocus_blur=0.6716 glass_blur=0.6608 motion_blur=0.7098 zoom_blur=0.5322
11
+ Processing Files (2 / 2) : 100%|██████████| 1.22GB / 1.22GB, 225MB/s
12
+ New Data Upload : 100%|██████████| 1.22GB / 1.22GB, 225MB/s
13
+ ...ch_0001/pytorch_model.bin: 100%|██████████| 1.21GB / 1.21GB
14
+ ...001/visual_projection.bin: 100%|██████████| 3.15MB / 3.15MB
15
+ epoch 2/3: 100%|██████████| 1875/1875 [1:12:28<00:00, 2.32s/it, loss=33, lr=3.02e-6]
16
+ epoch 2 validation: 100%|██████████| 235/235 [04:38<00:00, 1.19s/it]
17
+ epoch 2 classification metrics: 100%|██████████| 313/313 [09:08<00:00, 1.75s/it]
18
+ epoch=2 classification clean=0.7520 perturbed=0.6605 defocus_blur=0.6882 glass_blur=0.6884 motion_blur=0.7268 zoom_blur=0.5386
19
+ Processing Files (2 / 2) : 100%|██████████| 1.22GB / 1.22GB, 121MB/s
20
+ New Data Upload : 100%|██████████| 1.22GB / 1.22GB, 121MB/s
21
+ ...ch_0002/pytorch_model.bin: 100%|██████████| 1.21GB / 1.21GB
22
+ ...002/visual_projection.bin: 100%|██████████| 3.15MB / 3.15MB
23
+ epoch 3/3: 100%|██████████| 1875/1875 [1:12:29<00:00, 2.32s/it, loss=24.4, lr=0]
24
+ epoch 3 validation: 100%|██████████| 235/235 [04:39<00:00, 1.19s/it]
25
+ epoch 3 classification metrics: 100%|██████████| 313/313 [09:08<00:00, 1.75s/it]
26
+ epoch=3 classification clean=0.7568 perturbed=0.6666 defocus_blur=0.6950 glass_blur=0.6914 motion_blur=0.7296 zoom_blur=0.5506
27
+ Processing Files (2 / 2) : 100%|██████████| 1.22GB / 1.22GB, 184MB/s
28
+ New Data Upload : 100%|██████████| 1.22GB / 1.22GB, 184MB/s
29
+ ...ch_0003/pytorch_model.bin: 100%|██████████| 1.21GB / 1.21GB
30
+ ...003/visual_projection.bin: 100%|██████████| 3.15MB / 3.15MB
31
+ Processing Files (9 / 9) : 100%|██████████| 4.87GB / 4.87GB, 1.49GB/s
32
+ New Data Upload : 100%|█████████▉| 3.59MB / 3.60MB, 1.50MB/s
33
+ ...002/visual_projection.bin: 100%|██████████| 3.15MB / 3.15MB
34
+ ...001/visual_projection.bin: 100%|██████████| 3.15MB / 3.15MB
35
+ ...003/visual_projection.bin: 100%|██████████| 3.15MB / 3.15MB
36
+ ...ch_0001/pytorch_model.bin: 100%|██████████| 1.21GB / 1.21GB
37
+ ...ch_0002/pytorch_model.bin: 100%|██████████| 1.21GB / 1.21GB
38
+ ...840/visual_projection.bin: 100%|██████████| 3.15MB / 3.15MB
39
+ ...ch_0003/pytorch_model.bin: 100%|██████████| 1.21GB / 1.21GB
40
+ ...-050840/pytorch_model.bin: 100%|██████████| 1.21GB / 1.21GB
41
+ ...wwcf5b/run-4qwwcf5b.wandb: 100%|██████████| 3.60MB / 3.60MB
wandb/run-20260821_051826-4qwwcf5b/files/requirements.txt ADDED
@@ -0,0 +1,277 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ setuptools==80.9.0
2
+ wheel==0.45.1
3
+ webencodings==0.5.1
4
+ pure_eval==0.2.3
5
+ ptyprocess==0.7.0
6
+ fastjsonschema==2.21.2
7
+ zipp==3.23.0
8
+ websocket-client==1.9.0
9
+ webcolors==24.11.1
10
+ wcwidth==0.2.14
11
+ urllib3==2.5.0
12
+ uri-template==1.3.0
13
+ tzdata==2025.2
14
+ typing_extensions==4.15.0
15
+ traitlets==5.14.3
16
+ tornado==6.5.2
17
+ tomli==2.3.0
