jbostock commited on
Commit
554f02a
·
verified ·
1 Parent(s): fd611f2

Full-parameter AFT (midtrain4): full_aft_midtrain4/balanced

Browse files
This view is limited to 50 files because it contains too many changes.   See raw diff
Files changed (50) hide show
  1. .gitattributes +8 -0
  2. full_aft_midtrain4/balanced/checkpoint-128/chat_template.jinja +47 -0
  3. full_aft_midtrain4/balanced/checkpoint-128/config.json +125 -0
  4. full_aft_midtrain4/balanced/checkpoint-128/generation_config.json +13 -0
  5. full_aft_midtrain4/balanced/checkpoint-128/model.safetensors +3 -0
  6. full_aft_midtrain4/balanced/checkpoint-128/preprocessor_config.json +29 -0
  7. full_aft_midtrain4/balanced/checkpoint-128/processor_config.json +4 -0
  8. full_aft_midtrain4/balanced/checkpoint-128/tokenizer.json +3 -0
  9. full_aft_midtrain4/balanced/checkpoint-128/tokenizer_config.json +26 -0
  10. full_aft_midtrain4/balanced/checkpoint-128/tokens_state.json +1 -0
  11. full_aft_midtrain4/balanced/checkpoint-128/trainer_state.json +1826 -0
  12. full_aft_midtrain4/balanced/checkpoint-128/training_args.bin +3 -0
  13. full_aft_midtrain4/balanced/checkpoint-16/chat_template.jinja +47 -0
  14. full_aft_midtrain4/balanced/checkpoint-16/config.json +125 -0
  15. full_aft_midtrain4/balanced/checkpoint-16/generation_config.json +13 -0
  16. full_aft_midtrain4/balanced/checkpoint-16/model.safetensors +3 -0
  17. full_aft_midtrain4/balanced/checkpoint-16/preprocessor_config.json +29 -0
  18. full_aft_midtrain4/balanced/checkpoint-16/processor_config.json +4 -0
  19. full_aft_midtrain4/balanced/checkpoint-16/tokenizer.json +3 -0
  20. full_aft_midtrain4/balanced/checkpoint-16/tokenizer_config.json +26 -0
  21. full_aft_midtrain4/balanced/checkpoint-16/tokens_state.json +1 -0
  22. full_aft_midtrain4/balanced/checkpoint-16/trainer_state.json +258 -0
  23. full_aft_midtrain4/balanced/checkpoint-16/training_args.bin +3 -0
  24. full_aft_midtrain4/balanced/checkpoint-256/chat_template.jinja +47 -0
  25. full_aft_midtrain4/balanced/checkpoint-256/config.json +125 -0
  26. full_aft_midtrain4/balanced/checkpoint-256/generation_config.json +13 -0
  27. full_aft_midtrain4/balanced/checkpoint-256/model.safetensors +3 -0
  28. full_aft_midtrain4/balanced/checkpoint-256/preprocessor_config.json +29 -0
  29. full_aft_midtrain4/balanced/checkpoint-256/processor_config.json +4 -0
  30. full_aft_midtrain4/balanced/checkpoint-256/tokenizer.json +3 -0
  31. full_aft_midtrain4/balanced/checkpoint-256/tokenizer_config.json +26 -0
  32. full_aft_midtrain4/balanced/checkpoint-256/tokens_state.json +1 -0
  33. full_aft_midtrain4/balanced/checkpoint-256/trainer_state.json +0 -0
  34. full_aft_midtrain4/balanced/checkpoint-256/training_args.bin +3 -0
  35. full_aft_midtrain4/balanced/checkpoint-32/chat_template.jinja +47 -0
  36. full_aft_midtrain4/balanced/checkpoint-32/config.json +125 -0
  37. full_aft_midtrain4/balanced/checkpoint-32/generation_config.json +13 -0
  38. full_aft_midtrain4/balanced/checkpoint-32/model.safetensors +3 -0
  39. full_aft_midtrain4/balanced/checkpoint-32/preprocessor_config.json +29 -0
  40. full_aft_midtrain4/balanced/checkpoint-32/processor_config.json +4 -0
  41. full_aft_midtrain4/balanced/checkpoint-32/tokenizer.json +3 -0
  42. full_aft_midtrain4/balanced/checkpoint-32/tokenizer_config.json +26 -0
  43. full_aft_midtrain4/balanced/checkpoint-32/tokens_state.json +1 -0
  44. full_aft_midtrain4/balanced/checkpoint-32/trainer_state.json +482 -0
  45. full_aft_midtrain4/balanced/checkpoint-32/training_args.bin +3 -0
  46. full_aft_midtrain4/balanced/checkpoint-4/chat_template.jinja +47 -0
  47. full_aft_midtrain4/balanced/checkpoint-4/config.json +125 -0
  48. full_aft_midtrain4/balanced/checkpoint-4/generation_config.json +13 -0
  49. full_aft_midtrain4/balanced/checkpoint-4/model.safetensors +3 -0
  50. full_aft_midtrain4/balanced/checkpoint-4/preprocessor_config.json +29 -0
.gitattributes CHANGED
@@ -111,3 +111,11 @@ full_aft_midtrain4/coin4/checkpoint-4/tokenizer.json filter=lfs diff=lfs merge=l
111
  full_aft_midtrain4/coin4/checkpoint-512/tokenizer.json filter=lfs diff=lfs merge=lfs -text
112
  full_aft_midtrain4/coin4/checkpoint-64/tokenizer.json filter=lfs diff=lfs merge=lfs -text
113
  full_aft_midtrain4/coin4/checkpoint-8/tokenizer.json filter=lfs diff=lfs merge=lfs -text
 
