hanngao commited on
Commit
2cbee80
·
verified ·
1 Parent(s): 831d9cd

Final faithful model: llava-1.5-7b-hf-r1-s3-faithful-ft-h200-20260818-185519

Browse files
.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-20260818_185616-oc7a60u3/run-oc7a60u3.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 +1,27 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
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
+ "resample": 3,
23
+ "rescale_factor": 0.00392156862745098,
24
+ "size": {
25
+ "shortest_edge": 336
26
+ }
27
+ }
checkpoints/epoch_0001/pytorch_model.bin ADDED
@@ -0,0 +1,3 @@
 
 
 
 
1
+ version https://git-lfs.github.com/spec/v1
2
+ oid sha256:344a06e9e960beafcb390dcaf4e2363204399258ce3e410985e9b57e77afefeb
3
+ size 1214155007
checkpoints/epoch_0001/visual_projection.bin ADDED
@@ -0,0 +1,3 @@
 
 
 
 
1
+ version https://git-lfs.github.com/spec/v1
2
+ oid sha256:8ff2092b534966251c8ab876354213309f41bb1e22b61a3960908129ed99a780
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 +1,27 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
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
+ "resample": 3,
23
+ "rescale_factor": 0.00392156862745098,
24
+ "size": {
25
+ "shortest_edge": 336
26
+ }
27
+ }
checkpoints/epoch_0002/pytorch_model.bin ADDED
@@ -0,0 +1,3 @@
 
 
 
 
1
+ version https://git-lfs.github.com/spec/v1
2
+ oid sha256:3a199f9511fbb6c08ed11711668846670224dfba80fe0206958c178cbb71014d
3
+ size 1214155007
checkpoints/epoch_0002/visual_projection.bin ADDED
@@ -0,0 +1,3 @@
 
 
 
 
1
+ version https://git-lfs.github.com/spec/v1
2
+ oid sha256:40aa3538bbb74e1b4f8045c797d9bc16b72dbafc4cab9962e419126060ad4915
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 +1,27 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
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
+ "resample": 3,
23
+ "rescale_factor": 0.00392156862745098,
24
+ "size": {
25
+ "shortest_edge": 336
26
+ }
27
+ }
checkpoints/epoch_0003/pytorch_model.bin ADDED
@@ -0,0 +1,3 @@
 
 
 
 
1
+ version https://git-lfs.github.com/spec/v1
2
+ oid sha256:9e68d7d5dc0d592e4fb408cdac846f868166c3a83518c04a42a31fe95e25c386
3
+ size 1214155007
checkpoints/epoch_0003/visual_projection.bin ADDED
@@ -0,0 +1,3 @@
 
 
 
 
1
+ version https://git-lfs.github.com/spec/v1
2
+ oid sha256:9065ee95b76cb2c327fe824238c4e992c1912e0afb154b7e231bf8bf1295b081
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 +1,27 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
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
+ "resample": 3,
23
+ "rescale_factor": 0.00392156862745098,
24
+ "size": {
25
+ "shortest_edge": 336
26
+ }
27
+ }
pytorch_model.bin ADDED
@@ -0,0 +1,3 @@
 
 
 
 
1
+ version https://git-lfs.github.com/spec/v1
2
+ oid sha256:9e68d7d5dc0d592e4fb408cdac846f868166c3a83518c04a42a31fe95e25c386
3
+ size 1214155007
splits.json ADDED
The diff for this file is too large to render. See raw diff
 
