4gate commited on
Commit
1cacd3f
·
verified ·
1 Parent(s): e48ec63

Add files using upload-large-folder tool

Browse files
This view is limited to 50 files because it contains too many changes.   See raw diff
Files changed (50) hide show
  1. proper_wanda_sparsities_2026_08_10_32k/gemma3/calibration_biology/sparsity_0p05/checkpoint_results.json +54 -0
  2. proper_wanda_sparsities_2026_08_10_32k/gemma3/calibration_biology/sparsity_0p05/recovered_model/added_tokens.json +3 -0
  3. proper_wanda_sparsities_2026_08_10_32k/gemma3/calibration_biology/sparsity_0p05/recovered_model/chat_template.jinja +47 -0
  4. proper_wanda_sparsities_2026_08_10_32k/gemma3/calibration_biology/sparsity_0p05/recovered_model/config.json +111 -0
  5. proper_wanda_sparsities_2026_08_10_32k/gemma3/calibration_biology/sparsity_0p05/recovered_model/generation_config.json +14 -0
  6. proper_wanda_sparsities_2026_08_10_32k/gemma3/calibration_biology/sparsity_0p05/recovered_model/model.safetensors.index.json +0 -0
  7. proper_wanda_sparsities_2026_08_10_32k/gemma3/calibration_biology/sparsity_0p05/recovered_model/special_tokens_map.json +33 -0
  8. proper_wanda_sparsities_2026_08_10_32k/gemma3/calibration_biology/sparsity_0p05/recovered_model/tokenizer_config.json +0 -0
  9. proper_wanda_sparsities_2026_08_10_32k/gemma3/calibration_biology/sparsity_0p1/checkpoint_results.json +54 -0
  10. proper_wanda_sparsities_2026_08_10_32k/gemma3/calibration_biology/sparsity_0p1/recovered_model/added_tokens.json +3 -0
  11. proper_wanda_sparsities_2026_08_10_32k/gemma3/calibration_biology/sparsity_0p1/recovered_model/chat_template.jinja +47 -0
  12. proper_wanda_sparsities_2026_08_10_32k/gemma3/calibration_biology/sparsity_0p1/recovered_model/config.json +111 -0
  13. proper_wanda_sparsities_2026_08_10_32k/gemma3/calibration_biology/sparsity_0p1/recovered_model/generation_config.json +14 -0
  14. proper_wanda_sparsities_2026_08_10_32k/gemma3/calibration_biology/sparsity_0p1/recovered_model/model.safetensors.index.json +0 -0
  15. proper_wanda_sparsities_2026_08_10_32k/gemma3/calibration_biology/sparsity_0p1/recovered_model/special_tokens_map.json +33 -0
  16. proper_wanda_sparsities_2026_08_10_32k/gemma3/calibration_biology/sparsity_0p1/recovered_model/tokenizer_config.json +0 -0
  17. proper_wanda_sparsities_2026_08_10_32k/gemma3/calibration_biology/sparsity_0p15/checkpoint_results.json +54 -0
  18. proper_wanda_sparsities_2026_08_10_32k/gemma3/calibration_biology/sparsity_0p15/recovered_model/added_tokens.json +3 -0
  19. proper_wanda_sparsities_2026_08_10_32k/gemma3/calibration_biology/sparsity_0p15/recovered_model/chat_template.jinja +47 -0
  20. proper_wanda_sparsities_2026_08_10_32k/gemma3/calibration_biology/sparsity_0p15/recovered_model/config.json +111 -0
  21. proper_wanda_sparsities_2026_08_10_32k/gemma3/calibration_biology/sparsity_0p15/recovered_model/generation_config.json +14 -0
  22. proper_wanda_sparsities_2026_08_10_32k/gemma3/calibration_biology/sparsity_0p15/recovered_model/model.safetensors.index.json +0 -0
  23. proper_wanda_sparsities_2026_08_10_32k/gemma3/calibration_biology/sparsity_0p15/recovered_model/special_tokens_map.json +33 -0
  24. proper_wanda_sparsities_2026_08_10_32k/gemma3/calibration_biology/sparsity_0p15/recovered_model/tokenizer_config.json +0 -0
  25. proper_wanda_sparsities_2026_08_10_32k/gemma3/calibration_biology/sparsity_0p2/checkpoint_results.json +54 -0
  26. proper_wanda_sparsities_2026_08_10_32k/gemma3/calibration_biology/sparsity_0p2/recovered_model/added_tokens.json +3 -0
  27. proper_wanda_sparsities_2026_08_10_32k/gemma3/calibration_biology/sparsity_0p2/recovered_model/chat_template.jinja +47 -0
  28. proper_wanda_sparsities_2026_08_10_32k/gemma3/calibration_biology/sparsity_0p2/recovered_model/config.json +111 -0
  29. proper_wanda_sparsities_2026_08_10_32k/gemma3/calibration_biology/sparsity_0p2/recovered_model/generation_config.json +14 -0
  30. proper_wanda_sparsities_2026_08_10_32k/gemma3/calibration_biology/sparsity_0p2/recovered_model/model.safetensors.index.json +0 -0
  31. proper_wanda_sparsities_2026_08_10_32k/gemma3/calibration_biology/sparsity_0p2/recovered_model/special_tokens_map.json +33 -0
  32. proper_wanda_sparsities_2026_08_10_32k/gemma3/calibration_biology/sparsity_0p2/recovered_model/tokenizer_config.json +0 -0
  33. proper_wanda_sparsities_2026_08_10_32k/gemma3/calibration_biology/sparsity_0p25/checkpoint_results.json +54 -0
  34. proper_wanda_sparsities_2026_08_10_32k/gemma3/calibration_biology/sparsity_0p25/recovered_model/added_tokens.json +3 -0
  35. proper_wanda_sparsities_2026_08_10_32k/gemma3/calibration_biology/sparsity_0p25/recovered_model/chat_template.jinja +47 -0
  36. proper_wanda_sparsities_2026_08_10_32k/gemma3/calibration_biology/sparsity_0p25/recovered_model/config.json +111 -0
  37. proper_wanda_sparsities_2026_08_10_32k/gemma3/calibration_biology/sparsity_0p25/recovered_model/generation_config.json +14 -0
  38. proper_wanda_sparsities_2026_08_10_32k/gemma3/calibration_biology/sparsity_0p25/recovered_model/model.safetensors.index.json +0 -0
  39. proper_wanda_sparsities_2026_08_10_32k/gemma3/calibration_biology/sparsity_0p25/recovered_model/special_tokens_map.json +33 -0
  40. proper_wanda_sparsities_2026_08_10_32k/gemma3/calibration_biology/sparsity_0p25/recovered_model/tokenizer_config.json +0 -0
  41. proper_wanda_sparsities_2026_08_10_32k/gemma3/calibration_biology/sparsity_0p3/checkpoint_results.json +54 -0
  42. proper_wanda_sparsities_2026_08_10_32k/gemma3/calibration_biology/sparsity_0p3/recovered_model/added_tokens.json +3 -0
  43. proper_wanda_sparsities_2026_08_10_32k/gemma3/calibration_biology/sparsity_0p3/recovered_model/chat_template.jinja +47 -0
  44. proper_wanda_sparsities_2026_08_10_32k/gemma3/calibration_biology/sparsity_0p3/recovered_model/config.json +111 -0
  45. proper_wanda_sparsities_2026_08_10_32k/gemma3/calibration_biology/sparsity_0p3/recovered_model/generation_config.json +14 -0
  46. proper_wanda_sparsities_2026_08_10_32k/gemma3/calibration_biology/sparsity_0p3/recovered_model/model.safetensors.index.json +0 -0
  47. proper_wanda_sparsities_2026_08_10_32k/gemma3/calibration_biology/sparsity_0p3/recovered_model/special_tokens_map.json +33 -0
  48. proper_wanda_sparsities_2026_08_10_32k/gemma3/calibration_biology/sparsity_0p3/recovered_model/tokenizer_config.json +0 -0
  49. proper_wanda_sparsities_2026_08_10_32k/gemma3/calibration_biology/sparsity_0p35/checkpoint_results.json +54 -0
  50. proper_wanda_sparsities_2026_08_10_32k/gemma3/calibration_biology/sparsity_0p35/recovered_model/added_tokens.json +3 -0
proper_wanda_sparsities_2026_08_10_32k/gemma3/calibration_biology/sparsity_0p05/checkpoint_results.json ADDED
@@ -0,0 +1,54 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ {
2
+ "model": "gemma3",
3
+ "domain": "biology",
4
+ "sparsity": 0.05,
5
+ "pgd_steps": 2000,
6
+ "training_examples": 32000,
7
+ "per_device_train_batch_size": 2,
8
+ "gradient_accumulation_steps": 8,
9
+ "reference_loss": 3.1282850253582,
10
+ "loss_history": [
11
+ {
12
+ "step": 200,
13
+ "loss": 1.0705748826265336
14
+ },
15
+ {
16
+ "step": 400,
17
+ "loss": 1.0435286790132523
18
+ },
19
+ {
20
+ "step": 600,
21
+ "loss": 1.0298420244455337
22
+ },
23
+ {
24
+ "step": 800,
25
+ "loss": 1.0174534922838212
26
+ },
27
+ {
28
+ "step": 1000,
29
+ "loss": 1.0074299648404121
30
+ },
31
+ {
32
+ "step": 1200,
33
+ "loss": 1.0010298210382462
34
+ },
35
+ {
36
+ "step": 1400,
37
+ "loss": 0.9957343101501465
38
+ },
39
+ {
40
+ "step": 1600,
41
+ "loss": 0.9930229991674423
42
+ },
43
+ {
44
+ "step": 1800,
45
+ "loss": 0.9922317489981651
46
+ },
47
+ {
48
+ "step": 2000,
49
+ "loss": 0.992279257774353
50
+ }
51
+ ],
52
+ "sparsity_after_pgd": 0.049986293859649124,
53
+ "masked_nonzero_after_pgd": 0
54
+ }
proper_wanda_sparsities_2026_08_10_32k/gemma3/calibration_biology/sparsity_0p05/recovered_model/added_tokens.json ADDED
@@ -0,0 +1,3 @@
 