18
+ tinycss2==1.4.0
19
+ soupsieve==2.8
20
+ sniffio==1.3.1
21
+ six==1.17.0
22
+ Send2Trash==1.8.3
23
+ rpds-py==0.27.1
24
+ rfc3986-validator==0.1.1
25
+ pyzmq==27.1.0
26
+ PyYAML==6.0.3
27
+ Pygments==2.19.2
28
+ pycparser==2.23
29
+ psutil==7.1.2
30
+ prometheus_client==0.23.1
31
+ platformdirs==4.4.0
32
+ pexpect==4.9.0
33
+ parso==0.8.5
34
+ pandocfilters==1.5.1
35
+ packaging==25.0
36
+ overrides==7.7.0
37
+ nest-asyncio==1.6.0
38
+ MarkupSafe==3.0.3
39
+ jupyterlab_pygments==0.3.0
40
+ jsonpointer==3.0.0
41
+ json5==0.12.1
42
+ idna==3.11
43
+ h11==0.16.0
44
+ fqdn==1.5.1
45
+ executing==2.2.1
46
+ defusedxml==0.7.1
47
+ decorator==5.2.1
48
+ debugpy==1.8.17
49
+ comm==0.2.3
50
+ charset-normalizer==3.4.4
51
+ certifi==2025.10.5
52
+ bleach==6.2.0
53
+ babel==2.17.0
54
+ attrs==25.4.0
55
+ asttokens==3.0.0
56
+ terminado==0.18.1
57
+ stack-data==0.6.3
58
+ rfc3987-syntax==1.1.0
59
+ rfc3339-validator==0.1.4
60
+ requests==2.32.5
61
+ referencing==0.36.2
62
+ python-json-logger==4.0.0
63
+ python-dateutil==2.9.0.post0
64
+ prompt_toolkit==3.0.52
65
+ mistune==3.1.4
66
+ matplotlib-inline==0.2.1
67
+ jupyter_core==5.8.1
68
+ Jinja2==3.1.6
69
+ jedi==0.19.2
70
+ importlib_metadata==8.7.0
71
+ httpcore==1.0.9
72
+ exceptiongroup==1.3.0
73
+ cffi==2.0.0
74
+ beautifulsoup4==4.14.2
75
+ async-lru==2.0.5
76
+ jupyter_server_terminals==0.5.3
77
+ jupyter_client==8.6.3
78
+ jsonschema-specifications==2025.9.1
79
+ ipython==8.18.1
80
+ arrow==1.4.0
81
+ argon2-cffi-bindings==25.1.0
82
+ anyio==4.11.0
83
+ jsonschema==4.25.1
84
+ isoduration==20.11.0
85
+ ipykernel==6.31.0
86
+ httpx==0.28.1
87
+ argon2-cffi==25.1.0
88
+ nbformat==5.10.4
89
+ nbclient==0.10.2
90
+ jupyter-events==0.12.0
91
+ nbconvert==7.16.6
92
+ jupyter_server==2.17.0
93
+ notebook_shim==0.2.4
94
+ jupyterlab_server==2.28.0
95
+ jupyter-lsp==2.3.0
96
+ jupyterlab==4.4.10
97
+ notebook==7.4.7
98
+ pytz==2025.2
99
+ nvidia-cusparselt-cu12==0.7.1
100
+ mpmath==1.3.0
101
+ tqdm==4.67.1
102
+ threadpoolctl==3.6.0
103
+ sympy==1.14.0
104
+ scipy==1.13.1
105
+ safetensors==0.6.2
106
+ regex==2025.10.23
107
+ nvidia-nvtx-cu12==12.8.90
108
+ nvidia-nvjitlink-cu12==12.8.93
109
+ nvidia-nccl-cu12==2.27.3
110
+ nvidia-curand-cu12==10.3.9.90
111
+ nvidia-cufile-cu12==1.13.1.3
112
+ nvidia-cuda-runtime-cu12==12.8.90
113
+ nvidia-cuda-nvrtc-cu12==12.8.93
114
+ nvidia-cuda-cupti-cu12==12.8.90
115
+ nvidia-cublas-cu12==12.8.4.1
116
+ networkx==3.2.1
117
+ joblib==1.5.2
118
+ hf-xet==1.2.0
119
+ fsspec==2025.9.0
120
+ filelock==3.19.1
121
+ triton==3.4.0
122
+ scikit-learn==1.6.1
123
+ pandas==2.3.3
124
+ nvidia-cusparse-cu12==12.5.8.93
125
+ nvidia-cufft-cu12==11.3.3.83
126
+ nvidia-cudnn-cu12==9.10.2.21
127
+ huggingface-hub==0.36.0
128
+ tokenizers==0.22.1
129
+ nvidia-cusolver-cu12==11.7.3.90
130
+ torch==2.8.0
131
+ pillow==11.3.0
132
+ pyparsing==3.2.5
133
+ kiwisolver==1.4.7
134
+ importlib_resources==6.5.2