training_history.json ADDED
@@ -0,0 +1,101 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ {
2
+ "training_script": "finetuning_test_faithful.py",
3
+ "run_name": "llava-1.5-7b-hf-r1-s3-faithful-ft-h200-20260818-185519",
4
+ "rotation": "R1",
5
+ "model_name": "llava-hf/llava-1.5-7b-hf",
6
+ "clip_model_name": "openai/clip-vit-large-patch14-336",
7
+ "device": "cuda",
8
+ "objective": "corruption_fare",
9
+ "embedding_dim": 768,
10
+ "projection_trained": true,
11
+ "num_train_examples": 120000,
12
+ "num_val_examples": 15000,
13
+ "args": {
14
+ "smoke_test": false,
15
+ "data_smoke_test": false,
16
+ "model_name": "llava-hf/llava-1.5-7b-hf",
17
+ "clip_model_name": "openai/clip-vit-large-patch14-336",
18
+ "original_path": "/home/hanngao/orcd/pool/ImageNet",
19
+ "perturbed_path": "/home/hanngao/orcd/pool/ImageNet-C",
20
+ "output_dir": "/orcd/home/002/hanngao/adv_ft_interp/finetuning_runs",
21
+ "run_name": "llava-1.5-7b-hf-r1-s3-faithful-ft-h200-20260818-185519",
22
+ "split_metadata_path": "/orcd/home/002/hanngao/adv_ft_interp/finetuning_runs/llava-1.5-7b-hf-r1-s3-faithful-ft-h200-20260818-185519/splits.json",
23
+ "corruptions": null,
24
+ "rotation": "R1",
25
+ "severities": "3",
26
+ "max_pairs": null,
27
+ "train_frac": 0.8,
28
+ "val_frac": 0.1,
29
+ "seed": 0,
30
+ "num_epochs": 3,
31
+ "batch_size": 64,
32
+ "num_workers": 8,
33
+ "lr": 1e-05,
34
+ "weight_decay": 0.0,
35
+ "max_grad_norm": 1.0,
36
+ "torch_dtype": "float16",
37
+ "device": "cuda",
38
+ "wandb_project": "llava-1.5-7b-hf-faithful-ft",
39
+ "wandb_entity": "adv-ft-interp",
40
+ "wandb_mode": "online",
41
+ "wandb_run_name": "llava-1.5-7b-hf-r1-s3-faithful-ft-h200-20260818-185519",
42
+ "wandb_log_every_steps": 125,
43
+ "push_to_hub": true,
44
+ "repo_id": "Adv-FT-Interp/llava-1.5-7b-hf-r1-s3-faithful-ft"
45
+ },
46
+ "checkpoint": {
47
+ "format": "llava_vision_tower_plus_faithful_clip_projection",
48
+ "vision_checkpoint": "pytorch_model.bin",
49
+ "projection_checkpoint": "visual_projection.bin",
50
+ "clip_model_name": "openai/clip-vit-large-patch14-336",
51
+ "projection_is_used_by_llava_inference": false,
52
+ "vision_tower_is_loadable_by_existing_inference_loader": true
53
+ },
54
+ "history": {
55
+ "epoch": [
56
+ 1,
57
+ 2,
58
+ 3
59
+ ],
60
+ "train": [
61
+ {
62
+ "loss": 52.27631518554688,
63
+ "embedding_l2": 52.27631518554688,
64
+ "embedding_cosine_similarity": 0.9183515048662821,
65
+ "examples": 120000
66
+ },
67
+ {
68
+ "loss": 34.65345450744629,
69
+ "embedding_l2": 34.65345450744629,
70
+ "embedding_cosine_similarity": 0.9465240153948467,
71
+ "examples": 120000
72
+ },
73
+ {
74
+ "loss": 26.34983610636393,
75
+ "embedding_l2": 26.34983610636393,
76
+ "embedding_cosine_similarity": 0.959579998811086,
77
+ "examples": 120000
78
+ }
79
+ ],
80
+ "val": [
81
+ {
82
+ "loss": 44.25731761271159,
83
+ "embedding_l2": 44.257317610677084,
84
+ "embedding_cosine_similarity": 0.9310897902170817,
85
+ "examples": 15000
86
+ },
87
+ {
88
+ "loss": 36.40801206461589,
89
+ "embedding_l2": 36.40801206461589,
90
+ "embedding_cosine_similarity": 0.9436133759816487,
91
+ "examples": 15000
92
+ },
93
+ {
94
+ "loss": 31.37981117451986,
95
+ "embedding_l2": 31.379811173502603,
96
+ "embedding_cosine_similarity": 0.9515272616068522,
97
+ "examples": 15000
98
+ }
99
+ ]
100
+ }
101
+ }
visual_projection.bin ADDED
@@ -0,0 +1,3 @@
 
 
 
 
1
+ version https://git-lfs.github.com/spec/v1
2
+ oid sha256:9065ee95b76cb2c327fe824238c4e992c1912e0afb154b7e231bf8bf1295b081
3
+ size 3147503
wandb/debug-internal.log ADDED
The diff for this file is too large to render. See raw diff
 