 
 
 
1
+ {
2
+ "<image_soft_token>": 262144
3
+ }
proper_wanda_sparsities_2026_08_10_32k/gemma3/calibration_biology/sparsity_0p05/recovered_model/chat_template.jinja ADDED
@@ -0,0 +1,47 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ {{ bos_token }}
2
+ {%- if messages[0]['role'] == 'system' -%}
3
+ {%- if messages[0]['content'] is string -%}
4
+ {%- set first_user_prefix = messages[0]['content'] + '
5
+
6
+ ' -%}
7
+ {%- else -%}
8
+ {%- set first_user_prefix = messages[0]['content'][0]['text'] + '
9
+
10
+ ' -%}
11
+ {%- endif -%}
12
+ {%- set loop_messages = messages[1:] -%}
13
+ {%- else -%}
14
+ {%- set first_user_prefix = "" -%}
15
+ {%- set loop_messages = messages -%}
16
+ {%- endif -%}
17
+ {%- for message in loop_messages -%}
18
+ {%- if (message['role'] == 'user') != (loop.index0 % 2 == 0) -%}
19
+ {{ raise_exception("Conversation roles must alternate user/assistant/user/assistant/...") }}
20
+ {%- endif -%}
21
+ {%- if (message['role'] == 'assistant') -%}
22
+ {%- set role = "model" -%}
23
+ {%- else -%}
24
+ {%- set role = message['role'] -%}
25
+ {%- endif -%}
26
+ {{ '<start_of_turn>' + role + '
27
+ ' + (first_user_prefix if loop.first else "") }}
28
+ {%- if message['content'] is string -%}
29
+ {{ message['content'] | trim }}
30
+ {%- elif message['content'] is iterable -%}
31
+ {%- for item in message['content'] -%}
32
+ {%- if item['type'] == 'image' -%}
33
+ {{ '<start_of_image>' }}
34
+ {%- elif item['type'] == 'text' -%}
35
+ {{ item['text'] | trim }}
36
+ {%- endif -%}
37
+ {%- endfor -%}
38
+ {%- else -%}
39
+ {{ raise_exception("Invalid content type") }}
40
+ {%- endif -%}
41
+ {{ '<end_of_turn>
42
+ ' }}
43
+ {%- endfor -%}
44
+ {%- if add_generation_prompt -%}
45
+ {{'<start_of_turn>model
46
+ '}}
47
+ {%- endif -%}
proper_wanda_sparsities_2026_08_10_32k/gemma3/calibration_biology/sparsity_0p05/recovered_model/config.json ADDED
@@ -0,0 +1,111 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ {
2
+ "architectures": [
3
+ "Gemma3ForConditionalGeneration"
4
+ ],
5
+ "boi_token_index": 255999,
6
+ "bos_token_id": 2,
7
+ "dtype": "bfloat16",
8
+ "eoi_token_index": 256000,
9
+ "eos_token_id": 1,
10
+ "image_token_index": 262144,
11
+ "initializer_range": 0.02,
12
+ "mm_tokens_per_image": 256,
13
+ "model_type": "gemma3",
14
+ "pad_token_id": 0,
15
+ "text_config": {
16
+ "_sliding_window_pattern": 6,
17
+ "attention_bias": false,
18
+ "attention_dropout": 0.0,
19
+ "attn_logit_softcapping": null,
20
+ "dtype": "bfloat16",
21
+ "final_logit_softcapping": null,
22
+ "head_dim": 256,
23
+ "hidden_activation": "gelu_pytorch_tanh",
24
+ "hidden_size": 3840,
25
+ "initializer_range": 0.02,
26
+ "intermediate_size": 15360,
27
+ "layer_types": [
28
+ "sliding_attention",
29
+ "sliding_attention",
30
+ "sliding_attention",
31
+ "sliding_attention",
32
+ "sliding_attention",
33
+ "full_attention",
34
+ "sliding_attention",
35
+ "sliding_attention",
36
+ "sliding_attention",
37
+ "sliding_attention",
38
+ "sliding_attention",
39
+ "full_attention",
40
+ "sliding_attention",
41
+ "sliding_attention",
42
+ "sliding_attention",
43
+ "sliding_attention",
44
+ "sliding_attention",
45
+ "full_attention",
46
+ "sliding_attention",
47
+ "sliding_attention",
48
+ "sliding_attention",
49
+ "sliding_attention",
50
+ "sliding_attention",
51
+ "full_attention",
52
+ "sliding_attention",
53
+ "sliding_attention",
54
+ "sliding_attention",
55
+ "sliding_attention",
56
+ "sliding_attention",
57
+ "full_attention",
58
+ "sliding_attention",
59
+ "sliding_attention",
60
+ "sliding_attention",
61
+ "sliding_attention",
62
+ "sliding_attention",
63
+ "full_attention",
64
+ "sliding_attention",
65
+ "sliding_attention",
66
+ "sliding_attention",
67
+ "sliding_attention",
68
+ "sliding_attention",
69
+ "full_attention",
70
+ "sliding_attention",
71
+ "sliding_attention",
72
+ "sliding_attention",
73
+ "sliding_attention",
74
+ "sliding_attention",
75
+ "full_attention"
76
+ ],
77
+ "max_position_embeddings": 131072,
78
+ "model_type": "gemma3_text",
79
+ "num_attention_heads": 16,
80
+ "num_hidden_layers": 48,
81
+ "num_key_value_heads": 8,
82
+ "query_pre_attn_scalar": 256,
83
+ "rms_norm_eps": 1e-06,
84
+ "rope_local_base_freq": 10000.0,
85
+ "rope_scaling": {
86
+ "factor": 8.0,
87
+ "rope_type": "linear"
88
+ },
89
+ "rope_theta": 1000000.0,
90
+ "sliding_window": 1024,
91
+ "use_cache": true,
92
+ "vocab_size": 262208
93
+ },
94
+ "transformers_version": "4.56.1",
95
+ "use_cache": false,
96
+ "vision_config": {
97
+ "attention_dropout": 0.0,
98
+ "dtype": "bfloat16",
99
+ "hidden_act": "gelu_pytorch_tanh",
100
+ "hidden_size": 1152,
101
+ "image_size": 896,
102
+ "intermediate_size": 4304,
103
+ "layer_norm_eps": 1e-06,
104
+ "model_type": "siglip_vision_model",
105
+ "num_attention_heads": 16,
106
+ "num_channels": 3,
107
+ "num_hidden_layers": 27,
108
+ "patch_size": 14,
109
+ "vision_use_head": false
110
+ }
111
+ }
proper_wanda_sparsities_2026_08_10_32k/gemma3/calibration_biology/sparsity_0p05/recovered_model/generation_config.json ADDED
@@ -0,0 +1,14 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ {
2
+ "bos_token_id": 2,
3
+ "cache_implementation": "hybrid",
4
+ "do_sample": true,
5
+ "eos_token_id": [
6
+ 1,
7
+ 1,
8
+ 106
9
+ ],
10
+ "pad_token_id": 0,
11
+ "top_k": 64,
12
+ "top_p": 0.95,
13
+ "transformers_version": "4.56.1"
14
+ }
proper_wanda_sparsities_2026_08_10_32k/gemma3/calibration_biology/sparsity_0p05/recovered_model/model.safetensors.index.json ADDED
The diff for this file is too large to render. See raw diff
 
proper_wanda_sparsities_2026_08_10_32k/gemma3/calibration_biology/sparsity_0p05/recovered_model/special_tokens_map.json ADDED
@@ -0,0 +1,33 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ {
2
+ "boi_token": "<start_of_image>",
3
+ "bos_token": {
4
+ "content": "<bos>",
5
+ "lstrip": false,
6
+ "normalized": false,
7
+ "rstrip": false,
8
+ "single_word": false
9
+ },
10
+ "eoi_token": "<end_of_image>",
11
+ "eos_token": {
12
+ "content": "<eos>",
13
+ "lstrip": false,
14
+ "normalized": false,
15
+ "rstrip": false,
16
+ "single_word": false
17
+ },
18
+ "image_token": "<image_soft_token>",
19
+ "pad_token": {
20
+ "content": "<pad>",
21
+ "lstrip": false,
22
+ "normalized": false,
23
+ "rstrip": false,
24
+ "single_word": false
25
+ },
26
+ "unk_token": {
27
+ "content": "<unk>",
28
+ "lstrip": false,
29
+ "normalized": false,
30
+ "rstrip": false,
31
+ "single_word": false
32
+ }
33
+ }
proper_wanda_sparsities_2026_08_10_32k/gemma3/calibration_biology/sparsity_0p05/recovered_model/tokenizer_config.json ADDED
The diff for this file is too large to render. See raw diff
 