135
+ fonttools==4.60.1
136
+ cycler==0.12.1
137
+ contourpy==1.3.0
138
+ matplotlib==3.9.4
139
+ xxhash==3.6.0
140
+ pyarrow==21.0.0
141
+ propcache==0.4.1
142
+ multidict==6.7.0
143
+ frozenlist==1.8.0
144
+ dill==0.4.0
145
+ async-timeout==5.0.1
146
+ aiohappyeyeballs==2.6.1
147
+ yarl==1.22.0
148
+ multiprocess==0.70.18
149
+ aiosignal==1.4.0
150
+ aiohttp==3.13.2
151
+ typing-inspection==0.4.2
152
+ smmap==5.0.2
153
+ sentry-sdk==2.48.0
154
+ pydantic_core==2.41.5
155
+ protobuf==6.33.6
156
+ eval_type_backport==0.3.1
157
+ click==8.1.8
158
+ annotated-types==0.7.0
159
+ pydantic==2.12.5
160
+ gitdb==4.0.12
161
+ GitPython==3.1.45
162
+ wandb==0.23.1
163
+ bitsandbytes==0.48.2
164
+ accelerate==1.10.1
165
+ peft==0.17.1
166
+ trl==0.24.0
167
+ wrapt==2.0.1
168
+ wasabi==1.1.3
169
+ typer-slim==0.21.1
170
+ spacy-loggers==1.0.5
171
+ spacy-legacy==3.0.12
172
+ murmurhash==1.0.15
173
+ cymem==2.0.13
174
+ catalogue==2.0.10
175
+ srsly==2.5.2
176
+ preshed==3.0.12
177
+ confection==0.1.5
178
+ thinc==8.3.9
179
+ typer==0.23.2
180
+ sentence-transformers==5.1.2
181
+ pathlib_abc==0.1.1
182
+ torchvision==0.23.0
183
+ smart-open==6.4.0
184
+ langcodes==3.5.1
185
+ pathy==0.11.0
186
+ cloudpathlib==0.16.0
187
+ spacy==3.8.11
188
+ weasel==0.4.3
189
+ py-cpuinfo==9.0.0
190
+ scikit-learn-extra==0.3.0
191
+ llvmlite==0.43.0
192
+ numba==0.60.0
193
+ pynndescent==0.6.0
194
+ umap-learn==0.5.11
195
+ tifffile==2024.8.30
196
+ lazy-loader==0.5
197
+ ImageIO==2.37.2
198
+ scikit-image==0.24.0
199
+ python-dotenv==1.2.1
200
+ dotenv==0.9.9
201
+ evaluate==0.4.6
202
+ en_core_web_md==3.8.0
203
+ blis==1.3.3
204
+ fastrlock==0.8.3
205
+ websockets==15.0.1
206
+ uvloop==0.22.1
207
+ uvicorn==0.39.0
208
+ shellingham==1.5.4
209
+ setproctitle==1.3.7
210
+ transformers==4.57.6
211
+ rignore==0.7.6
212
+ python-multipart==0.0.20
213
+ pycountry==24.6.1
214
+ pybase64==1.4.3
215
+ partial-json-parser==0.2.1.1.post7
216
+ outlines_core==0.2.11
217
+ numpy==2.0.2
218
+ ninja==1.13.0
219
+ msgspec==0.20.0
220
+ msgpack==1.1.2
221
+ mdurl==0.1.2
222
+ llguidance==0.7.30
223
+ jiter==0.14.0
224
+ lark==1.2.2
225
+ interegular==0.3.3
226
+ httptools==0.7.1
227
+ frozendict==2.4.7
228
+ fastar==0.10.0
229
+ einops==0.8.2
230
+ dnspython==2.7.0
231
+ distro==1.9.0
232
+ diskcache==5.6.3
233
+ cloudpickle==3.1.2
234
+ cbor2==5.9.0
235
+ cachetools==6.2.6
236
+ blake3==1.0.8
237
+ astor==0.8.1
238
+ annotated-doc==0.0.4
239
+ tiktoken==0.12.0
240
+ soxr==1.0.0
241
+ soundfile==0.13.1
242
+ opencv-python-headless==4.13.0.92
243
+ markdown-it-py==3.0.0
244
+ gguf==0.18.0
245
+ email-validator==2.3.0
246
+ depyf==0.19.0
247
+ cupy-cuda12x==13.6.0
248
+ watchfiles==1.1.1
249
+ starlette==0.49.3
250
+ rich==15.0.0
251
+ pydantic-settings==2.11.0
252
+ pydantic-extra-types==2.11.2
253
+ openai-harmony==0.0.8
254
+ lm-format-enforcer==0.11.3
255
+ xformers==0.0.32.post1
256