wandb/debug.log ADDED
@@ -0,0 +1,20 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ 2026-08-18 18:56:16,799 INFO MainThread:409376 [wandb_setup.py:_flush():81] Current SDK version is 0.28.0
2
+ 2026-08-18 18:56:16,799 INFO MainThread:409376 [wandb_setup.py:_flush():81] Configure stats pid to 409376
3
+ 2026-08-18 18:56:16,799 INFO MainThread:409376 [wandb_setup.py:_flush():81] Loading settings from environment variables
4
+ 2026-08-18 18:56:16,799 INFO MainThread:409376 [wandb_init.py:setup_run_log_directory():725] Logging user logs to /orcd/home/002/hanngao/adv_ft_interp/finetuning_runs/llava-1.5-7b-hf-r1-s3-faithful-ft-h200-20260818-185519/wandb/run-20260818_185616-oc7a60u3/logs/debug.log
5
+ 2026-08-18 18:56:16,799 INFO MainThread:409376 [wandb_init.py:setup_run_log_directory():726] Logging internal logs to /orcd/home/002/hanngao/adv_ft_interp/finetuning_runs/llava-1.5-7b-hf-r1-s3-faithful-ft-h200-20260818-185519/wandb/run-20260818_185616-oc7a60u3/logs/debug-internal.log
6
+ 2026-08-18 18:56:16,799 INFO MainThread:409376 [wandb_init.py:init():768] calling init triggers
7
+ 2026-08-18 18:56:16,799 INFO MainThread:409376 [wandb_init.py:init():773] wandb.init called with sweep_config: {}
8
+ config: {'smoke_test': False, 'data_smoke_test': False, 'model_name': 'llava-hf/llava-1.5-7b-hf', 'clip_model_name': 'openai/clip-vit-large-patch14-336', 'original_path': '/home/hanngao/orcd/pool/ImageNet', 'perturbed_path': '/home/hanngao/orcd/pool/ImageNet-C', 'output_dir': '/orcd/home/002/hanngao/adv_ft_interp/finetuning_runs', 'run_name': 'llava-1.5-7b-hf-r1-s3-faithful-ft-h200-20260818-185519', 'split_metadata_path': '/orcd/home/002/hanngao/adv_ft_interp/finetuning_runs/llava-1.5-7b-hf-r1-s3-faithful-ft-h200-20260818-185519/splits.json', '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, 'num_workers': 8, 'lr': 1e-05, 'weight_decay': 0.0, 'max_grad_norm': 1.0, 'torch_dtype': 'float16', 'device': 'cuda', 'wandb_project': 'llava-1.5-7b-hf-faithful-ft', 'wandb_entity': 'adv-ft-interp', 'wandb_mode': 'online', 'wandb_run_name': 'llava-1.5-7b-hf-r1-s3-faithful-ft-h200-20260818-185519', 'wandb_log_every_steps': 125, 'push_to_hub': True, 'repo_id': 'Adv-FT-Interp/llava-1.5-7b-hf-r1-s3-faithful-ft', '_wandb': {}}
9
+ 2026-08-18 18:56:16,799 INFO MainThread:409376 [wandb_init.py:init():816] starting backend
10
+ 2026-08-18 18:56:16,799 INFO MainThread:409376 [wandb_init.py:init():831] sending inform_init request
11
+ 2026-08-18 18:56:17,149 INFO MainThread:409376 [wandb_init.py:init():836] backend started and connected
12
+ 2026-08-18 18:56:17,151 INFO MainThread:409376 [wandb_init.py:init():906] updated telemetry
13
+ 2026-08-18 18:56:17,172 INFO MainThread:409376 [wandb_init.py:init():929] communicating run to backend with 90.0 second timeout
14
+ 2026-08-18 18:56:17,467 INFO MainThread:409376 [wandb_init.py:init():974] starting run threads in backend
15
+ 2026-08-18 18:56:17,956 INFO MainThread:409376 [wandb_run.py:_console_start():2523] atexit reg
16
+ 2026-08-18 18:56:17,957 INFO MainThread:409376 [wandb_run.py:_redirect():2373] redirect: wrap_raw
17
+ 2026-08-18 18:56:17,957 INFO MainThread:409376 [wandb_run.py:_redirect():2442] Wrapping output streams.
18
+ 2026-08-18 18:56:17,957 INFO MainThread:409376 [wandb_run.py:_redirect():2465] Redirects installed.
19
+ 2026-08-18 18:56:17,960 INFO MainThread:409376 [wandb_init.py:init():1012] run started, returning control to user process
20
+ 2026-08-18 18:56:17,961 INFO MainThread:409376 [wandb_run.py:_config_callback():1346] config_cb None None {'training_script': 'finetuning_test_faithful.py', 'objective': 'corruption_fare', 'embedding_dim': 768, 'projection_trained': True, 'run_name': 'llava-1.5-7b-hf-r1-s3-faithful-ft-h200-20260818-185519'}
wandb/run-20260818_185616-oc7a60u3/files/output.log ADDED