proper_wanda_sparsities_2026_08_10_32k/gemma3/calibration_biology/sparsity_0p1/checkpoint_results.json ADDED
@@ -0,0 +1,54 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ {
2
+ "model": "gemma3",
3
+ "domain": "biology",
4
+ "sparsity": 0.1,
5
+ "pgd_steps": 2000,
6
+ "training_examples": 32000,
7
+ "per_device_train_batch_size": 2,
8
+ "gradient_accumulation_steps": 8,
9
+ "reference_loss": 3.1282850253582,
10
+ "loss_history": [
11
+ {
12
+ "step": 200,
13
+ "loss": 1.070497879385948
14
+ },
15
+ {
16
+ "step": 400,
17
+ "loss": 1.04411971449852
18
+ },
19
+ {
20
+ "step": 600,
21
+ "loss": 1.0304869377613068
22
+ },
23
+ {
24
+ "step": 800,
25
+ "loss": 1.019303821325302
26
+ },
27
+ {
28
+ "step": 1000,
29
+ "loss": 1.0100166726112365
30
+ },
31
+ {
32
+ "step": 1200,
33
+ "loss": 1.0044535639882088
34
+ },
35
+ {
36
+ "step": 1400,
37
+ "loss": 1.0001539278030396
38
+ },
39
+ {
40
+ "step": 1600,
41
+ "loss": 0.9982488462328911
42
+ },
43
+ {
44
+ "step": 1800,
45
+ "loss": 0.9972800877690315
46
+ },
47
+ {
48
+ "step": 2000,
49
+ "loss": 0.997349861562252
50
+ }
51
+ ],
52
+ "sparsity_after_pgd": 0.09998972039473684,
53
+ "masked_nonzero_after_pgd": 0
54
+ }
proper_wanda_sparsities_2026_08_10_32k/gemma3/calibration_biology/sparsity_0p1/recovered_model/added_tokens.json ADDED
@@ -0,0 +1,3 @@
 
 
 
 
1
+ {
2
+ "<image_soft_token>": 262144
3
+ }
proper_wanda_sparsities_2026_08_10_32k/gemma3/calibration_biology/sparsity_0p1/recovered_model/chat_template.jinja ADDED
@@ -0,0 +1,47 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ {{ bos_token }}
2
+ {%- if messages[0]['role'] == 'system' -%}
3
+ {%- if messages[0]['content'] is string -%}
4
+ {%- set first_user_prefix = messages[0]['content'] + '
5
+
6
+ ' -%}
7
+ {%- else -%}
8
+ {%- set first_user_prefix = messages[0]['content'][0]['text'] + '
9
+
10
+ ' -%}
11
+ {%- endif -%}
12
+ {%- set loop_messages = messages[1:] -%}
13
+ {%- else -%}
14
+ {%- set first_user_prefix = "" -%}
15
+ {%- set loop_messages = messages -%}
16
+ {%- endif -%}
17
+ {%- for message in loop_messages -%}
18
+ {%- if (message['role'] == 'user') != (loop.index0 % 2 == 0) -%}
19
+ {{ raise_exception("Conversation roles must alternate user/assistant/user/assistant/...") }}
20
+ {%- endif -%}
21
+ {%- if (message['role'] == 'assistant') -%}
22
+ {%- set role = "model" -%}
23
+ {%- else -%}
24
+ {%- set role = message['role'] -%}
25
+ {%- endif -%}
26
+ {{ '<start_of_turn>' + role + '
27
+ ' + (first_user_prefix if loop.first else "") }}
28
+ {%- if message['content'] is string -%}
29
+ {{ message['content'] | trim }}
30
+ {%- elif message['content'] is iterable -%}
31
+ {%- for item in message['content'] -%}
32
+ {%- if item['type'] == 'image' -%}
33
+ {{ '<start_of_image>' }}
34
+ {%- elif item['type'] == 'text' -%}
35
+ {{ item['text'] | trim }}
36
+ {%- endif -%}
37
+ {%- endfor -%}
38
+ {%- else -%}
39
+ {{ raise_exception("Invalid content type") }}
40
+ {%- endif -%}
41
+ {{ '<end_of_turn>
42
+ ' }}
43
+ {%- endfor -%}
44
+ {%- if add_generation_prompt -%}
45
+ {{'<start_of_turn>model
46
+ '}}
47
+ {%- endif -%}
proper_wanda_sparsities_2026_08_10_32k/gemma3/calibration_biology/sparsity_0p1/recovered_model/config.json ADDED
@@ -0,0 +1,111 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ {
2
+ "architectures": [
3
+ "Gemma3ForConditionalGeneration"
4
+ ],
5
+ "boi_token_index": 255999,
6
+ "bos_token_id": 2,
7
+ "dtype": "bfloat16",
8
+ "eoi_token_index": 256000,
9
+ "eos_token_id": 1,
10
+ "image_token_index": 262144,
11
+ "initializer_range": 0.02,
12
+ "mm_tokens_per_image": 256,
13
+ "model_type": "gemma3",
14
+ "pad_token_id": 0,
15
+ "text_config": {
16
+ "_sliding_window_pattern": 6,
17
+ "attention_bias": false,
18
+ "attention_dropout": 0.0,
19
+ "attn_logit_softcapping": null,
20
+ "dtype": "bfloat16",
21
+ "final_logit_softcapping": null,
22
+ "head_dim": 256,
23
+ "hidden_activation": "gelu_pytorch_tanh",
24
+ "hidden_size": 3840,
25
+ "initializer_range": 0.02,
26
+ "intermediate_size": 15360,
27
+ "layer_types": [
28
+ "sliding_attention",
29
+ "sliding_attention",
30
+ "sliding_attention",
31
+ "sliding_attention",
32
+ "sliding_attention",
33
+ "full_attention",
34
+ "sliding_attention",
35
+ "sliding_attention",
36
+ "sliding_attention",
37
+ "sliding_attention",
38
+ "sliding_attention",
39
+ "full_attention",
40
+ "sliding_attention",
41
+ "sliding_attention",
42
+ "sliding_attention",
43
+ "sliding_attention",
44
+ "sliding_attention",
45
+ "full_attention",
46
+ "sliding_attention",
47
+ "sliding_attention",
48
+ "sliding_attention",
49
+ "sliding_attention",
50
+ "sliding_attention",
51
+ "full_attention",
52
+ "sliding_attention",
53
+ "sliding_attention",
54
+ "sliding_attention",
55
+ "sliding_attention",
56
+ "sliding_attention",
57
+ "full_attention",
58
+ "sliding_attention",
59
+ "sliding_attention",
60
+ "sliding_attention",
61
+ "sliding_attention",
62
+ "sliding_attention",
63
+ "full_attention",
64
+ "sliding_attention",
65
+ "sliding_attention",
66
+ "sliding_attention",
67
+ "sliding_attention",
68
+ "sliding_attention",
69
+ "full_attention",
70
+ "sliding_attention",
71
+ "sliding_attention",
72
+ "sliding_attention",
73
+ "sliding_attention",
74
+ "sliding_attention",
75
+ "full_attention"
76
+ ],
77
+ "max_position_embeddings": 131072,
78
+ "model_type": "gemma3_text",
79
+ "num_attention_heads": 16,
80
+ "num_hidden_layers": 48,
81
+ "num_key_value_heads": 8,
82
+ "query_pre_attn_scalar": 256,
83
+ "rms_norm_eps": 1e-06,
84
+ "rope_local_base_freq": 10000.0,
85
+ "rope_scaling": {
86
+ "factor": 8.0,
87
+ "rope_type": "linear"
88
+ },
89
+ "rope_theta": 1000000.0,
90
+ "sliding_window": 1024,
91
+ "use_cache": true,
92
+ "vocab_size": 262208
93
+ },
94
+ "transformers_version": "4.56.1",
95
+ "use_cache": false,
96
+ "vision_config": {
97
+ "attention_dropout": 0.0,
98
+ "dtype": "bfloat16",
99
+ "hidden_act": "gelu_pytorch_tanh",
100
+ "hidden_size": 1152,
101
+ "image_size": 896,
102
+ "intermediate_size": 4304,
103
+ "layer_norm_eps": 1e-06,
104
+ "model_type": "siglip_vision_model",
105
+ "num_attention_heads": 16,
106
+ "num_channels": 3,
107
+ "num_hidden_layers": 27,
108
+ "patch_size": 14,
109
+ "vision_use_head": false
110
+ }
111
+ }
proper_wanda_sparsities_2026_08_10_32k/gemma3/calibration_biology/sparsity_0p1/recovered_model/generation_config.json ADDED
@@ -0,0 +1,14 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ {
2
+ "bos_token_id": 2,
3
+ "cache_implementation": "hybrid",
4
+ "do_sample": true,
5
+ "eos_token_id": [
6
+ 1,
7
+ 1,
8
+ 106
9
+ ],
10
+ "pad_token_id": 0,
11
+ "top_k": 64,
12
+ "top_p": 0.95,
13
+ "transformers_version": "4.56.1"
14
+ }
proper_wanda_sparsities_2026_08_10_32k/gemma3/calibration_biology/sparsity_0p1/recovered_model/model.safetensors.index.json ADDED
The diff for this file is too large to render. See raw diff
 