+ torchaudio==2.8.0
257
+ rich-toolkit==0.19.7
258
+ ray==2.51.2
259
+ prometheus-fastapi-instrumentator==7.1.0
260
+ openai==2.31.0
261
+ fastapi==0.128.8
262
+ xgrammar==0.1.25
263
+ mistral_common==1.8.5
264
+ fastapi-cloud-cli==0.12.0
265
+ fastapi-cli==0.0.22
266
+ compressed-tensors==0.11.0
267
+ vllm==0.11.0
268
+ en_core_web_sm==3.8.0
269
+ seaborn==0.13.2
270
+ transformers-stream-generator==0.0.5
271
+ uv==0.11.15
272
+ hf_transfer==0.1.9
273
+ kagglehub==0.3.13
274
+ datasets==4.5.0
275
+ pip==26.0.1
276
+ zstandard==0.25.0
277
+ sentencepiece==0.2.2
wandb/run-20260821_051826-4qwwcf5b/files/wandb-metadata.json ADDED
@@ -0,0 +1,150 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ {
2
+ "os": "Linux-4.18.0-553.147.1.el8_10.x86_64-x86_64-with-glibc2.28",
3
+ "python": "CPython 3.9.7",
4
+ "startedAt": "2026-08-21T09:18:26.294211Z",
5
+ "args": [
6
+ "--model-name",
7
+ "llava-hf/llava-1.5-7b-hf",
8
+ "--clip-model-name",
9
+ "openai/clip-vit-large-patch14-336",
10
+ "--original-path",
11
+ "/home/agarwali/orcd/datasets/imagenet/images_complete/ilsvrc/val",
12
+ "--perturbed-path",
13
+ "/home/agarwali/orcd/pool/ImageNet-C",
14
+ "--output-dir",
15
+ "/orcd/home/002/agarwali/projects/adv_ft_interp/finetuning_runs",
16
+ "--run-name",
17
+ "llava-1.5-7b-hf-r1-s3-patch_loss-h200-20260821-050840",
18
+ "--split-metadata-path",
19
+ "/orcd/home/002/agarwali/projects/adv_ft_interp/scripts/finetuning_source_splits.json",
20
+ "--rotation",
21
+ "R1",
22
+ "--severities",
23
+ "3",
24
+ "--train-frac",
25
+ "0.8",
26
+ "--val-frac",
27
+ "0.1",
28
+ "--seed",
29
+ "0",
30
+ "--num-epochs",
31
+ "3",
32
+ "--batch-size",
33
+ "64",
34
+ "--expected-train-examples",
35
+ "120000",
36
+ "--expected-steps-per-epoch",
37
+ "1875",
38
+ "--num-workers",
39
+ "8",
40
+ "--lr",
41
+ "1e-5",
42
+ "--weight-decay",
43
+ "0.01",
44
+ "--lr-scheduler",
45
+ "cosine",
46
+ "--warmup-ratio",
47
+ "0.1",
48
+ "--max-grad-norm",
49
+ "1.0",
50
+ "--torch-dtype",
51
+ "float16",
52
+ "--device",
53
+ "cuda",
54
+ "--wandb-project",
55
+ "llava-1-5-7b-hf-patch_loss-ft",
56
+ "--wandb-entity",
57
+ "adv-ft-interp",
58
+ "--wandb-run-name",
59
+ "llava-1.5-7b-hf-r1-s3-patch_loss-h200-20260821-050840",
60
+ "--wandb-mode",
61
+ "online",
62
+ "--wandb-log-every-steps",
63
+ "125",
64
+ "--push-to-hub",
65
+ "--repo-id",
66
+ "Adv-FT-Interp/llava-1.5-7b-hf-s3-patch_loss",
67
+ "--lambda-patch",
68
+ "0.1"
69
+ ],
70
+ "program": "/orcd/home/002/agarwali/projects/adv_ft_interp/scripts/finetuning_tests/finetuning_patch_loss.py",
71
+ "codePath": "scripts/finetuning_tests/finetuning_patch_loss.py",
72
+ "codePathLocal": "scripts/finetuning_tests/finetuning_patch_loss.py",
73
+ "git": {
74
+ "remote": "https://github.com/hannah-y-gao/adv_ft_interp.git",
75
+ "commit": "f6f529d1a7f71ed73fce602d0974ecaed4f17101"
76
+ },
77
+ "email": "agarwali@mit.edu",
78