@@ -0,0 +1,15 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ epoch 1 validation: 100%|██████████| 235/235 [04:56<00:00, 1.26s/it]
2
+ Processing Files (2 / 2) : 100%|██████████| 1.22GB / 1.22GB, 53.6MB/s
3
+ New Data Upload : 100%|██████████| 1.22GB / 1.22GB, 53.6MB/s
4
+ ...ch_0001/pytorch_model.bin: 100%|██████████| 1.21GB / 1.21GB
5
+ ...001/visual_projection.bin: 100%|██████████| 3.15MB / 3.15MB
6
+ epoch 2 validation: 100%|██████████| 235/235 [04:56<00:00, 1.26s/it]
7
+ Processing Files (2 / 2) : 100%|██████████| 1.22GB / 1.22GB, 160MB/s
8
+ New Data Upload : 100%|██████████| 1.22GB / 1.22GB, 160MB/s
9
+ ...ch_0002/pytorch_model.bin: 100%|██████████| 1.21GB / 1.21GB
10
+ ...002/visual_projection.bin: 100%|██████████| 3.15MB / 3.15MB
11
+ epoch 3 validation: 100%|██████████| 235/235 [04:54<00:00, 1.25s/it]
12
+ Processing Files (2 / 2) : 100%|██████████| 1.22GB / 1.22GB, 179MB/s
13
+ New Data Upload : 100%|██████████| 1.22GB / 1.22GB, 179MB/s
14
+ ...ch_0003/pytorch_model.bin: 100%|██████████| 1.21GB / 1.21GB
15
+ ...003/visual_projection.bin: 100%|██████████| 3.15MB / 3.15MB
wandb/run-20260818_185616-oc7a60u3/files/requirements.txt ADDED
@@ -0,0 +1,261 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ sentencepiece==0.1.99
2
+ pydub==0.25.1
3
+ py-cpuinfo==9.0.0
4
+ nvidia-ml-py==13.580.82
5
+ hjson==3.1.0
6
+ websockets==11.0.3
7
+ tomlkit==0.12.0
8
+ svgwrite==1.4.3
9
+ smmap==5.0.2
10
+ shortuuid==1.0.13
11
+ sentry-sdk==2.46.0
12
+ semantic-version==2.10.0
13
+ ruff==0.14.7
14
+ python-multipart==0.0.20
15
+ jiter==0.13.0
16
+ typing_extensions==4.15.0
17
+ einops==0.8.2
18
+ timm==1.0.26
19
+ openai==2.29.0
20
+ ninja==1.13.0
21
+ narwhals==2.12.0
22
+ markdown2==2.5.4
23
+ latex2mathml==3.78.1
24
+ ffmpy==1.0.0
25
+ aiofiles==23.2.1
26
+ wavedrom==2.0.3.post3
27
+ namex==0.1.0
28
+ gitdb==4.0.12
29
+ einops-exts==0.0.4
30
+ GitPython==3.1.45
31
+ hf-xet==1.4.3
32
+ prismatic==0.0.2
33
+ gradio_client==0.8.1
34
+ deepspeed==0.12.6
35
+ altair==5.5.0
36
+ peft==0.18.0
37
+ gradio==4.16.0
38
+ llava==1.2.2.post1
39
+ optree==0.19.0
40
+ accelerate==1.13.0
41
+ tokenizers==0.22.2
42
+ huggingface_hub==1.10.1
43
+ libclang==18.1.1
44
+ flatbuffers==25.12.19
45
+ termcolor==3.3.0
46
+ h5py==3.14.0
47
+ google-pasta==0.2.0
48
+ gast==0.7.0
49
+ astunparse==1.6.3
50
+ absl-py==2.4.0
51
+ keras==3.12.1
52
+ tensorflow==2.21.0
53
+ regex==2026.4.4
54
+ packaging==26.0
55
+ wheel==0.47.0
56
+ pip==26.1.1
57
+ z3-solver==4.15.4.0
58
+ torchaudio==2.11.0
59
+ supervisor==4.3.0
60
+ nvidia-cusparselt-cu13==0.8.0
61
+ mpmath==1.3.0
62
+ cuda-toolkit==13.0.2
63
+ uvloop==0.22.1
64
+ urllib3==2.7.0
65
+ typing-inspection==0.4.2
66
+ triton==3.6.0
67
+ tqdm==4.68.3
68
+ tomli==2.4.1
69
+ tokenspeed-triton==3.7.10.post20260531
70
+ tabulate==0.10.0
71
+ sympy==1.14.0
72
+ sniffio==1.3.1
73
+ shellingham==1.5.4
74
+ setproctitle==1.3.7
75
+ setuptools==81.0.0
76
+ safetensors==0.8.0
77
+ rpds-py==0.30.0
78
+ rignore==0.7.6
79
+ pyzmq==27.1.0
80
+ PyYAML==6.0.3
81
+ python-json-logger==4.1.0
82
+ python-dotenv==1.2.2
83
+ PyJWT==2.13.0
84
+ Pygments==2.20.0
85
+ pyelftools==0.33
86
+ pydantic_core==2.41.5
87
+ pycparser==3.0
88
+ pycountry==26.2.16
89
+ pybase64==1.4.3
90
+ psutil==7.2.2
91
+ protobuf==6.33.6
92
+ propcache==0.5.2
93
+ prometheus_client==0.25.0
94
+ pillow==12.2.0
95
+ partial-json-parser==0.2.1.1.post7
96
+ outlines_core==0.2.14
97
+ opentelemetry-api==1.42.1
98
+ nvidia-nvvm==13.2.78
99
+ nvidia-nvtx==13.0.85
100
+ nvidia-nvshmem-cu13==3.4.5
101
+ nvidia-nvjitlink==13.0.88
102
+ nvidia-nccl-cu13==2.28.9
103
+ nvidia-curand==10.4.0.35
104
+ nvidia-cufile==1.15.1.6
105
+ nvidia-cudnn-frontend==1.25.0
106
+ nvidia-cuda-runtime==13.0.96
107
+ nvidia-cuda-nvrtc==13.0.88
108
+ nvidia-cuda-cupti==13.0.85
109
+ nvidia-cuda-crt==13.3.33
110
+ nvidia-cuda-cccl==13.3.3.3.1
111
+ nvidia-cublas==13.1.0.3
112
+ numpy==2.2.6
113
+ numpy==2.2.5
114
+ networkx==3.4.2
115