 
 
 
 
 
 
 
 
111
  full_aft_midtrain4/coin4/checkpoint-512/tokenizer.json filter=lfs diff=lfs merge=lfs -text
112
  full_aft_midtrain4/coin4/checkpoint-64/tokenizer.json filter=lfs diff=lfs merge=lfs -text
113
  full_aft_midtrain4/coin4/checkpoint-8/tokenizer.json filter=lfs diff=lfs merge=lfs -text
114
+ full_aft_midtrain4/balanced/checkpoint-128/tokenizer.json filter=lfs diff=lfs merge=lfs -text
115
+ full_aft_midtrain4/balanced/checkpoint-16/tokenizer.json filter=lfs diff=lfs merge=lfs -text
116
+ full_aft_midtrain4/balanced/checkpoint-256/tokenizer.json filter=lfs diff=lfs merge=lfs -text
117
+ full_aft_midtrain4/balanced/checkpoint-32/tokenizer.json filter=lfs diff=lfs merge=lfs -text
118
+ full_aft_midtrain4/balanced/checkpoint-4/tokenizer.json filter=lfs diff=lfs merge=lfs -text
119
+ full_aft_midtrain4/balanced/checkpoint-512/tokenizer.json filter=lfs diff=lfs merge=lfs -text
120
+ full_aft_midtrain4/balanced/checkpoint-64/tokenizer.json filter=lfs diff=lfs merge=lfs -text
121
+ full_aft_midtrain4/balanced/checkpoint-8/tokenizer.json filter=lfs diff=lfs merge=lfs -text
full_aft_midtrain4/balanced/checkpoint-128/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 -%}
full_aft_midtrain4/balanced/checkpoint-128/config.json ADDED
@@ -0,0 +1,125 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
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
+ "bos_token_id": 2,
21
+ "cache_implementation": "hybrid",
22
+ "dtype": "bfloat16",
23
+ "eos_token_id": 1,
24
+ "final_logit_softcapping": null,
25
+ "head_dim": 256,
26
+ "hidden_activation": "gelu_pytorch_tanh",
27
+ "hidden_size": 3840,
28
+ "initializer_range": 0.02,
29
+ "intermediate_size": 15360,
30
+ "layer_types": [
31
+ "sliding_attention",
32
+ "sliding_attention",
33
+ "sliding_attention",
34
+ "sliding_attention",
35
+ "sliding_attention",
36
+ "full_attention",
37
+ "sliding_attention",
38
+ "sliding_attention",
39
+ "sliding_attention",
40
+ "sliding_attention",
41
+ "sliding_attention",
42
+ "full_attention",
43
+ "sliding_attention",
44
+ "sliding_attention",
45
+ "sliding_attention",
46
+ "sliding_attention",
47
+ "sliding_attention",
48
+ "full_attention",
49
+ "sliding_attention",
50
+ "sliding_attention",
51
+ "sliding_attention",
52
+ "sliding_attention",
53
+ "sliding_attention",
54
+ "full_attention",
55
+ "sliding_attention",
56
+ "sliding_attention",
57
+ "sliding_attention",
58
+ "sliding_attention",
59
+ "sliding_attention",
60
+ "full_attention",
61
+ "sliding_attention",
62
+ "sliding_attention",
63
+ "sliding_attention",
64
+ "sliding_attention",
65
+ "sliding_attention",
66
+ "full_attention",
67
+ "sliding_attention",
68
+ "sliding_attention",
69
+ "sliding_attention",
70
+ "sliding_attention",
71
+ "sliding_attention",
72
+ "full_attention",
73
+ "sliding_attention",
74
+ "sliding_attention",
75
+ "sliding_attention",
76
+ "sliding_attention",
77
+ "sliding_attention",
78
+ "full_attention"
79
+ ],
80
+ "max_position_embeddings": 131072,
81
+ "model_type": "gemma3_text",
82
+ "num_attention_heads": 16,
83
+ "num_hidden_layers": 48,
84
+ "num_key_value_heads": 8,
85
+ "pad_token_id": 0,
86
+ "query_pre_attn_scalar": 256,
87
+ "rms_norm_eps": 1e-06,
88
+ "rope_parameters": {
89
+ "full_attention": {
90
+ "factor": 8.0,
91
+ "rope_theta": 1000000.0,
92
+ "rope_type": "linear"
93
+ },
94
+ "sliding_attention": {
95
+ "rope_theta": 10000.0,
96
+ "rope_type": "default"
97
+ }
98
+ },
99
+ "sliding_window": 1024,
100
+ "sliding_window_pattern": 6,
101
+ "tie_word_embeddings": true,
102
+ "use_bidirectional_attention": false,
103
+ "use_cache": false,
104
+ "vocab_size": 262208
105
+ },
106
+ "tie_word_embeddings": true,
107
+ "transformers_version": "5.9.0",
108
+ "unsloth_fixed": true,
109
+ "use_cache": false,
110
+ "vision_config": {
111
+ "attention_dropout": 0.0,
112
+ "dtype": "bfloat16",
113
+ "hidden_act": "gelu_pytorch_tanh",
114
+ "hidden_size": 1152,
115
+ "image_size": 896,
116
+ "intermediate_size": 4304,
117
+ "layer_norm_eps": 1e-06,
118
+ "model_type": "siglip_vision_model",
119
+ "num_attention_heads": 16,
120
+ "num_channels": 3,
121
+ "num_hidden_layers": 27,
122
+ "patch_size": 14,
123
+ "vision_use_head": false
124
+ }
125
+ }
full_aft_midtrain4/balanced/checkpoint-128/generation_config.json ADDED
@@ -0,0 +1,13 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ {
2
+ "bos_token_id": 2,
3
+ "cache_implementation": "hybrid",
4
+ "do_sample": true,
5
+ "eos_token_id": [
6
+ 1,
7
+ 106
8
+ ],
9
+ "pad_token_id": 0,
10
+ "top_k": 64,
11
+ "top_p": 0.95,
12
+ "transformers_version": "5.9.0"
13
+ }
full_aft_midtrain4/balanced/checkpoint-128/model.safetensors ADDED
@@ -0,0 +1,3 @@
 