+ "root": "/orcd/home/002/agarwali/projects/adv_ft_interp/finetuning_runs/llava-1.5-7b-hf-r1-s3-patch_loss-h200-20260821-050840",
79
+ "host": "node4900",
80
+ "executable": "/home/agarwali/anaconda3/envs/proj/bin/python",
81
+ "cpu_count": 120,
82
+ "cpu_count_logical": 240,
83
+ "gpu": "NVIDIA H200",
84
+ "gpu_count": 1,
85
+ "disk": {
86
+ "/": {
87
+ "total": "944364355584",
88
+ "used": "17049993216"
89
+ }
90
+ },
91
+ "memory": {
92
+ "total": "2163419840512"
93
+ },
94
+ "gpu_nvidia": [
95
+ {
96
+ "name": "NVIDIA H200",
97
+ "memoryTotal": "150754820096",
98
+ "cudaCores": 16896,
99
+ "architecture": "Hopper",
100
+ "uuid": "GPU-03e3ae62-87aa-dce4-ecc2-380b8497ccb2"
101
+ }
102
+ ],
103
+ "cudaVersion": "13.1",
104
+ "slurm": {
105
+ "cluster_name": "eofe7",
106
+ "conf": "/etc/slurm/slurm.conf",
107
+ "cpus_on_node": "8",
108
+ "cpus_per_task": "8",
109
+ "export_env": "ALL,FINETUNING_STRATEGY=patch_loss",
110
+ "get_user_env": "1",
111
+ "gpus_on_node": "1",
112
+ "gtids": "0",
113
+ "job_account": "mit_general",
114
+ "job_cpus_per_node": "8",
115
+ "job_end_time": "1787324917",
116
+ "job_gid": "100241592",
117
+ "job_gpus": "5",
118
+ "job_id": "20893027",
119
+ "job_name": "clip-ft-patch-loss",
120
+ "job_nodelist": "node4900",
121
+ "job_num_nodes": "1",
122
+ "job_partition": "mit_preemptable",
123
+ "job_qos": "normal",
124
+ "job_start_time": "1787303317",
125
+ "job_uid": "241592",
126
+ "job_user": "agarwali",
127
+ "jobid": "20893027",
128
+ "localid": "0",
129
+ "mem_per_node": "98304",
130
+ "nnodes": "1",
131
+ "nodeid": "0",
132
+ "nodelist": "node4900",
133
+ "nprocs": "1",
134
+ "ntasks": "1",
135
+ "ntasks_per_node": "1",
136
+ "oom_kill_step": "0",
137
+ "prio_process": "0",
138
+ "procid": "0",
139
+ "restart_count": "1",
140
+ "script_context": "prolog_task",
141
+ "submit_dir": "/orcd/home/002/agarwali/projects/adv_ft_interp",
142
+ "submit_host": "login006",
143
+ "task_pid": "769673",
144
+ "tasks_per_node": "1",
145
+ "topology_addr": "node4900",
146
+ "topology_addr_pattern": "node",
147
+ "tres_per_task": "cpu=8"
148
+ },
149
+ "writerId": "z9n3ltr0dbq5ggw93l1ck75zwyarlqcm"
150
+ }
wandb/run-20260821_051826-4qwwcf5b/logs/debug-core.log ADDED
@@ -0,0 +1,6 @@
 
 
 
 
 
 
 
1
+ {"time":"2026-08-21T05:18:27.349574231-04:00","level":"INFO","msg":"main: starting server","port-filename":"/tmp/tmpiq9ej01a/port-769791.txt","pid":769791,"log-level":0,"disable-analytics":false,"shutdown-on-parent-exit":false,"enable-dcgm-profiling":false}
2
+ {"time":"2026-08-21T05:18:27.352310377-04:00","level":"INFO","msg":"server: will exit if parent process dies","ppid":769791}
3
+ {"time":"2026-08-21T05:18:27.352299877-04:00","level":"INFO","msg":"server: accepting connections","addr":{"Name":"/tmp/wandb-769791-817478-1318887650/socket","Net":"unix"}}
4
+ {"time":"2026-08-21T05:18:27.538911431-04:00","level":"INFO","msg":"connection: ManageConnectionData: new connection created","id":"1(@)"}