+ multidict==6.7.1
116
+ msgspec==0.21.1
117
+ mdurl==0.1.2
118
+ MarkupSafe==3.0.3
119
+ loguru==0.7.3
120
+ llvmlite==0.47.0
121
+ llguidance==1.7.6
122
+ lark==1.2.2
123
+ jmespath==1.1.0
124
+ interegular==0.3.3
125
+ ijson==3.5.0
126
+ idna==3.18
127
+ httpx-sse==0.4.3
128
+ httptools==0.8.0
129
+ h11==0.16.0
130
+ grpcio==1.81.1
131
+ frozenlist==1.8.0
132
+ flashinfer-cubin==0.6.12
133
+ filelock==3.29.4
134
+ fastar==0.11.0
135
+ exceptiongroup==1.3.1
136
+ docstring_parser==0.18.0
137
+ dnspython==2.8.0
138
+ distro==1.9.0
139
+ diskcache==5.6.3
140
+ dill==0.4.1
141
+ detect-installer==0.1.0
142
+ cuda-tile==1.3.0
143
+ cuda-pathfinder==1.5.5
144
+ cloudpickle==3.1.2
145
+ click==8.4.1
146
+ charset-normalizer==3.4.7
147
+ certifi==2026.6.17
148
+ cbor2==6.1.2
149
+ cachetools==7.1.4
150
+ blake3==1.0.8
151
+ backports.strenum==1.3.1
152
+ attrs==26.1.0
153
+ async-timeout==5.0.1
154
+ astor==0.8.1
155
+ apache-tvm-ffi==0.1.9
156
+ annotated-types==0.7.0
157
+ annotated-doc==0.0.4
158
+ aiohappyeyeballs==2.6.2
159
+ yarl==1.24.2
160
+ uvicorn==0.49.0
161
+ requests==2.34.2
162
+ referencing==0.37.0
163
+ opentelemetry-semantic-conventions==0.63b1
164
+ opentelemetry-proto==1.42.1
165
+ opencv-python-headless==4.13.0.92
166
+ nvidia-cusparse==12.6.3.3
167
+ nvidia-cufft==12.0.0.61
168
+ nvidia-cudnn-cu13==9.19.0.56
169
+ nvidia-cuda-nvcc==13.2.78
170
+ numba==0.65.0
171
+ ml_dtypes==0.5.4
172
+ markdown-it-py==4.2.0
173
+ Jinja2==3.1.6
174
+ httpcore==1.0.9
175
+ googleapis-common-protos==1.75.0
176
+ email-validator==2.3.0
177
+ depyf==0.20.0
178
+ cuda-core==1.0.1
179
+ cuda-bindings==13.3.1
180
+ cffi==2.0.0
181
+ anyio==4.14.0
182
+ aiosignal==1.4.0
183
+ watchfiles==1.2.0
184
+ tiktoken==0.13.0
185
+ starlette==1.3.1
186
+ rich==15.0.0
187
+ pydantic-settings==2.14.2
188
+ pydantic-extra-types==2.11.1
189
+ opentelemetry-sdk==1.42.1
190
+ opentelemetry-exporter-otlp-proto-common==1.42.1
191
+ openai-harmony==0.0.8
192
+ nvidia-cusolver==12.0.4.66
193
+ nvidia-cuda-tileiras==13.2.78
194
+ lm-format-enforcer==0.11.3
195
+ jsonschema-specifications==2025.9.1
196
+ httpx==0.28.1
197
+ gguf==0.19.0
198
+ cuda-python==13.3.1
199
+ cryptography==49.0.0
200
+ aiohttp==3.14.1
201
+ typer==0.26.7
202
+ sse-starlette==3.4.5
203
+ rich-toolkit==0.20.1
204
+ prometheus-fastapi-instrumentator==8.0.0
205
+ opentelemetry-semantic-conventions-ai==0.5.1
206
+ opentelemetry-exporter-otlp-proto-http==1.42.1
207
+ opentelemetry-exporter-otlp-proto-grpc==1.42.1
208
+ nvidia-cutlass-dsl-libs-base==4.5.2
209
+ jsonschema==4.26.0
210
+ fastapi==0.136.3
211
+ anthropic==0.111.0
212
+ torch==2.11.0
213
+ opentelemetry-exporter-otlp==1.42.1
214
+ nvidia-cutlass-dsl-libs-cu13==4.5.2
215
+ nvidia-cutlass-dsl==4.5.2
216
+ model-hosting-container-standards==0.1.16
217
+ mistral_common==1.11.3
218
+ mcp==1.28.0
219
+ fastsafetensors==0.3.2
220
+ fastapi-cloud-cli==0.20.0
221
+ fastapi-cli==0.0.27
222
+ torchvision==0.26.0
223
+ torch_c_dlpack_ext==0.1.5
224
+ tokenspeed-mla==0.1.2
225
+ humming-kernels==0.1.4
226
+ flashinfer-python==0.6.12
227
+ transformers==5.12.1
228
+ tilelang==0.1.9
229
+ quack-kernels==0.5.0
230
+ xgrammar==0.2.2
231
+ compressed-tensors==0.17.0
232
+ vllm==0.23.0
233
+ tzdata==2026.2
234
+ pytz==2026.2
235
+ six==1.17.0
236
+ python-dateutil==2.9.0.post0
237
+ Bottleneck==1.4.2
238
+ numexpr==2.14.1
239
+ pandas==2.3.3
240
+ xxhash==3.7.0
241
+ pyarrow==24.0.0
242
+ multiprocess==0.70.19
243
+ datasets==5.0.0
244
+ fsspec==2026.4.0
245
+ platformdirs==4.10.0
246
+ wandb==0.28.0
247
+ tifffile==2025.5.10
248
+ scipy==1.15.3
249
+ lazy-loader==0.5
250
+ ImageIO==2.37.3
251
+ scikit-image==0.25.2
252
+ Wand==0.7.2
253
+ pyparsing==3.3.2
254
+ kiwisolver==1.5.0
255
+ fonttools==4.63.0
256
+ cycler==0.12.1
257
+ contourpy==1.3.2
258
+ matplotlib==3.10.9
259
+ jsonpath-python==1.1.6
260
+ openrouter==1.1.2
261