proper_wanda_sparsities_2026_08_10_32k/gemma3/calibration_biology/sparsity_0p1/recovered_model/special_tokens_map.json ADDED
@@ -0,0 +1,33 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ {
2
+ "boi_token": "<start_of_image>",
3
+ "bos_token": {
4
+ "content": "<bos>",
5
+ "lstrip": false,
6
+ "normalized": false,
7
+ "rstrip": false,
8
+ "single_word": false
9
+ },
10
+ "eoi_token": "<end_of_image>",
11
+ "eos_token": {
12
+ "content": "<eos>",
13
+ "lstrip": false,
14
+ "normalized": false,
15
+ "rstrip": false,
16
+ "single_word": false
17
+ },
18
+ "image_token": "<image_soft_token>",
19
+ "pad_token": {
20
+ "content": "<pad>",
21
+ "lstrip": false,
22
+ "normalized": false,
23
+ "rstrip": false,
24
+ "single_word": false
25
+ },
26
+ "unk_token": {
27
+ "content": "<unk>",
28
+ "lstrip": false,
29
+ "normalized": false,
30
+ "rstrip": false,
31
+ "single_word": false
32
+ }
33
+ }
proper_wanda_sparsities_2026_08_10_32k/gemma3/calibration_biology/sparsity_0p1/recovered_model/tokenizer_config.json ADDED
The diff for this file is too large to render. See raw diff
 
proper_wanda_sparsities_2026_08_10_32k/gemma3/calibration_biology/sparsity_0p15/checkpoint_results.json ADDED
@@ -0,0 +1,54 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ {
2
+ "model": "gemma3",
3
+ "domain": "biology",
4
+ "sparsity": 0.15,
5
+ "pgd_steps": 2000,
6
+ "training_examples": 32000,
7
+ "per_device_train_batch_size": 2,
8
+ "gradient_accumulation_steps": 8,
9
+ "reference_loss": 3.1282850253582,
10
+ "loss_history": [
11
+ {
12
+ "step": 200,
13
+ "loss": 1.0733222764730455
14
+ },
15
+ {
16
+ "step": 400,
17
+ "loss": 1.046976425051689
18
+ },
19
+ {
20
+ "step": 600,
21
+ "loss": 1.0339794331789016
22
+ },
23
+ {
24
+ "step": 800,
25
+ "loss": 1.0235021966695785
26
+ },
27
+ {
28
+ "step": 1000,
29
+ "loss": 1.0148386600613595
30
+ },
31
+ {
32
+ "step": 1200,
33
+ "loss": 1.0101912748813628
34
+ },
35
+ {
36
+ "step": 1400,
37
+ "loss": 1.0064997225999832
38
+ },
39
+ {
40
+ "step": 1600,
41
+ "loss": 1.0048882424831391
42
+ },
43
+ {
44
+ "step": 1800,
45
+ "loss": 1.0043698224425315
46
+ },
47
+ {
48
+ "step": 2000,
49
+ "loss": 1.004220250248909
50
+ }
51
+ ],
52
+ "sparsity_after_pgd": 0.14999314692982457,
53
+ "masked_nonzero_after_pgd": 0
54
+ }
proper_wanda_sparsities_2026_08_10_32k/gemma3/calibration_biology/sparsity_0p15/recovered_model/added_tokens.json ADDED
@@ -0,0 +1,3 @@
 
 
 
 
1
+ {
2
+ "<image_soft_token>": 262144
3
+ }
proper_wanda_sparsities_2026_08_10_32k/gemma3/calibration_biology/sparsity_0p15/recovered_model/chat_template.jinja ADDED
@@ -0,0 +1,47 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ {{ bos_token }}
2
+ {%- if messages[0]['role'] == 'system' -%}
3
+ {%- if messages[0]['content'] is string -%}
4
+ {%- set first_user_prefix = messages[0]['content'] + '
5
+
6
+ ' -%}
7
+ {%- else -%}
8
+ {%- set first_user_prefix = messages[0]['content'][0]['text'] + '
9
+
10
+ ' -%}
11
+ {%- endif -%}
12
+ {%- set loop_messages = messages[1:] -%}
13
+ {%- else -%}
14
+ {%- set first_user_prefix = "" -%}
15
+ {%- set loop_messages = messages -%}
16
+ {%- endif -%}
17
+ {%- for message in loop_messages -%}
18
+ {%- if (message['role'] == 'user') != (loop.index0 % 2 == 0) -%}
19
+ {{ raise_exception("Conversation roles must alternate user/assistant/user/assistant/...") }}
20
+ {%- endif -%}
21
+ {%- if (message['role'] == 'assistant') -%}
22
+ {%- set role = "model" -%}
23
+ {%- else -%}
24
+ {%- set role = message['role'] -%}
25
+ {%- endif -%}
26
+ {{ '<start_of_turn>' + role + '
27
+ ' + (first_user_prefix if loop.first else "") }}
28
+ {%- if message['content'] is string -%}
29
+ {{ message['content'] | trim }}
30
+ {%- elif message['content'] is iterable -%}
31
+ {%- for item in message['content'] -%}
32
+ {%- if item['type'] == 'image' -%}
33
+ {{ '<start_of_image>' }}
34
+ {%- elif item['type'] == 'text' -%}
35
+ {{ item['text'] | trim }}
36
+ {%- endif -%}
37
+ {%- endfor -%}
38
+ {%- else -%}
39
+ {{ raise_exception("Invalid content type") }}
40
+ {%- endif -%}
41
+ {{ '<end_of_turn>
42
+ ' }}
43
+ {%- endfor -%}
44
+ {%- if add_generation_prompt -%}
45
+ {{'<start_of_turn>model
46
+ '}}
47
+ {%- endif -%}
proper_wanda_sparsities_2026_08_10_32k/gemma3/calibration_biology/sparsity_0p15/recovered_model/config.json ADDED
@@ -0,0 +1,111 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ {
2
+ "architectures": [
3
+ "Gemma3ForConditionalGeneration"
4
+ ],
5
+ "boi_token_index": 255999,
6
+ "bos_token_id": 2,
7
+ "dtype": "bfloat16",
8
+ "eoi_token_index": 256000,
9
+ "eos_token_id": 1,
10
+ "image_token_index": 262144,
11
+ "initializer_range": 0.02,
12
+ "mm_tokens_per_image": 256,
13
+ "model_type": "gemma3",
14
+ "pad_token_id": 0,
15
+ "text_config": {
16
+ "_sliding_window_pattern": 6,
17
+ "attention_bias": false,
18
+ "attention_dropout": 0.0,
19
+ "attn_logit_softcapping": null,
20
+ "dtype": "bfloat16",
21
+ "final_logit_softcapping": null,
22
+ "head_dim": 256,
23
+ "hidden_activation": "gelu_pytorch_tanh",
24
+ "hidden_size": 3840,
25
+ "initializer_range": 0.02,
26
+ "intermediate_size": 15360,
27
+ "layer_types": [
28
+ "sliding_attention",
29
+ "sliding_attention",
30
+ "sliding_attention",
31
+ "sliding_attention",
32
+ "sliding_attention",
33
+ "full_attention",
34
+ "sliding_attention",
35
+ "sliding_attention",
36
+ "sliding_attention",
37
+ "sliding_attention",
38
+ "sliding_attention",
39
+ "full_attention",
40
+ "sliding_attention",
41
+ "sliding_attention",
42
+ "sliding_attention",
43
+ "sliding_attention",
44
+ "sliding_attention",
45
+ "full_attention",
46
+ "sliding_attention",
47
+ "sliding_attention",
48
+ "sliding_attention",
49
+ "sliding_attention",
50
+ "sliding_attention",
51
+ "full_attention",
52
+ "sliding_attention",
53
+ "sliding_attention",
54
+ "sliding_attention",
55
+ "sliding_attention",
56
+ "sliding_attention",
57
+ "full_attention",
58
+ "sliding_attention",
59
+ "sliding_attention",
60
+ "sliding_attention",
61
+ "sliding_attention",
62
+ "sliding_attention",
63
+ "full_attention",
64
+ "sliding_attention",
65
+ "sliding_attention",
66
+ "sliding_attention",
67
+ "sliding_attention",
68
+ "sliding_attention",
69
+ "full_attention",
70
+ "sliding_attention",
71
+ "sliding_attention",
72
+ "sliding_attention",
73
+ "sliding_attention",
74
+ "sliding_attention",
75
+ "full_attention"
76
+ ],
77
+ "max_position_embeddings": 131072,
78
+ "model_type": "gemma3_text",
79
+ "num_attention_heads": 16,
80
+ "num_hidden_layers": 48,
81
+ "num_key_value_heads": 8,
82
+ "query_pre_attn_scalar": 256,
83
+ "rms_norm_eps": 1e-06,
84
+ "rope_local_base_freq": 10000.0,
85
+ "rope_scaling": {
86
+ "factor": 8.0,
87
+ "rope_type": "linear"
88
+ },
89
+ "rope_theta": 1000000.0,
90
+ "sliding_window": 1024,
91
+ "use_cache": true,
92
+ "vocab_size": 262208
93
+ },
94
+ "transformers_version": "4.56.1",
95
+ "use_cache": false,
96
+ "vision_config": {
97
+ "attention_dropout": 0.0,
98
+ "dtype": "bfloat16",
99
+ "hidden_act": "gelu_pytorch_tanh",
100
+ "hidden_size": 1152,
101
+ "image_size": 896,
102
+ "intermediate_size": 4304,
103
+ "layer_norm_eps": 1e-06,
104
+ "model_type": "siglip_vision_model",
105
+ "num_attention_heads": 16,
106
+ "num_channels": 3,
107
+ "num_hidden_layers": 27,
108
+ "patch_size": 14,
109
+ "vision_use_head": false
110
+ }
111
+ }
proper_wanda_sparsities_2026_08_10_32k/gemma3/calibration_biology/sparsity_0p15/recovered_model/generation_config.json ADDED
@@ -0,0 +1,14 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ {
2
+ "bos_token_id": 2,
3
+ "cache_implementation": "hybrid",
4
+ "do_sample": true,
5
+ "eos_token_id": [
6
+ 1,
7
+ 1,
8
+ 106
9
+ ],
10
+ "pad_token_id": 0,
11
+ "top_k": 64,
12
+ "top_p": 0.95,
13
+ "transformers_version": "4.56.1"
14
+ }
proper_wanda_sparsities_2026_08_10_32k/gemma3/calibration_biology/sparsity_0p15/recovered_model/model.safetensors.index.json ADDED
The diff for this file is too large to render. See raw diff
 