 
 
 
1
+ version https://git-lfs.github.com/spec/v1
2
+ oid sha256:8a5e0af959288727d79345f566740842ad2f95dd4de24a5cf311bee9dbda9635
3
+ size 26388552360
full_aft_midtrain4/balanced/checkpoint-128/preprocessor_config.json ADDED
@@ -0,0 +1,29 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ {
2
+ "do_convert_rgb": null,
3
+ "do_normalize": true,
4
+ "do_pan_and_scan": null,
5
+ "do_rescale": true,
6
+ "do_resize": true,
7
+ "image_mean": [
8
+ 0.5,
9
+ 0.5,
10
+ 0.5
11
+ ],
12
+ "image_processor_type": "Gemma3ImageProcessor",
13
+ "image_seq_length": 256,
14
+ "image_std": [
15
+ 0.5,
16
+ 0.5,
17
+ 0.5
18
+ ],
19
+ "pan_and_scan_max_num_crops": null,
20
+ "pan_and_scan_min_crop_size": null,
21
+ "pan_and_scan_min_ratio_to_activate": null,
22
+ "processor_class": "Gemma3Processor",
23
+ "resample": 2,
24
+ "rescale_factor": 0.00392156862745098,
25
+ "size": {
26
+ "height": 896,
27
+ "width": 896
28
+ }
29
+ }
full_aft_midtrain4/balanced/checkpoint-128/processor_config.json ADDED
@@ -0,0 +1,4 @@
 
 
 
 
 
1
+ {
2
+ "image_seq_length": 256,
3
+ "processor_class": "Gemma3Processor"
4
+ }
full_aft_midtrain4/balanced/checkpoint-128/tokenizer.json ADDED
@@ -0,0 +1,3 @@
 
 
 
 
1
+ version https://git-lfs.github.com/spec/v1
2
+ oid sha256:daab2354f8a74e70d70b4d1f804939b68a8c9624dd06cb7858e52dd8970e9726
3
+ size 33384567
full_aft_midtrain4/balanced/checkpoint-128/tokenizer_config.json ADDED
@@ -0,0 +1,26 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ {
2
+ "backend": "tokenizers",
3
+ "boi_token": "<start_of_image>",
4
+ "bos_token": "<bos>",
5
+ "clean_up_tokenization_spaces": false,
6
+ "eoi_token": "<end_of_image>",
7
+ "eos_token": "<eos>",
8
+ "image_token": "<image_soft_token>",
9
+ "is_local": false,
10
+ "local_files_only": false,
11
+ "mask_token": "<mask>",
12
+ "model_max_length": 131072,
13
+ "model_specific_special_tokens": {
14
+ "boi_token": "<start_of_image>",
15
+ "eoi_token": "<end_of_image>",
16
+ "image_token": "<image_soft_token>"
17
+ },
18
+ "pad_token": "<pad>",
19
+ "padding_side": "left",
20
+ "processor_class": "Gemma3Processor",
21
+ "sp_model_kwargs": null,
22
+ "spaces_between_special_tokens": false,
23
+ "tokenizer_class": "GemmaTokenizer",
24
+ "unk_token": "<unk>",
25
+ "use_default_system_prompt": false
26
+ }
full_aft_midtrain4/balanced/checkpoint-128/tokens_state.json ADDED
@@ -0,0 +1 @@
 