+ pydantic==2.12.5
wandb/run-20260818_185616-oc7a60u3/files/wandb-metadata.json ADDED
@@ -0,0 +1,137 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ {
2
+ "os": "Linux-4.18.0-553.147.1.el8_10.x86_64-x86_64-with-glibc2.28",
3
+ "python": "CPython 3.10.20",
4
+ "startedAt": "2026-08-18T22:56:16.782656Z",
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/hanngao/orcd/pool/ImageNet",
12
+ "--perturbed-path",
13
+ "/home/hanngao/orcd/pool/ImageNet-C",
14
+ "--output-dir",
15
+ "/orcd/home/002/hanngao/adv_ft_interp/finetuning_runs",
16
+ "--run-name",
17
+ "llava-1.5-7b-hf-r1-s3-faithful-ft-h200-20260818-185519",
18
+ "--split-metadata-path",
19
+ "/orcd/home/002/hanngao/adv_ft_interp/finetuning_runs/llava-1.5-7b-hf-r1-s3-faithful-ft-h200-20260818-185519/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
+ "--num-workers",
35
+ "8",
36
+ "--lr",
37
+ "1e-5",
38
+ "--weight-decay",
39
+ "0.0",
40
+ "--max-grad-norm",
41
+ "1.0",
42
+ "--torch-dtype",
43
+ "float16",
44
+ "--device",
45
+ "cuda",
46
+ "--wandb-project",
47
+ "llava-1.5-7b-hf-faithful-ft",
48
+ "--wandb-entity",
49
+ "adv-ft-interp",
50
+ "--wandb-run-name",
51
+ "llava-1.5-7b-hf-r1-s3-faithful-ft-h200-20260818-185519",
52
+ "--wandb-mode",
53
+ "online",
54
+ "--wandb-log-every-steps",
55
+ "125",
56
+ "--push-to-hub",
57
+ "--repo-id",
58
+ "Adv-FT-Interp/llava-1.5-7b-hf-r1-s3-faithful-ft"
59
+ ],
60
+ "program": "/orcd/home/002/hanngao/adv_ft_interp/scripts/finetuning_test_faithful.py",
61
+ "codePath": "scripts/finetuning_test_faithful.py",
62
+ "codePathLocal": "scripts/finetuning_test_faithful.py",
63
+ "git": {
64
+ "remote": "git@github.com:hannah-y-gao/adv_ft_interp.git",
65
+ "commit": "6b5271f485a5a1d8470d6580d1a6e717f7a1e792"
66
+ },
67
+ "email": "hann.yun.gao@gmail.com",
68
+ "root": "/orcd/home/002/hanngao/adv_ft_interp/finetuning_runs/llava-1.5-7b-hf-r1-s3-faithful-ft-h200-20260818-185519",
69
+ "host": "node4002",
70
+ "executable": "/home/hanngao/.conda/envs/vllm_env/bin/python",
71
+ "cpu_count": 96,
72
+ "cpu_count_logical": 192,
73
+ "gpu": "NVIDIA H200 NVL",
74
+ "gpu_count": 1,
75
+ "disk": {
76
+ "/": {
77
+ "total": "1885560414208",
78
+ "used": "32551866368"
79
+ }
80
+ },
81
+ "memory": {
82
+ "total": "1621773463552"
83
+ },
84
+ "gpu_nvidia": [
85
+ {
86
+ "name": "NVIDIA H200 NVL",
87
+ "memoryTotal": "150754820096",
88
+ "cudaCores": 16896,
89
+ "architecture": "Hopper",
90
+ "uuid": "GPU-4a26e7fe-ec3f-f565-3510-398017f05a22"
91
+ }
92
+ ],
93
+ "cudaVersion": "13.1",
94
+ "slurm": {
95
+ "cluster_name": "eofe7",
96
+ "conf": "/etc/slurm/slurm.conf",
97
+ "cpus_on_node": "8",
98
+ "cpus_per_task": "8",
99
+ "gpus_on_node": "1",
100
+ "gtids": "0",
101
+ "job_account": "mit_general",
102
+ "job_cpus_per_node": "8",
103
+ "job_end_time": "1787122519",
104
+ "job_gid": "241272",
105
+ "job_gpus": "0",
106
+ "job_id": "20713309",
107
+ "job_name": "clip-ft-faithful-h200",
108
+ "job_nodelist": "node4002",
109
+ "job_num_nodes": "1",
110
+ "job_partition": "mit_preemptable",
111
+ "job_qos": "normal",
112
+ "job_start_time": "1787093719",
113
+ "job_uid": "241272",
114
+ "job_user": "hanngao",
115
+ "jobid": "20713309",
116
+ "localid": "0",
117
+ "mem_per_node": "98304",
118
+ "nnodes": "1",
119
+ "nodeid": "0",
120
+ "nodelist": "node4002",
121
+ "nprocs": "1",
122
+ "ntasks": "1",
123
+ "ntasks_per_node": "1",
124
+ "oom_kill_step": "0",
125
+ "prio_process": "0",
126
+ "procid": "0",
127
+ "script_context": "prolog_task",
128
+ "submit_dir": "/orcd/home/002/hanngao/adv_ft_interp",
129
+ "submit_host": "node3003",
130
+ "task_pid": "409294",
131
+ "tasks_per_node": "1",
132
+ "topology_addr": "node4002",
133
+ "topology_addr_pattern": "node",
134
+ "tres_per_task": "cpu=8"
135
+ },
136
+ "writerId": "nqy7rmccexz6xh88zxyicjkej2v0ursw"
137
+ }
wandb/run-20260818_185616-oc7a60u3/logs/debug-core.log ADDED
@@ -0,0 +1,7 @@
 