proper_wanda_sparsities_2026_08_10_32k/gemma3/calibration_biology/sparsity_0p15/recovered_model/special_tokens_map.json ADDED
@@ -0,0 +1,33 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ {
2
+ "boi_token": "<start_of_image>",
3
+ "bos_token": {
4
+ "content": "<bos>",
5
+ "lstrip": false,
6
+ "normalized": false,
7
+ "rstrip": false,
8
+ "single_word": false
9
+ },
10
+ "eoi_token": "<end_of_image>",
11
+ "eos_token": {
12
+ "content": "<eos>",
13
+ "lstrip": false,
14
+ "normalized": false,
15
+ "rstrip": false,
16
+ "single_word": false
17
+ },
18
+ "image_token": "<image_soft_token>",
19
+ "pad_token": {
20
+ "content": "<pad>",
21
+ "lstrip": false,
22
+ "normalized": false,
23
+ "rstrip": false,
24
+ "single_word": false
25
+ },
26
+ "unk_token": {
27
+ "content": "<unk>",
28
+ "lstrip": false,
29
+ "normalized": false,
30
+ "rstrip": false,
31
+ "single_word": false
32
+ }
33
+ }
proper_wanda_sparsities_2026_08_10_32k/gemma3/calibration_biology/sparsity_0p15/recovered_model/tokenizer_config.json ADDED
The diff for this file is too large to render. See raw diff
 
proper_wanda_sparsities_2026_08_10_32k/gemma3/calibration_biology/sparsity_0p2/checkpoint_results.json ADDED
@@ -0,0 +1,54 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ {
2
+ "model": "gemma3",
3
+ "domain": "biology",
4
+ "sparsity": 0.2,
5
+ "pgd_steps": 2000,
6
+ "training_examples": 32000,
7
+ "per_device_train_batch_size": 2,
8
+ "gradient_accumulation_steps": 8,
9
+ "reference_loss": 3.1282850253582,
10
+ "loss_history": [
11
+ {
12
+ "step": 200,
13
+ "loss": 1.0786311346292496
14
+ },
15
+ {
16
+ "step": 400,
17
+ "loss": 1.0519474625587464
18
+ },
19
+ {
20
+ "step": 600,
21
+ "loss": 1.0393102306127548
22
+ },
23
+ {
24
+ "step": 800,
25
+ "loss": 1.029550707936287
26
+ },
27
+ {
28
+ "step": 1000,
29
+ "loss": 1.021350647211075
30
+ },
31
+ {
32
+ "step": 1200,
33
+ "loss": 1.0177906182408334
34
+ },
35
+ {
36
+ "step": 1400,
37
+ "loss": 1.0143386253714561
38
+ },
39
+ {
40
+ "step": 1600,
41
+ "loss": 1.0128045457601547
42
+ },
43
+ {
44
+ "step": 1800,
45
+ "loss": 1.012356966137886
46
+ },
47
+ {
48
+ "step": 2000,
49
+ "loss": 1.0120375660061836
50
+ }
51
+ ],
52
+ "sparsity_after_pgd": 0.1999965734649123,
53
+ "masked_nonzero_after_pgd": 0
54
+ }
proper_wanda_sparsities_2026_08_10_32k/gemma3/calibration_biology/sparsity_0p2/recovered_model/added_tokens.json ADDED
@@ -0,0 +1,3 @@
 
 
 
 
1
+ {
2
+ "<image_soft_token>": 262144
3
+ }
proper_wanda_sparsities_2026_08_10_32k/gemma3/calibration_biology/sparsity_0p2/recovered_model/chat_template.jinja ADDED
@@ -0,0 +1,47 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ {{ bos_token }}
2
+ {%- if messages[0]['role'] == 'system' -%}
3
+ {%- if messages[0]['content'] is string -%}
4
+ {%- set first_user_prefix = messages[0]['content'] + '
5
+
6
+ ' -%}
7
+ {%- else -%}
8
+ {%- set first_user_prefix = messages[0]['content'][0]['text'] + '
9
+
10
+ ' -%}
11
+ {%- endif -%}
12
+ {%- set loop_messages = messages[1:] -%}
13
+ {%- else -%}
14
+ {%- set first_user_prefix = "" -%}
15
+ {%- set loop_messages = messages -%}
16
+ {%- endif -%}
17
+ {%- for message in loop_messages -%}
18
+ {%- if (message['role'] == 'user') != (loop.index0 % 2 == 0) -%}
19
+ {{ raise_exception("Conversation roles must alternate user/assistant/user/assistant/...") }}
20
+ {%- endif -%}
21
+ {%- if (message['role'] == 'assistant') -%}
22
+ {%- set role = "model" -%}
23
+ {%- else -%}
24
+ {%- set role = message['role'] -%}
25
+ {%- endif -%}
26
+ {{ '<start_of_turn>' + role + '
27
+ ' + (first_user_prefix if loop.first else "") }}
28
+ {%- if message['content'] is string -%}
29
+ {{ message['content'] | trim }}
30
+ {%- elif message['content'] is iterable -%}
31
+ {%- for item in message['content'] -%}
32
+ {%- if item['type'] == 'image' -%}
33
+ {{ '<start_of_image>' }}
34
+ {%- elif item['type'] == 'text' -%}
35
+ {{ item['text'] | trim }}
36
+ {%- endif -%}
37
+ {%- endfor -%}
38
+ {%- else -%}
39
+ {{ raise_exception("Invalid content type") }}
40
+ {%- endif -%}
41
+ {{ '<end_of_turn>
42
+ ' }}
43
+ {%- endfor -%}
44
+ {%- if add_generation_prompt -%}
45
+ {{'<start_of_turn>model
46
+ '}}
47
+ {%- endif -%}
proper_wanda_sparsities_2026_08_10_32k/gemma3/calibration_biology/sparsity_0p2/recovered_model/config.json ADDED
@@ -0,0 +1,111 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ {
2
+ "architectures": [
3
+ "Gemma3ForConditionalGeneration"
4
+ ],
5
+ "boi_token_index": 255999,
6
+ "bos_token_id": 2,
7
+ "dtype": "bfloat16",
8
+ "eoi_token_index": 256000,
9
+ "eos_token_id": 1,
10
+ "image_token_index": 262144,
11
+ "initializer_range": 0.02,
12
+ "mm_tokens_per_image": 256,
13
+ "model_type": "gemma3",
14
+ "pad_token_id": 0,
15
+ "text_config": {
16
+ "_sliding_window_pattern": 6,
17
+ "attention_bias": false,
18
+ "attention_dropout": 0.0,
19
+ "attn_logit_softcapping": null,
20
+ "dtype": "bfloat16",
21
+ "final_logit_softcapping": null,
22
+ "head_dim": 256,
23
+ "hidden_activation": "gelu_pytorch_tanh",
24
+ "hidden_size": 3840,
25
+ "initializer_range": 0.02,
26
+ "intermediate_size": 15360,
27
+ "layer_types": [
28
+ "sliding_attention",
29
+ "sliding_attention",
30
+ "sliding_attention",
31
+ "sliding_attention",
32
+ "sliding_attention",
33
+ "full_attention",
34
+ "sliding_attention",
35
+ "sliding_attention",
36
+ "sliding_attention",
37
+ "sliding_attention",
38
+ "sliding_attention",
39
+ "full_attention",
40
+ "sliding_attention",
41
+ "sliding_attention",
42
+ "sliding_attention",
43
+ "sliding_attention",
44
+ "sliding_attention",
45
+ "full_attention",
46
+ "sliding_attention",
47
+ "sliding_attention",
48
+ "sliding_attention",
49
+ "sliding_attention",
50
+ "sliding_attention",
51
+ "full_attention",
52
+ "sliding_attention",
53
+ "sliding_attention",
54
+ "sliding_attention",
55
+ "sliding_attention",
56
+ "sliding_attention",
57
+ "full_attention",
58
+ "sliding_attention",
59
+ "sliding_attention",
60
+ "sliding_attention",
61
+ "sliding_attention",
62
+ "sliding_attention",
63
+ "full_attention",
64
+ "sliding_attention",
65
+ "sliding_attention",
66
+ "sliding_attention",
67
+ "sliding_attention",
68
+ "sliding_attention",
69
+ "full_attention",
70
+ "sliding_attention",
71
+ "sliding_attention",
72
+ "sliding_attention",
73
+ "sliding_attention",
74
+ "sliding_attention",
75
+ "full_attention"
76
+ ],
77
+ "max_position_embeddings": 131072,
78
+ "model_type": "gemma3_text",
79
+ "num_attention_heads": 16,
80
+ "num_hidden_layers": 48,
81
+ "num_key_value_heads": 8,
82
+ "query_pre_attn_scalar": 256,
83
+ "rms_norm_eps": 1e-06,
84
+ "rope_local_base_freq": 10000.0,
85
+ "rope_scaling": {
86
+ "factor": 8.0,
87
+ "rope_type": "linear"
88
+ },
89
+ "rope_theta": 1000000.0,
90
+ "sliding_window": 1024,
91
+ "use_cache": true,
92
+ "vocab_size": 262208
93
+ },
94
+ "transformers_version": "4.56.1",
95
+ "use_cache": false,
96
+ "vision_config": {
97
+ "attention_dropout": 0.0,
98
+ "dtype": "bfloat16",
99
+ "hidden_act": "gelu_pytorch_tanh",
100
+ "hidden_size": 1152,
101
+ "image_size": 896,
102
+ "intermediate_size": 4304,
103
+ "layer_norm_eps": 1e-06,
104
+ "model_type": "siglip_vision_model",
105
+ "num_attention_heads": 16,
106
+ "num_channels": 3,
107
+ "num_hidden_layers": 27,
108
+ "patch_size": 14,
109
+ "vision_use_head": false
110
+ }
111
+ }
proper_wanda_sparsities_2026_08_10_32k/gemma3/calibration_biology/sparsity_0p2/recovered_model/generation_config.json ADDED
@@ -0,0 +1,14 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ {
2
+ "bos_token_id": 2,
3
+ "cache_implementation": "hybrid",
4
+ "do_sample": true,
5
+ "eos_token_id": [
6
+ 1,
7
+ 1,
8
+ 106
9
+ ],
10
+ "pad_token_id": 0,
11
+ "top_k": 64,
12
+ "top_p": 0.95,
13
+ "transformers_version": "4.56.1"
14
+ }
proper_wanda_sparsities_2026_08_10_32k/gemma3/calibration_biology/sparsity_0p2/recovered_model/model.safetensors.index.json ADDED
The diff for this file is too large to render. See raw diff
 