5
+ {"time":"2026-08-21T05:18:27.602060209-04:00","level":"INFO","msg":"handleInformInit: received","streamId":"4qwwcf5b","id":"1(@)"}
6
+ {"time":"2026-08-21T05:18:28.099185511-04:00","level":"INFO","msg":"handleInformInit: stream started","streamId":"4qwwcf5b","id":"1(@)"}
wandb/run-20260821_051826-4qwwcf5b/logs/debug-internal.log ADDED
@@ -0,0 +1,6 @@
 
 
 
 
 
 
 
1
+ {"time":"2026-08-21T05:18:27.602575633-04:00","level":"INFO","msg":"stream: starting","core version":"0.23.1"}
2
+ {"time":"2026-08-21T05:18:28.098338792-04:00","level":"INFO","msg":"stream: created new stream","id":"4qwwcf5b"}
3
+ {"time":"2026-08-21T05:18:28.098401348-04:00","level":"INFO","msg":"handler: started","stream_id":"4qwwcf5b"}
4
+ {"time":"2026-08-21T05:18:28.099180044-04:00","level":"INFO","msg":"stream: started","id":"4qwwcf5b"}
5
+ {"time":"2026-08-21T05:18:28.099187832-04:00","level":"INFO","msg":"writer: started","stream_id":"4qwwcf5b"}
6
+ {"time":"2026-08-21T05:18:28.099198132-04:00","level":"INFO","msg":"sender: started","stream_id":"4qwwcf5b"}
wandb/run-20260821_051826-4qwwcf5b/logs/debug.log ADDED
@@ -0,0 +1,22 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ 2026-08-21 05:18:26,314 INFO MainThread:769791 [wandb_setup.py:_flush():80] Current SDK version is 0.23.1
2
+ 2026-08-21 05:18:26,314 INFO MainThread:769791 [wandb_setup.py:_flush():80] Configure stats pid to 769791
3
+ 2026-08-21 05:18:26,314 INFO MainThread:769791 [wandb_setup.py:_flush():80] Loading settings from /home/agarwali/.config/wandb/settings
4
+ 2026-08-21 05:18:26,314 INFO MainThread:769791 [wandb_setup.py:_flush():80] Loading settings from /orcd/home/002/agarwali/projects/adv_ft_interp/wandb/settings
5
+ 2026-08-21 05:18:26,314 INFO MainThread:769791 [wandb_setup.py:_flush():80] Loading settings from environment variables
6
+ 2026-08-21 05:18:26,314 INFO MainThread:769791 [wandb_init.py:setup_run_log_directory():714] Logging user logs to /orcd/home/002/agarwali/projects/adv_ft_interp/finetuning_runs/llava-1.5-7b-hf-r1-s3-patch_loss-h200-20260821-050840/wandb/run-20260821_051826-4qwwcf5b/logs/debug.log
7
+ 2026-08-21 05:18:26,314 INFO MainThread:769791 [wandb_init.py:setup_run_log_directory():715] Logging internal logs to /orcd/home/002/agarwali/projects/adv_ft_interp/finetuning_runs/llava-1.5-7b-hf-r1-s3-patch_loss-h200-20260821-050840/wandb/run-20260821_051826-4qwwcf5b/logs/debug-internal.log
8
+ 2026-08-21 05:18:26,314 INFO MainThread:769791 [wandb_init.py:init():841] calling init triggers
9
+ 2026-08-21 05:18:26,314 INFO MainThread:769791 [wandb_init.py:init():846] wandb.init called with sweep_config: {}
10