 
 
 
 
 
 
 
1
+ {"time":"2026-08-18T18:56:16.43365495-04:00","level":"INFO","msg":"main: starting server","port-filename":"/tmp/tmp80sb01bj/port-409376.txt","pid":409376,"detached":false,"idle-timeout":600000000000,"log-level":0,"disable-analytics":false,"shutdown-on-parent-exit":false,"enable-dcgm-profiling":false}
2
+ {"time":"2026-08-18T18:56:16.436198183-04:00","level":"INFO","msg":"server: will exit if parent process dies","ppid":409376}
3
+ {"time":"2026-08-18T18:56:16.436204082-04:00","level":"INFO","msg":"server: accepting connections","addr":{"Name":"/tmp/wandb-409376-409641-3219226532/socket","Net":"unix"}}
4
+ {"time":"2026-08-18T18:56:16.470257764-04:00","level":"INFO","msg":"connection: ManageConnectionData: new connection created","id":"1(@)"}
5
+ {"time":"2026-08-18T18:56:16.800009077-04:00","level":"INFO","msg":"handleInformInit: received","streamId":"oc7a60u3","id":"1(@)"}
6
+ {"time":"2026-08-18T18:56:17.149555179-04:00","level":"INFO","msg":"handleInformInit: stream started","streamId":"oc7a60u3","id":"1(@)"}
7
+ {"time":"2026-08-18T18:56:22.964052032-04:00","level":"INFO","msg":"connection: cancelling request","id":"1(@)","requestId":"4xjba8wrk8qf"}
wandb/run-20260818_185616-oc7a60u3/logs/debug-internal.log ADDED
The diff for this file is too large to render. See raw diff
 