 
1
+ {"total": 2796262, "trainable": 58534}
full_aft_midtrain4/balanced/checkpoint-128/trainer_state.json ADDED
@@ -0,0 +1,1826 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ {
2
+ "best_global_step": null,
3
+ "best_metric": null,
4
+ "best_model_checkpoint": null,
5
+ "epoch": 0.5,
6
+ "eval_steps": 500,
7
+ "global_step": 128,
8
+ "is_hyper_param_search": false,
9
+ "is_local_process_zero": true,
10
+ "is_world_process_zero": true,
11
+ "log_history": [
12
+ {
13
+ "epoch": 0.00390625,
14
+ "grad_norm": 17.375,
15
+ "learning_rate": 5e-06,
16
+ "loss": 0.11627960205078125,
17
+ "memory/device_reserved (GiB)": 46.13,
18
+ "memory/max_active (GiB)": 35.94,
19
+ "memory/max_allocated (GiB)": 35.94,
20
+ "ppl": 1.12331,
21
+ "step": 1,
22
+ "tokens/total": 22614,
23
+ "tokens/train_per_sec_per_gpu": 1.99,
24
+ "tokens/trainable": 470
25
+ },
26
+ {
27
+ "epoch": 0.0078125,
28
+ "grad_norm": 1472.0,
29
+ "learning_rate": 5e-06,
30
+ "loss": 0.1474212110042572,
31
+ "memory/device_reserved (GiB)": 58.08,
32
+ "memory/max_active (GiB)": 46.9,
33
+ "memory/max_allocated (GiB)": 46.9,
34
+ "ppl": 1.15884,
35
+ "step": 2,
36
+ "tokens/total": 44116,
37
+ "tokens/train_per_sec_per_gpu": 4.14,
38
+ "tokens/trainable": 922
39
+ },
40
+ {
41
+ "epoch": 0.01171875,
42
+ "grad_norm": 57.75,
43
+ "learning_rate": 5e-06,
44
+ "loss": 0.2154746651649475,
45
+ "memory/device_reserved (GiB)": 58.08,
46
+ "memory/max_active (GiB)": 46.9,
47
+ "memory/max_allocated (GiB)": 46.9,
48
+ "ppl": 1.24045,
49
+ "step": 3,
50
+ "tokens/total": 67092,
51
+ "tokens/train_per_sec_per_gpu": 3.77,
52
+ "tokens/trainable": 1396
53
+ },
54
+ {
55
+ "epoch": 0.015625,
56
+ "grad_norm": 37.75,
57
+ "learning_rate": 5e-06,
58
+ "loss": 0.10307750105857849,
59
+ "memory/device_reserved (GiB)": 58.09,
60
+ "memory/max_active (GiB)": 46.9,
61
+ "memory/max_allocated (GiB)": 46.9,
62
+ "ppl": 1.10858,
63
+ "step": 4,
64
+ "tokens/total": 88800,
65
+ "tokens/train_per_sec_per_gpu": 5.16,
66
+ "tokens/trainable": 1850
67
+ },
68
+ {
69
+ "epoch": 0.01953125,
70
+ "grad_norm": 25.0,
71
+ "learning_rate": 5e-06,
72
+ "loss": 0.10710409283638,
73
+ "memory/device_reserved (GiB)": 58.09,
74
+ "memory/max_active (GiB)": 46.9,
75
+ "memory/max_allocated (GiB)": 46.9,
76
+ "ppl": 1.11305,
77
+ "step": 5,
78
+ "tokens/total": 110742,
79
+ "tokens/train_per_sec_per_gpu": 3.92,
80
+ "tokens/trainable": 2302
81
+ },
82
+ {
83
+ "epoch": 0.0234375,
84
+ "grad_norm": 16.25,
85
+ "learning_rate": 5e-06,
86
+ "loss": 0.05220422148704529,
87
+ "memory/device_reserved (GiB)": 57.19,
88
+ "memory/max_active (GiB)": 46.9,
89
+ "memory/max_allocated (GiB)": 46.9,
90
+ "ppl": 1.05359,
91
+ "step": 6,
92
+ "tokens/total": 131497,
93
+ "tokens/train_per_sec_per_gpu": 4.45,
94
+ "tokens/trainable": 2742
95
+ },
96
+ {
97
+ "epoch": 0.02734375,
98
+ "grad_norm": 18.0,
99
+ "learning_rate": 5e-06,
100
+ "loss": 0.06188321113586426,
101
+ "memory/device_reserved (GiB)": 59.04,
102
+ "memory/max_active (GiB)": 46.9,
103
+ "memory/max_allocated (GiB)": 46.9,
104
+ "ppl": 1.06384,
105
+ "step": 7,
106
+ "tokens/total": 153846,
107
+ "tokens/train_per_sec_per_gpu": 4.99,
108
+ "tokens/trainable": 3208
109
+ },
110
+ {
111
+ "epoch": 0.03125,
112
+ "grad_norm": 16.75,
113
+ "learning_rate": 5e-06,
114
+ "loss": 0.036078453063964844,
115
+ "memory/device_reserved (GiB)": 60.12,
116
+ "memory/max_active (GiB)": 46.9,
117
+ "memory/max_allocated (GiB)": 46.9,
118
+ "ppl": 1.03674,
119
+ "step": 8,
120
+ "tokens/total": 175713,
121
+ "tokens/train_per_sec_per_gpu": 4.47,
122
+ "tokens/trainable": 3655
123
+ },
124
+ {
125
+ "epoch": 0.03515625,
126
+ "grad_norm": 13.5625,
127