proper_wanda_sparsities_2026_08_10_32k/gemma3/calibration_biology/sparsity_0p2/recovered_model/special_tokens_map.json ADDED
@@ -0,0 +1,33 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ {
2
+ "boi_token": "<start_of_image>",
3
+ "bos_token": {
4
+ "content": "<bos>",
5
+ "lstrip": false,
6
+ "normalized": false,
7
+ "rstrip": false,
8
+ "single_word": false
9
+ },
10
+ "eoi_token": "<end_of_image>",
11
+ "eos_token": {
12
+ "content": "<eos>",
13
+ "lstrip": false,
14
+ "normalized": false,
15
+ "rstrip": false,
16
+ "single_word": false
17
+ },
18
+ "image_token": "<image_soft_token>",
19
+ "pad_token": {
20
+ "content": "<pad>",
21
+ "lstrip": false,
22
+ "normalized": false,
23
+ "rstrip": false,
24
+ "single_word": false
25
+ },
26
+ "unk_token": {
27
+ "content": "<unk>",
28
+ "lstrip": false,
29
+ "normalized": false,
30
+ "rstrip": false,
31
+ "single_word": false
32
+ }
33
+ }
proper_wanda_sparsities_2026_08_10_32k/gemma3/calibration_biology/sparsity_0p2/recovered_model/tokenizer_config.json ADDED
The diff for this file is too large to render. See raw diff
 
proper_wanda_sparsities_2026_08_10_32k/gemma3/calibration_biology/sparsity_0p25/checkpoint_results.json ADDED
@@ -0,0 +1,54 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ {
2
+ "model": "gemma3",
3
+ "domain": "biology",
4
+ "sparsity": 0.25,
5
+ "pgd_steps": 2000,
6
+ "training_examples": 32000,
7
+ "per_device_train_batch_size": 2,
8
+ "gradient_accumulation_steps": 8,
9
+ "reference_loss": 3.1282850253582,
10
+ "loss_history": [
11
+ {
12
+ "step": 200,
13
+ "loss": 1.0890753972530365
14
+ },
15
+ {
16
+ "step": 400,
17
+ "loss": 1.0626357805728912
18
+ },
19
+ {
20
+ "step": 600,
21
+ "loss": 1.0497788339853287
22
+ },
23
+ {
24
+ "step": 800,
25
+ "loss": 1.0407772344350814
26
+ },
27
+ {
28
+ "step": 1000,
29
+ "loss": 1.0334627419710158
30
+ },
31
+ {
32
+ "step": 1200,
33
+ "loss": 1.0298733413219452
34
+ },
35
+ {
36
+ "step": 1400,
37
+ "loss": 1.0266114890575408
38
+ },
39
+ {
40
+ "step": 1600,
41
+ "loss": 1.025310166478157
42
+ },
43
+ {
44
+ "step": 1800,
45
+ "loss": 1.0249032694101334
46
+ },
47
+ {
48
+ "step": 2000,
49
+ "loss": 1.0250673323869706
50
+ }
51
+ ],
52
+ "sparsity_after_pgd": 0.25,
53
+ "masked_nonzero_after_pgd": 0
54
+ }
proper_wanda_sparsities_2026_08_10_32k/gemma3/calibration_biology/sparsity_0p25/recovered_model/added_tokens.json ADDED
@@ -0,0 +1,3 @@
 
 
 