wandb/run-20260818_185616-oc7a60u3/logs/debug.log ADDED
@@ -0,0 +1,20 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ 2026-08-18 18:56:16,799 INFO MainThread:409376 [wandb_setup.py:_flush():81] Current SDK version is 0.28.0
2
+ 2026-08-18 18:56:16,799 INFO MainThread:409376 [wandb_setup.py:_flush():81] Configure stats pid to 409376
3
+ 2026-08-18 18:56:16,799 INFO MainThread:409376 [wandb_setup.py:_flush():81] Loading settings from environment variables
4
+ 2026-08-18 18:56:16,799 INFO MainThread:409376 [wandb_init.py:setup_run_log_directory():725] Logging user logs to /orcd/home/002/hanngao/adv_ft_interp/finetuning_runs/llava-1.5-7b-hf-r1-s3-faithful-ft-h200-20260818-185519/wandb/run-20260818_185616-oc7a60u3/logs/debug.log
5
+ 2026-08-18 18:56:16,799 INFO MainThread:409376 [wandb_init.py:setup_run_log_directory():726] Logging internal logs to /orcd/home/002/hanngao/adv_ft_interp/finetuning_runs/llava-1.5-7b-hf-r1-s3-faithful-ft-h200-20260818-185519/wandb/run-20260818_185616-oc7a60u3/logs/debug-internal.log
6
+ 2026-08-18 18:56:16,799 INFO MainThread:409376 [wandb_init.py:init():768] calling init triggers
7
+ 2026-08-18 18:56:16,799 INFO MainThread:409376 [wandb_init.py:init():773] wandb.init called with sweep_config: {}
8
+ config: {'smoke_test': False, 'data_smoke_test': False, 'model_name': 'llava-hf/llava-1.5-7b-hf', 'clip_model_name': 'openai/clip-vit-large-patch14-336', 'original_path': '/home/hanngao/orcd/pool/ImageNet', 'perturbed_path': '/home/hanngao/orcd/pool/ImageNet-C', 'output_dir': '/orcd/home/002/hanngao/adv_ft_interp/finetuning_runs', 'run_name': 'llava-1.5-7b-hf-r1-s3-faithful-ft-h200-20260818-185519', 'split_metadata_path': '/orcd/home/002/hanngao/adv_ft_interp/finetuning_runs/llava-1.5-7b-hf-r1-s3-faithful-ft-h200-20260818-185519/splits.json', '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, 'num_workers': 8, 'lr': 1e-05, 'weight_decay': 0.0, 'max_grad_norm': 1.0, 'torch_dtype': 'float16', 'device': 'cuda', 'wandb_project': 'llava-1.5-7b-hf-faithful-ft', 'wandb_entity': 'adv-ft-interp', 'wandb_mode': 'online', 'wandb_run_name': 'llava-1.5-7b-hf-r1-s3-faithful-ft-h200-20260818-185519', 'wandb_log_every_steps': 125, 'push_to_hub': True, 'repo_id': 'Adv-FT-Interp/llava-1.5-7b-hf-r1-s3-faithful-ft', '_wandb': {}}
9
+ 2026-08-18 18:56:16,799 INFO MainThread:409376 [wandb_init.py:init():816] starting backend
10
+ 2026-08-18 18:56:16,799 INFO MainThread:409376 [wandb_init.py:init():831] sending inform_init request
11
+ 2026-08-18 18:56:17,149 INFO MainThread:409376 [wandb_init.py:init():836] backend started and connected
12
+ 2026-08-18 18:56:17,151 INFO MainThread:409376 [wandb_init.py:init():906] updated telemetry
13
+ 2026-08-18 18:56:17,172 INFO MainThread:409376 [wandb_init.py:init():929] communicating run to backend with 90.0 second timeout
14
+ 2026-08-18 18:56:17,467 INFO MainThread:409376 [wandb_init.py:init():974] starting run threads in backend
15
+ 2026-08-18 18:56:17,956 INFO MainThread:409376 [wandb_run.py:_console_start():2523] atexit reg
16
+ 2026-08-18 18:56:17,957 INFO MainThread:409376 [wandb_run.py:_redirect():2373] redirect: wrap_raw
17
+ 2026-08-18 18:56:17,957 INFO MainThread:409376 [wandb_run.py:_redirect():2442] Wrapping output streams.
18
+ 2026-08-18 18:56:17,957 INFO MainThread:409376 [wandb_run.py:_redirect():2465] Redirects installed.
19
+ 2026-08-18 18:56:17,960 INFO MainThread:409376 [wandb_init.py:init():1012] run started, returning control to user process
20
+ 2026-08-18 18:56:17,961 INFO MainThread:409376 [wandb_run.py:_config_callback():1346] config_cb None None {'training_script': 'finetuning_test_faithful.py', 'objective': 'corruption_fare', 'embedding_dim': 768, 'projection_trained': True, 'run_name': 'llava-1.5-7b-hf-r1-s3-faithful-ft-h200-20260818-185519'}
wandb/run-20260818_185616-oc7a60u3/run-oc7a60u3.wandb ADDED
@@ -0,0 +1,3 @@
 
 
 
 
1
+ version https://git-lfs.github.com/spec/v1
2
+ oid sha256:f09026b9b009250a5ac824cced7aee1b4348b4dd8fa00dc3ad5cff302ee4729e
3
+ size 1507328