+ "learning_rate": 5e-06,
128
+ "loss": 0.05783689022064209,
129
+ "memory/device_reserved (GiB)": 60.12,
130
+ "memory/max_active (GiB)": 46.9,
131
+ "memory/max_allocated (GiB)": 46.9,
132
+ "ppl": 1.05954,
133
+ "step": 9,
134
+ "tokens/total": 196358,
135
+ "tokens/train_per_sec_per_gpu": 3.39,
136
+ "tokens/trainable": 4091
137
+ },
138
+ {
139
+ "epoch": 0.0390625,
140
+ "grad_norm": 16.375,
141
+ "learning_rate": 5e-06,
142
+ "loss": 0.05202341079711914,
143
+ "memory/device_reserved (GiB)": 60.1,
144
+ "memory/max_active (GiB)": 46.9,
145
+ "memory/max_allocated (GiB)": 46.9,
146
+ "ppl": 1.0534,
147
+ "step": 10,
148
+ "tokens/total": 218622,
149
+ "tokens/train_per_sec_per_gpu": 4.33,
150
+ "tokens/trainable": 4553
151
+ },
152
+ {
153
+ "epoch": 0.04296875,
154
+ "grad_norm": 22.5,
155
+ "learning_rate": 5e-06,
156
+ "loss": 0.03920435905456543,
157
+ "memory/device_reserved (GiB)": 60.1,
158
+ "memory/max_active (GiB)": 46.9,
159
+ "memory/max_allocated (GiB)": 46.9,
160
+ "ppl": 1.03998,
161
+ "step": 11,
162
+ "tokens/total": 239835,
163
+ "tokens/train_per_sec_per_gpu": 4.54,
164
+ "tokens/trainable": 4992
165
+ },
166
+ {
167
+ "epoch": 0.046875,
168
+ "grad_norm": 43.75,
169
+ "learning_rate": 5e-06,
170
+ "loss": 0.0486750602722168,
171
+ "memory/device_reserved (GiB)": 60.1,
172
+ "memory/max_active (GiB)": 46.9,
173
+ "memory/max_allocated (GiB)": 46.9,
174
+ "ppl": 1.04988,
175
+ "step": 12,
176
+ "tokens/total": 264428,
177
+ "tokens/train_per_sec_per_gpu": 4.93,
178
+ "tokens/trainable": 5494
179
+ },
180
+ {
181
+ "epoch": 0.05078125,
182
+ "grad_norm": 7.59375,
183
+ "learning_rate": 5e-06,
184
+ "loss": 0.023757457733154297,
185
+ "memory/device_reserved (GiB)": 60.1,
186
+ "memory/max_active (GiB)": 46.9,
187
+ "memory/max_allocated (GiB)": 46.9,
188
+ "ppl": 1.02404,
189
+ "step": 13,
190
+ "tokens/total": 284859,
191
+ "tokens/train_per_sec_per_gpu": 3.95,
192
+ "tokens/trainable": 5933
193
+ },
194
+ {
195
+ "epoch": 0.0546875,
196
+ "grad_norm": 19.0,
197
+ "learning_rate": 5e-06,
198
+ "loss": 0.04330146312713623,
199
+ "memory/device_reserved (GiB)": 60.1,
200
+ "memory/max_active (GiB)": 46.9,
201
+ "memory/max_allocated (GiB)": 46.9,
202
+ "ppl": 1.04425,
203
+ "step": 14,
204
+ "tokens/total": 305640,
205
+ "tokens/train_per_sec_per_gpu": 3.17,
206
+ "tokens/trainable": 6369
207
+ },
208
+ {
209
+ "epoch": 0.05859375,
210
+ "grad_norm": 9.6875,
211
+ "learning_rate": 5e-06,
212
+ "loss": 0.03816962242126465,
213
+ "memory/device_reserved (GiB)": 60.1,
214
+ "memory/max_active (GiB)": 46.9,
215
+ "memory/max_allocated (GiB)": 46.9,
216
+ "ppl": 1.03891,
217
+ "step": 15,
218
+ "tokens/total": 327063,
219
+ "tokens/train_per_sec_per_gpu": 4.35,
220
+ "tokens/trainable": 6820
221
+ },
222
+ {
223
+ "epoch": 0.0625,
224
+ "grad_norm": 11.8125,
225
+ "learning_rate": 5e-06,
226
+ "loss": 0.03375864028930664,
227
+ "memory/device_reserved (GiB)": 60.1,
228
+ "memory/max_active (GiB)": 46.9,
229
+ "memory/max_allocated (GiB)": 46.9,
230
+ "ppl": 1.03433,
231
+ "step": 16,
232
+ "tokens/total": 347838,
233
+ "tokens/train_per_sec_per_gpu": 3.77,
234
+ "tokens/trainable": 7258
235
+ },
236
+ {
237
+ "epoch": 0.06640625,
238
+ "grad_norm": 12.4375,
239
+ "learning_rate": 5e-06,
240
+ "loss": 0.023131251335144043,
241
+ "memory/device_reserved (GiB)": 60.1,
242
+ "memory/max_active (GiB)": 46.9,
243
+ "memory/max_allocated (GiB)": 46.9,
244
+ "ppl": 1.0234,
245
+ "step": 17,
246
+ "tokens/total": 369888,
247
+ "tokens/train_per_sec_per_gpu": 4.4,
248
+ "tokens/trainable": 7710
249
+ },
250
+ {
251
+ "epoch": 0.0703125,
252
+ "grad_norm": 10.75,
253
+ "learning_rate": 5e-06,
254
+ "loss": 0.02522754669189453,