+ config: {'objective_smoke_test': False, 'model_name': 'llava-hf/llava-1.5-7b-hf', 'clip_model_name': 'openai/clip-vit-large-patch14-336', 'original_path': '/home/agarwali/orcd/datasets/imagenet/images_complete/ilsvrc/val', 'perturbed_path': '/home/agarwali/orcd/pool/ImageNet-C', 'output_dir': '/orcd/home/002/agarwali/projects/adv_ft_interp/finetuning_runs', 'run_name': 'llava-1.5-7b-hf-r1-s3-patch_loss-h200-20260821-050840', 'split_metadata_path': '/orcd/home/002/agarwali/projects/adv_ft_interp/scripts/finetuning_source_splits.json', 'corruptions': None, 'eval_corruptions': None, 'rotation': 'R1', 'severities': '3', 'max_pairs': None, 'train_frac': 0.8, 'val_frac': 0.1, 'seed': 0, 'num_epochs': 3, 'batch_size': 64, 'skip_initial_metrics': False, 'expected_train_examples': 120000, 'expected_steps_per_epoch': 1875, 'num_workers': 8, 'lr': 1e-05, 'weight_decay': 0.01, 'lr_scheduler': 'cosine', 'warmup_ratio': 0.1, 'max_grad_norm': 1.0, 'torch_dtype': 'float16', 'device': 'cuda', 'freeze_projection': False, 'wandb_project': 'llava-1-5-7b-hf-patch_loss-ft', 'wandb_entity': 'adv-ft-interp', 'wandb_mode': 'online', 'wandb_run_name': 'llava-1.5-7b-hf-r1-s3-patch_loss-h200-20260821-050840', 'wandb_log_every_steps': 125, 'push_to_hub': True, 'repo_id': 'Adv-FT-Interp/llava-1.5-7b-hf-s3-patch_loss', 'lambda_patch': 0.1, '_wandb': {}}
11
+ 2026-08-21 05:18:26,314 INFO MainThread:769791 [wandb_init.py:init():889] starting backend
12
+ 2026-08-21 05:18:27,536 INFO MainThread:769791 [wandb_init.py:init():892] sending inform_init request
13
+ 2026-08-21 05:18:27,591 INFO MainThread:769791 [wandb_init.py:init():900] backend started and connected
14
+ 2026-08-21 05:18:27,595 INFO MainThread:769791 [wandb_init.py:init():970] updated telemetry
15
+ 2026-08-21 05:18:27,675 INFO MainThread:769791 [wandb_init.py:init():994] communicating run to backend with 90.0 second timeout
16
+ 2026-08-21 05:18:28,457 INFO MainThread:769791 [wandb_init.py:init():1041] starting run threads in backend
17
+ 2026-08-21 05:18:29,373 INFO MainThread:769791 [wandb_run.py:_console_start():2521] atexit reg
18
+ 2026-08-21 05:18:29,373 INFO MainThread:769791 [wandb_run.py:_redirect():2369] redirect: wrap_raw
19
+ 2026-08-21 05:18:29,373 INFO MainThread:769791 [wandb_run.py:_redirect():2438] Wrapping output streams.
20
+ 2026-08-21 05:18:29,373 INFO MainThread:769791 [wandb_run.py:_redirect():2461] Redirects installed.
21
+ 2026-08-21 05:18:29,383 INFO MainThread:769791 [wandb_init.py:init():1081] run started, returning control to user process
22
+ 2026-08-21 05:18:29,383 INFO MainThread:769791 [wandb_run.py:_config_callback():1396] config_cb None None {'training_script': 'finetuning_patch_loss.py', 'experiment_name': 'final_layer_patch_loss', 'objective': {'formula': 'projected_global_l2 + lambda_patch * mean_normalized_final_patch_l2', 'lambda_patch': 0.1, 'patch_correspondence': 'same_position', 'cls_excluded': True}, 'projection_frozen': False}
wandb/run-20260821_051826-4qwwcf5b/run-4qwwcf5b.wandb ADDED
@@ -0,0 +1,3 @@
 
 
 
 
1
+ version https://git-lfs.github.com/spec/v1
2
+ oid sha256:d845211d9cc6ce1dd064f289882f1bf7fc4cfb0a938e301a03dac06bbb015f5d
3
+ size 3604480