 
1
+ {
2
+ "<image_soft_token>": 262144
3
+ }
proper_wanda_sparsities_2026_08_10_32k/gemma3/calibration_biology/sparsity_0p25/recovered_model/chat_template.jinja ADDED
@@ -0,0 +1,47 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ {{ bos_token }}
2
+ {%- if messages[0]['role'] == 'system' -%}
3
+ {%- if messages[0]['content'] is string -%}
4
+ {%- set first_user_prefix = messages[0]['content'] + '
5
+
6
+ ' -%}
7
+ {%- else -%}
8
+ {%- set first_user_prefix = messages[0]['content'][0]['text'] + '
9
+
10
+ ' -%}
11
+ {%- endif -%}
12
+ {%- set loop_messages = messages[1:] -%}
13
+ {%- else -%}
14
+ {%- set first_user_prefix = "" -%}
15
+ {%- set loop_messages = messages -%}
16
+ {%- endif -%}
17
+ {%- for message in loop_messages -%}
18
+ {%- if (message['role'] == 'user') != (loop.index0 % 2 == 0) -%}
19
+ {{ raise_exception("Conversation roles must alternate user/assistant/user/assistant/...") }}
20
+ {%- endif -%}
21
+ {%- if (message['role'] == 'assistant') -%}
22
+ {%- set role = "model" -%}
23
+ {%- else -%}
24
+ {%- set role = message['role'] -%}
25
+ {%- endif -%}
26
+ {{ '<start_of_turn>' + role + '
27
+ ' + (first_user_prefix if loop.first else "") }}
28
+ {%- if message['content'] is string -%}
29
+ {{ message['content'] | trim }}
30
+ {%- elif message['content'] is iterable -%}
31
+ {%- for item in message['content'] -%}
32
+ {%- if item['type'] == 'image' -%}
33
+ {{ '<start_of_image>' }}
34
+ {%- elif item['type'] == 'text' -%}
35
+ {{ item['text'] | trim }}
36
+ {%- endif -%}
37
+ {%- endfor -%}
38
+ {%- else -%}
39
+ {{ raise_exception("Invalid content type") }}
40
+ {%- endif -%}
41
+ {{ '<end_of_turn>
42
+ ' }}
43
+ {%- endfor -%}
44
+ {%- if add_generation_prompt -%}
45
+ {{'<start_of_turn>model
46
+ '}}
47
+ {%- endif -%}
proper_wanda_sparsities_2026_08_10_32k/gemma3/calibration_biology/sparsity_0p25/recovered_model/config.json ADDED
@@ -0,0 +1,111 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ {
2
+ "architectures": [
3
+ "Gemma3ForConditionalGeneration"
4
+ ],
5
+ "boi_token_index": 255999,
6
+ "bos_token_id": 2,
7
+ "dtype": "bfloat16",
8
+ "eoi_token_index": 256000,
9
+ "eos_token_id": 1,
10
+ "image_token_index": 262144,
11
+ "initializer_range": 0.02,
12
+ "mm_tokens_per_image": 256,
13
+ "model_type": "gemma3",
14
+ "pad_token_id": 0,
15
+ "text_config": {
16
+ "_sliding_window_pattern": 6,
17
+ "attention_bias": false,
18
+ "attention_dropout": 0.0,
19
+ "attn_logit_softcapping": null,
20
+ "dtype": "bfloat16",
21
+ "final_logit_softcapping": null,
22
+ "head_dim": 256,
23
+ "hidden_activation": "gelu_pytorch_tanh",
24
+ "hidden_size": 3840,
25
+ "initializer_range": 0.02,
26
+ "intermediate_size": 15360,
27
+ "layer_types": [
28
+ "sliding_attention",
29
+ "sliding_attention",
30
+ "sliding_attention",
31
+ "sliding_attention",
32
+ "sliding_attention",
33
+ "full_attention",
34
+ "sliding_attention",
35
+ "sliding_attention",
36
+ "sliding_attention",
37
+ "sliding_attention",
38
+ "sliding_attention",
39
+ "full_attention",
40
+ "sliding_attention",
41
+ "sliding_attention",
42
+ "sliding_attention",
43
+ "sliding_attention",
44
+ "sliding_attention",
45
+ "full_attention",
46
+ "sliding_attention",
47
+ "sliding_attention",
48
+ "sliding_attention",
49
+ "sliding_attention",
50
+ "sliding_attention",
51
+ "full_attention",
52
+ "sliding_attention",
53
+ "sliding_attention",
54
+ "sliding_attention",
55
+ "sliding_attention",
56
+ "sliding_attention",
57
+ "full_attention",
58
+ "sliding_attention",
59
+ "sliding_attention",
60
+ "sliding_attention",
61
+ "sliding_attention",
62
+ "sliding_attention",
63
+ "full_attention",
64
+ "sliding_attention",
65
+ "sliding_attention",
66
+ "sliding_attention",
67
+ "sliding_attention",
68
+ "sliding_attention",
69
+ "full_attention",
70
+ "sliding_attention",
71
+ "sliding_attention",
72
+ "sliding_attention",
73
+ "sliding_attention",
74
+ "sliding_attention",
75
+ "full_attention"
76
+ ],
77
+ "max_position_embeddings": 131072,
78
+ "model_type": "gemma3_text",
79
+ "num_attention_heads": 16,
80
+ "num_hidden_layers": 48,
81
+ "num_key_value_heads": 8,
82
+ "query_pre_attn_scalar": 256,
83
+ "rms_norm_eps": 1e-06,
84
+ "rope_local_base_freq": 10000.0,
85
+ "rope_scaling": {
86
+ "factor": 8.0,
87
+ "rope_type": "linear"
88
+ },
89
+ "rope_theta": 1000000.0,
90
+ "sliding_window": 1024,
91
+ "use_cache": true,
92
+ "vocab_size": 262208
93
+ },
94
+ "transformers_version": "4.56.1",
95
+ "use_cache": false,
96
+ "vision_config": {
97
+ "attention_dropout": 0.0,
98
+ "dtype": "bfloat16",
99
+ "hidden_act": "gelu_pytorch_tanh",
100
+ "hidden_size": 1152,
101
+ "image_size": 896,
102
+ "intermediate_size": 4304,
103
+ "layer_norm_eps": 1e-06,
104
+ "model_type": "siglip_vision_model",
105
+ "num_attention_heads": 16,
106
+ "num_channels": 3,
107
+ "num_hidden_layers": 27,
108
+ "patch_size": 14,
109
+ "vision_use_head": false
110
+ }
111
+ }
proper_wanda_sparsities_2026_08_10_32k/gemma3/calibration_biology/sparsity_0p25/recovered_model/generation_config.json ADDED
@@ -0,0 +1,14 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ {
2
+ "bos_token_id": 2,
3
+ "cache_implementation": "hybrid",
4
+ "do_sample": true,
5
+ "eos_token_id": [
6
+ 1,
7
+ 1,
8
+ 106
9
+ ],
10
+ "pad_token_id": 0,
11
+ "top_k": 64,
12
+ "top_p": 0.95,
13
+ "transformers_version": "4.56.1"
14
+ }
proper_wanda_sparsities_2026_08_10_32k/gemma3/calibration_biology/sparsity_0p25/recovered_model/model.safetensors.index.json ADDED
The diff for this file is too large to render. See raw diff
 
proper_wanda_sparsities_2026_08_10_32k/gemma3/calibration_biology/sparsity_0p25/recovered_model/special_tokens_map.json ADDED
@@ -0,0 +1,33 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ {
2
+ "boi_token": "<start_of_image>",
3
+ "bos_token": {
4
+ "content": "<bos>",
5
+ "lstrip": false,
6
+ "normalized": false,
7
+ "rstrip": false,
8
+ "single_word": false
9
+ },
10
+ "eoi_token": "<end_of_image>",
11
+ "eos_token": {
12
+ "content": "<eos>",
13
+ "lstrip": false,
14
+ "normalized": false,
15
+ "rstrip": false,
16
+ "single_word": false
17
+ },
18
+ "image_token": "<image_soft_token>",
19
+ "pad_token": {
20
+ "content": "<pad>",
21
+ "lstrip": false,
22
+ "normalized": false,
23
+ "rstrip": false,
24
+ "single_word": false
25
+ },
26
+ "unk_token": {
27
+ "content": "<unk>",
28
+ "lstrip": false,
29
+ "normalized": false,
30
+ "rstrip": false,
31
+ "single_word": false
32
+ }
33
+ }
proper_wanda_sparsities_2026_08_10_32k/gemma3/calibration_biology/sparsity_0p25/recovered_model/tokenizer_config.json ADDED
The diff for this file is too large to render. See raw diff
 
proper_wanda_sparsities_2026_08_10_32k/gemma3/calibration_biology/sparsity_0p3/checkpoint_results.json ADDED
@@ -0,0 +1,54 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ {
2
+ "model": "gemma3",
3
+ "domain": "biology",
4
+ "sparsity": 0.3,
5
+ "pgd_steps": 2000,
6
+ "training_examples": 32000,
7
+ "per_device_train_batch_size": 2,
8
+ "gradient_accumulation_steps": 8,
9
+ "reference_loss": 3.1282850253582,
10
+ "loss_history": [
11
+ {
12
+ "step": 200,
13
+ "loss": 1.1052876561880112
14
+ },
15
+ {
16
+ "step": 400,
17
+ "loss": 1.0789224088191987
18
+ },
19
+ {
20
+ "step": 600,
21
+ "loss": 1.0665393048524856
22
+ },
23
+ {
24
+ "step": 800,
25
+ "loss": 1.0572929745912552
26
+ },
27
+ {
28
+ "step": 1000,
29
+ "loss": 1.049967291355133
30
+ },
31
+ {
32
+ "step": 1200,
33
+ "loss": 1.0470821452140808
34
+ },
35
+ {
36
+ "step": 1400,
37
+ "loss": 1.0443577200174332
38
+ },
39
+ {
40
+ "step": 1600,
41
+ "loss": 1.0429766035079957
42
+ },
43
+ {
44
+ "step": 1800,
45
+ "loss": 1.0427227801084518
46
+ },
47
+ {
48
+ "step": 2000,
49
+ "loss": 1.0426533663272857
50
+ }
51
+ ],
52
+ "sparsity_after_pgd": 0.29998629385964914,
53
+ "masked_nonzero_after_pgd": 0
54
+ }
proper_wanda_sparsities_2026_08_10_32k/gemma3/calibration_biology/sparsity_0p3/recovered_model/added_tokens.json ADDED
@@ -0,0 +1,3 @@
 
 
 