255
+ "memory/device_reserved (GiB)": 57.25,
256
+ "memory/max_active (GiB)": 46.9,
257
+ "memory/max_allocated (GiB)": 46.9,
258
+ "ppl": 1.02555,
259
+ "step": 18,
260
+ "tokens/total": 392455,
261
+ "tokens/train_per_sec_per_gpu": 3.72,
262
+ "tokens/trainable": 8174
263
+ },
264
+ {
265
+ "epoch": 0.07421875,
266
+ "grad_norm": 8.4375,
267
+ "learning_rate": 5e-06,
268
+ "loss": 0.028311967849731445,
269
+ "memory/device_reserved (GiB)": 58.4,
270
+ "memory/max_active (GiB)": 46.9,
271
+ "memory/max_allocated (GiB)": 46.9,
272
+ "ppl": 1.02872,
273
+ "step": 19,
274
+ "tokens/total": 413549,
275
+ "tokens/train_per_sec_per_gpu": 4.51,
276
+ "tokens/trainable": 8617
277
+ },
278
+ {
279
+ "epoch": 0.078125,
280
+ "grad_norm": 4.34375,
281
+ "learning_rate": 5e-06,
282
+ "loss": 0.012633323669433594,
283
+ "memory/device_reserved (GiB)": 58.4,
284
+ "memory/max_active (GiB)": 46.9,
285
+ "memory/max_allocated (GiB)": 46.9,
286
+ "ppl": 1.01271,
287
+ "step": 20,
288
+ "tokens/total": 434128,
289
+ "tokens/train_per_sec_per_gpu": 3.7,
290
+ "tokens/trainable": 9037
291
+ },
292
+ {
293
+ "epoch": 0.08203125,
294
+ "grad_norm": 3.34375,
295
+ "learning_rate": 5e-06,
296
+ "loss": 0.010331928730010986,
297
+ "memory/device_reserved (GiB)": 58.4,
298
+ "memory/max_active (GiB)": 46.9,
299
+ "memory/max_allocated (GiB)": 46.9,
300
+ "ppl": 1.01039,
301
+ "step": 21,
302
+ "tokens/total": 456131,
303
+ "tokens/train_per_sec_per_gpu": 5.15,
304
+ "tokens/trainable": 9503
305
+ },
306
+ {
307
+ "epoch": 0.0859375,
308
+ "grad_norm": 8.6875,
309
+ "learning_rate": 5e-06,
310
+ "loss": 0.011040687561035156,
311
+ "memory/device_reserved (GiB)": 58.4,
312
+ "memory/max_active (GiB)": 46.9,
313
+ "memory/max_allocated (GiB)": 46.9,
314
+ "ppl": 1.0111,
315
+ "step": 22,
316
+ "tokens/total": 478707,
317
+ "tokens/train_per_sec_per_gpu": 4.45,
318
+ "tokens/trainable": 9972
319
+ },
320
+ {
321
+ "epoch": 0.08984375,
322
+ "grad_norm": 17.75,
323
+ "learning_rate": 5e-06,
324
+ "loss": 0.030008912086486816,
325
+ "memory/device_reserved (GiB)": 58.4,
326
+ "memory/max_active (GiB)": 46.9,
327
+ "memory/max_allocated (GiB)": 46.9,
328
+ "ppl": 1.03046,
329
+ "step": 23,
330
+ "tokens/total": 501425,
331
+ "tokens/train_per_sec_per_gpu": 4.9,
332
+ "tokens/trainable": 10447
333
+ },
334
+ {
335
+ "epoch": 0.09375,
336
+ "grad_norm": 6.6875,
337
+ "learning_rate": 5e-06,
338
+ "loss": 0.007420599460601807,
339
+ "memory/device_reserved (GiB)": 60.14,
340
+ "memory/max_active (GiB)": 46.9,
341
+ "memory/max_allocated (GiB)": 46.9,
342
+ "ppl": 1.00745,
343
+ "step": 24,
344
+ "tokens/total": 522885,
345
+ "tokens/train_per_sec_per_gpu": 5.15,
346
+ "tokens/trainable": 10899
347
+ },
348
+ {
349
+ "epoch": 0.09765625,
350
+ "grad_norm": 24.75,
351
+ "learning_rate": 5e-06,
352
+ "loss": 0.05265447497367859,
353
+ "memory/device_reserved (GiB)": 60.14,
354
+ "memory/max_active (GiB)": 46.9,
355
+ "memory/max_allocated (GiB)": 46.9,
356
+ "ppl": 1.05407,
357
+ "step": 25,
358
+ "tokens/total": 544886,
359
+ "tokens/train_per_sec_per_gpu": 4.45,
360
+ "tokens/trainable": 11360
361
+ },
362
+ {
363
+ "epoch": 0.1015625,
364
+ "grad_norm": 12.625,
365
+ "learning_rate": 5e-06,
366
+ "loss": 0.04731038212776184,
367
+ "memory/device_reserved (GiB)": 60.14,
368
+ "memory/max_active (GiB)": 46.9,
369
+ "memory/max_allocated (GiB)": 46.9,
370
+ "ppl": 1.04845,
371
+ "step": 26,
372
+ "tokens/total": 564964,
373
+ "tokens/train_per_sec_per_gpu": 3.36,
374
+ "tokens/trainable": 11775
375
+ },
376
+ {
377
+ "epoch": 0.10546875,
378
+ "grad_norm": 13.1875,
379
+ "learning_rate": 5e-06,
380
+ "loss": 0.014921486377716064,
381
+ "memory/device_reserved (GiB)": 60.14,
382
+ "memory/max_active (GiB)": 46.9,
383
+ "memory/max_allocated (GiB)": 46.9,
384
+ "ppl": 1.01503,
385
+ "step": 27,
386
+ "tokens/total": 586567,
387
+ "tokens/train_per_sec_per_gpu": 5.29,
388
+ "tokens/trainable": 12242
389
+ },
390
+ {
391
+ "epoch": 0.109375,
392
+ "grad_norm": 18.625,
393
+ "learning_rate": 5e-06,
394
+ "loss": 0.013006031513214111,
395
+ "memory/device_reserved (GiB)": 60.14,
396
+ "memory/max_active (GiB)": 46.9,
397
+ "memory/max_allocated (GiB)": 46.9,
398
+ "ppl": 1.01309,
399
+ "step": 28,
400
+ "tokens/total": 608188,
401
+ "tokens/train_per_sec_per_gpu": 3.78,
402
+ "tokens/trainable": 12696
403
+ },
404
+ {
405
+ "epoch": 0.11328125,
406
+ "grad_norm": 21.25,
407
+ "learning_rate": 5e-06,
408
+ "loss": 0.028101369738578796,
409
+ "memory/device_reserved (GiB)": 60.14,
410
+ "memory/max_active (GiB)": 46.9,
411
+ "memory/max_allocated (GiB)": 46.9,
412
+ "ppl": 1.0285,
413
+ "step": 29,
414
+ "tokens/total": 630007,
415
+ "tokens/train_per_sec_per_gpu": 3.81,
416
+ "tokens/trainable": 13153
417
+ },
418
+ {
419
+ "epoch": 0.1171875,
420
+ "grad_norm": 8.3125,
421
+ "learning_rate": 5e-06,
422
+ "loss": 0.01690027117729187,
423
+ "memory/device_reserved (GiB)": 60.14,
424
+ "memory/max_active (GiB)": 46.9,
425
+ "memory/max_allocated (GiB)": 46.9,
426
+ "ppl": 1.01704,
427
+ "step": 30,
428
+ "tokens/total": 651656,
429
+ "tokens/train_per_sec_per_gpu": 5.16,
430
+ "tokens/trainable": 13624
431
+ },
432
+ {
433
+ "epoch": 0.12109375,
434
+ "grad_norm": 12.3125,
435
+ "learning_rate": 5e-06,
436
+ "loss": 0.015246853232383728,
437
+ "memory/device_reserved (GiB)": 60.14,
438
+ "memory/max_active (GiB)": 46.9,
439
+ "memory/max_allocated (GiB)": 46.9,
440
+ "ppl": 1.01536,
441
+ "step": 31,
442
+ "tokens/total": 672938,
443
+ "tokens/train_per_sec_per_gpu": 5.14,
444
+ "tokens/trainable": 14064
445
+ },
446
+ {
447
+ "epoch": 0.125,
448
+ "grad_norm": 11.125,
449
+ "learning_rate": 5e-06,
450
+ "loss": 0.017684906721115112,
451
+ "memory/device_reserved (GiB)": 60.14,
452
+ "memory/max_active (GiB)": 46.9,
453
+ "memory/max_allocated (GiB)": 46.9,
454
+ "ppl": 1.01784,
455
+ "step": 32,
456
+ "tokens/total": 696525,
457
+ "tokens/train_per_sec_per_gpu": 5.03,
458
+ "tokens/trainable": 14565
459
+ },
460
+ {
461
+ "epoch": 0.12890625,
462
+ "grad_norm": 15.5,
463
+ "learning_rate": 5e-06,
464
+ "loss": 0.02578166127204895,
465
+ "memory/device_reserved (GiB)": 60.14,
466
+ "memory/max_active (GiB)": 46.9,
467
+ "memory/max_allocated (GiB)": 46.9,
468
+ "ppl": 1.02612,
469
+ "step": 33,
470
+ "tokens/total": 717452,
471
+ "tokens/train_per_sec_per_gpu": 2.67,
472
+ "tokens/trainable": 15005
473
+ },
474
+ {
475
+ "epoch": 0.1328125,
476
+ "grad_norm": 10.5625,
477
+ "learning_rate": 5e-06,
478
+ "loss": 0.03840416669845581,
479
+ "memory/device_reserved (GiB)": 58.59,
480
+ "memory/max_active (GiB)": 46.9,
481
+ "memory/max_allocated (GiB)": 46.9,
482
+ "ppl": 1.03915,
483
+ "step": 34,
484
+ "tokens/total": 738762,
485
+ "tokens/train_per_sec_per_gpu": 3.32,
486
+ "tokens/trainable": 15459
487
+ },
488
+ {
489
+ "epoch": 0.13671875,
490
+ "grad_norm": 3.921875,
491
+ "learning_rate": 5e-06,
492
+ "loss": 0.017821550369262695,
493
+ "memory/device_reserved (GiB)": 58.59,
494
+ "memory/max_active (GiB)": 46.9,
495
+ "memory/max_allocated (GiB)": 46.9,
496
+ "ppl": 1.01798,
497
+ "step": 35,
498
+ "tokens/total": 762239,
499
+ "tokens/train_per_sec_per_gpu": 5.21,
500
+ "tokens/trainable": 15960
501
+ },
502
+ {
503
+ "epoch": 0.140625,
504
+ "grad_norm": 5.71875,
505
+ "learning_rate": 5e-06,
506
+ "loss": 0.02945883572101593,
507
+ "memory/device_reserved (GiB)": 58.59,