 
1
+ {
2
+ "<image_soft_token>": 262144
3
+ }
proper_wanda_sparsities_2026_08_10_32k/gemma3/calibration_biology/sparsity_0p3/recovered_model/chat_template.jinja ADDED
@@ -0,0 +1,47 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ {{ bos_token }}
2
+ {%- if messages[0]['role'] == 'system' -%}
3
+ {%- if messages[0]['content'] is string -%}
4
+ {%- set first_user_prefix = messages[0]['content'] + '
5
+
6
+ ' -%}
7
+ {%- else -%}
8
+ {%- set first_user_prefix = messages[0]['content'][0]['text'] + '
9
+
10
+ ' -%}
11
+ {%- endif -%}
12
+ {%- set loop_messages = messages[1:] -%}
13
+ {%- else -%}
14
+ {%- set first_user_prefix = "" -%}
15
+ {%- set loop_messages = messages -%}
16
+ {%- endif -%}
17
+ {%- for message in loop_messages -%}
18
+ {%- if (message['role'] == 'user') != (loop.index0 % 2 == 0) -%}
19
+ {{ raise_exception("Conversation roles must alternate user/assistant/user/assistant/...") }}
20
+ {%- endif -%}
21
+ {%- if (message['role'] == 'assistant') -%}
22
+ {%- set role = "model" -%}
23
+ {%- else -%}
24
+ {%- set role = message['role'] -%}
25
+ {%- endif -%}
26
+ {{ '<start_of_turn>' + role + '
27
+ ' + (first_user_prefix if loop.first else "") }}
28
+ {%- if message['content'] is string -%}
29
+ {{ message['content'] | trim }}
30
+ {%- elif message['content'] is iterable -%}
31
+ {%- for item in message['content'] -%}
32
+ {%- if item['type'] == 'image' -%}
33
+ {{ '<start_of_image>' }}
34
+ {%- elif item['type'] == 'text' -%}
35
+ {{ item['text'] | trim }}
36
+ {%- endif -%}
37
+ {%- endfor -%}
38
+ {%- else -%}
39
+ {{ raise_exception("Invalid content type") }}
40
+ {%- endif -%}
41
+ {{ '<end_of_turn>
42
+ ' }}
43
+ {%- endfor -%}
44
+ {%- if add_generation_prompt -%}
45
+ {{'<start_of_turn>model
46
+ '}}
47
+ {%- endif -%}
proper_wanda_sparsities_2026_08_10_32k/gemma3/calibration_biology/sparsity_0p3/recovered_model/config.json ADDED
@@ -0,0 +1,111 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ {
2
+ "architectures": [
3
+ "Gemma3ForConditionalGeneration"
4
+ ],
5
+ "boi_token_index": 255999,
6
+ "bos_token_id": 2,
7
+ "dtype": "bfloat16",
8
+ "eoi_token_index": 256000,
9
+ "eos_token_id": 1,
10
+ "image_token_index": 262144,
11
+ "initializer_range": 0.02,
12
+ "mm_tokens_per_image": 256,
13
+ "model_type": "gemma3",
14
+ "pad_token_id": 0,
15
+ "text_config": {
16
+ "_sliding_window_pattern": 6,
17
+ "attention_bias": false,
18
+ "attention_dropout": 0.0,
19
+ "attn_logit_softcapping": null,
20
+ "dtype": "bfloat16",
21
+ "final_logit_softcapping": null,
22
+ "head_dim": 256,
23
+ "hidden_activation": "gelu_pytorch_tanh",
24
+ "hidden_size": 3840,
25
+ "initializer_range": 0.02,
26
+ "intermediate_size": 15360,
27
+ "layer_types": [
28
+ "sliding_attention",
29
+ "sliding_attention",
30
+ "sliding_attention",
31
+ "sliding_attention",
32
+ "sliding_attention",
33
+ "full_attention",
34
+ "sliding_attention",
35
+ "sliding_attention",
36
+ "sliding_attention",
37
+ "sliding_attention",
38
+ "sliding_attention",
39
+ "full_attention",
40
+ "sliding_attention",
41
+ "sliding_attention",
42
+ "sliding_attention",
43
+ "sliding_attention",
44
+ "sliding_attention",
45
+ "full_attention",
46
+ "sliding_attention",
47
+ "sliding_attention",
48
+ "sliding_attention",
49
+ "sliding_attention",
50
+ "sliding_attention",
51
+ "full_attention",
52
+ "sliding_attention",
53
+ "sliding_attention",
54
+ "sliding_attention",
55
+ "sliding_attention",
56
+ "sliding_attention",
57
+ "full_attention",
58
+ "sliding_attention",
59
+ "sliding_attention",
60
+ "sliding_attention",
61
+ "sliding_attention",
62
+ "sliding_attention",
63
+ "full_attention",
64
+ "sliding_attention",
65
+ "sliding_attention",
66
+ "sliding_attention",
67
+ "sliding_attention",
68
+ "sliding_attention",
69
+ "full_attention",
70
+ "sliding_attention",
71
+ "sliding_attention",
72
+ "sliding_attention",
73
+ "sliding_attention",
74
+ "sliding_attention",
75
+ "full_attention"
76
+ ],
77
+ "max_position_embeddings": 131072,
78
+ "model_type": "gemma3_text",
79
+ "num_attention_heads": 16,
80
+ "num_hidden_layers": 48,
81
+ "num_key_value_heads": 8,
82
+ "query_pre_attn_scalar": 256,
83
+ "rms_norm_eps": 1e-06,
84
+ "rope_local_base_freq": 10000.0,
85
+ "rope_scaling": {
86
+ "factor": 8.0,
87
+ "rope_type": "linear"
88
+ },
89
+ "rope_theta": 1000000.0,
90
+ "sliding_window": 1024,
91
+ "use_cache": true,
92
+ "vocab_size": 262208
93
+ },
94
+ "transformers_version": "4.56.1",
95
+ "use_cache": false,
96
+ "vision_config": {
97
+ "attention_dropout": 0.0,
98
+ "dtype": "bfloat16",
99
+ "hidden_act": "gelu_pytorch_tanh",
100
+ "hidden_size": 1152,
101
+ "image_size": 896,
102
+ "intermediate_size": 4304,
103
+ "layer_norm_eps": 1e-06,
104
+ "model_type": "siglip_vision_model",
105
+ "num_attention_heads": 16,
106
+ "num_channels": 3,
107
+ "num_hidden_layers": 27,
108
+ "patch_size": 14,
109
+ "vision_use_head": false
110
+ }
111
+ }
proper_wanda_sparsities_2026_08_10_32k/gemma3/calibration_biology/sparsity_0p3/recovered_model/generation_config.json ADDED
@@ -0,0 +1,14 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ {
2
+ "bos_token_id": 2,
3
+ "cache_implementation": "hybrid",
4
+ "do_sample": true,
5
+ "eos_token_id": [
6
+ 1,
7
+ 1,
8
+ 106
9
+ ],
10
+ "pad_token_id": 0,
11
+ "top_k": 64,
12
+ "top_p": 0.95,
13
+ "transformers_version": "4.56.1"
14
+ }
proper_wanda_sparsities_2026_08_10_32k/gemma3/calibration_biology/sparsity_0p3/recovered_model/model.safetensors.index.json ADDED
The diff for this file is too large to render. See raw diff
 
proper_wanda_sparsities_2026_08_10_32k/gemma3/calibration_biology/sparsity_0p3/recovered_model/special_tokens_map.json ADDED
@@ -0,0 +1,33 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ {
2
+ "boi_token": "<start_of_image>",
3
+ "bos_token": {
4
+ "content": "<bos>",
5
+ "lstrip": false,
6
+ "normalized": false,
7
+ "rstrip": false,
8
+ "single_word": false
9
+ },
10
+ "eoi_token": "<end_of_image>",
11
+ "eos_token": {
12
+ "content": "<eos>",
13
+ "lstrip": false,
14
+ "normalized": false,
15
+ "rstrip": false,
16
+ "single_word": false
17
+ },
18
+ "image_token": "<image_soft_token>",
19
+ "pad_token": {
20
+ "content": "<pad>",
21
+ "lstrip": false,
22
+ "normalized": false,
23
+ "rstrip": false,
24
+ "single_word": false
25
+ },
26
+ "unk_token": {
27
+ "content": "<unk>",
28
+ "lstrip": false,
29
+ "normalized": false,
30
+ "rstrip": false,
31
+ "single_word": false
32
+ }
33
+ }
proper_wanda_sparsities_2026_08_10_32k/gemma3/calibration_biology/sparsity_0p3/recovered_model/tokenizer_config.json ADDED
The diff for this file is too large to render. See raw diff
 
proper_wanda_sparsities_2026_08_10_32k/gemma3/calibration_biology/sparsity_0p35/checkpoint_results.json ADDED
@@ -0,0 +1,54 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ {
2
+ "model": "gemma3",
3
+ "domain": "biology",
4
+ "sparsity": 0.35,
5
+ "pgd_steps": 2000,
6
+ "training_examples": 32000,
7
+ "per_device_train_batch_size": 2,
8
+ "gradient_accumulation_steps": 8,
9
+ "reference_loss": 3.1282850253582,
10
+ "loss_history": [
11
+ {
12
+ "step": 200,
13
+ "loss": 1.1277863103151322
14
+ },
15
+ {
16
+ "step": 400,
17
+ "loss": 1.1005483210086822
18
+ },
19
+ {
20
+ "step": 600,
21
+ "loss": 1.0880144131183624
22
+ },
23
+ {
24
+ "step": 800,
25
+ "loss": 1.0795432889461518
26
+ },
27
+ {
28
+ "step": 1000,
29
+ "loss": 1.0732172924280166
30
+ },
31
+ {
32
+ "step": 1200,
33
+ "loss": 1.0699898260831833
34
+ },
35
+ {
36
+ "step": 1400,
37
+ "loss": 1.0675747334957122
38
+ },
39
+ {
40
+ "step": 1600,
41
+ "loss": 1.0661312192678452
42
+ },
43
+ {
44
+ "step": 1800,
45
+ "loss": 1.0659805464744567
46
+ },
47
+ {
48
+ "step": 2000,
49
+ "loss": 1.0658278757333755
50
+ }
51
+ ],
52
+ "sparsity_after_pgd": 0.34998972039473686,
53
+ "masked_nonzero_after_pgd": 0
54
+ }
proper_wanda_sparsities_2026_08_10_32k/gemma3/calibration_biology/sparsity_0p35/recovered_model/added_tokens.json ADDED
@@ -0,0 +1,3 @@
 
 
 
 
1
+ {
2
+ "<image_soft_token>": 262144
3
+ }