508
+ "memory/max_active (GiB)": 46.9,
509
+ "memory/max_allocated (GiB)": 46.9,
510
+ "ppl": 1.0299,
511
+ "step": 36,
512
+ "tokens/total": 784340,
513
+ "tokens/train_per_sec_per_gpu": 3.76,
514
+ "tokens/trainable": 16428
515
+ },
516
+ {
517
+ "epoch": 0.14453125,
518
+ "grad_norm": 24.625,
519
+ "learning_rate": 5e-06,
520
+ "loss": 0.011434584856033325,
521
+ "memory/device_reserved (GiB)": 58.59,
522
+ "memory/max_active (GiB)": 46.9,
523
+ "memory/max_allocated (GiB)": 46.9,
524
+ "ppl": 1.0115,
525
+ "step": 37,
526
+ "tokens/total": 806003,
527
+ "tokens/train_per_sec_per_gpu": 4.55,
528
+ "tokens/trainable": 16884
529
+ },
530
+ {
531
+ "epoch": 0.1484375,
532
+ "grad_norm": 6.71875,
533
+ "learning_rate": 5e-06,
534
+ "loss": 0.016611874103546143,
535
+ "memory/device_reserved (GiB)": 60.08,
536
+ "memory/max_active (GiB)": 46.9,
537
+ "memory/max_allocated (GiB)": 46.9,
538
+ "ppl": 1.01675,
539
+ "step": 38,
540
+ "tokens/total": 828092,
541
+ "tokens/train_per_sec_per_gpu": 5.15,
542
+ "tokens/trainable": 17341
543
+ },
544
+ {
545
+ "epoch": 0.15234375,
546
+ "grad_norm": 5.53125,
547
+ "learning_rate": 5e-06,
548
+ "loss": 0.012658625841140747,
549
+ "memory/device_reserved (GiB)": 60.08,
550
+ "memory/max_active (GiB)": 46.9,
551
+ "memory/max_allocated (GiB)": 46.9,
552
+ "ppl": 1.01274,
553
+ "step": 39,
554
+ "tokens/total": 849305,
555
+ "tokens/train_per_sec_per_gpu": 4.44,
556
+ "tokens/trainable": 17786
557
+ },
558
+ {
559
+ "epoch": 0.15625,
560
+ "grad_norm": 8.8125,
561
+ "learning_rate": 5e-06,
562
+ "loss": 0.02379515767097473,
563
+ "memory/device_reserved (GiB)": 60.08,
564
+ "memory/max_active (GiB)": 46.9,
565
+ "memory/max_allocated (GiB)": 46.9,
566
+ "ppl": 1.02408,
567
+ "step": 40,
568
+ "tokens/total": 871803,
569
+ "tokens/train_per_sec_per_gpu": 3.79,
570
+ "tokens/trainable": 18262
571
+ },
572
+ {
573
+ "epoch": 0.16015625,
574
+ "grad_norm": 2.015625,
575
+ "learning_rate": 5e-06,
576
+ "loss": 0.0050678253173828125,
577
+ "memory/device_reserved (GiB)": 60.08,
578
+ "memory/max_active (GiB)": 46.9,
579
+ "memory/max_allocated (GiB)": 46.9,
580
+ "ppl": 1.00508,
581
+ "step": 41,
582
+ "tokens/total": 893895,
583
+ "tokens/train_per_sec_per_gpu": 3.89,
584
+ "tokens/trainable": 18716
585
+ },
586
+ {
587
+ "epoch": 0.1640625,
588
+ "grad_norm": 5.875,
589
+ "learning_rate": 5e-06,
590
+ "loss": 0.016993582248687744,
591
+ "memory/device_reserved (GiB)": 60.08,
592
+ "memory/max_active (GiB)": 46.9,
593
+ "memory/max_allocated (GiB)": 46.9,
594
+ "ppl": 1.01714,
595
+ "step": 42,
596
+ "tokens/total": 914803,
597
+ "tokens/train_per_sec_per_gpu": 3.89,
598
+ "tokens/trainable": 19157
599
+ },
600
+ {
601
+ "epoch": 0.16796875,
602
+ "grad_norm": 4.1875,
603
+ "learning_rate": 5e-06,
604
+ "loss": 0.025514155626296997,
605
+ "memory/device_reserved (GiB)": 60.08,
606
+ "memory/max_active (GiB)": 46.9,
607
+ "memory/max_allocated (GiB)": 46.9,
608
+ "ppl": 1.02584,
609
+ "step": 43,
610
+ "tokens/total": 938486,
611
+ "tokens/train_per_sec_per_gpu": 4.88,
612
+ "tokens/trainable": 19647
613
+ },
614
+ {
615
+ "epoch": 0.171875,
616
+ "grad_norm": 10.1875,
617
+ "learning_rate": 5e-06,
618
+ "loss": 0.013556092977523804,
619
+ "memory/device_reserved (GiB)": 60.08,
620
+ "memory/max_active (GiB)": 46.9,
621
+ "memory/max_allocated (GiB)": 46.9,
622
+ "ppl": 1.01365,
623
+ "step": 44,
624
+ "tokens/total": 961149,
625
+ "tokens/train_per_sec_per_gpu": 4.89,
626
+ "tokens/trainable": 20119
627
+ },
628
+ {
629
+ "epoch": 0.17578125,
630
+ "grad_norm": 3.9375,
631
+ "learning_rate": 5e-06,
632
+ "loss": 0.012663766741752625,
633
+ "memory/device_reserved (GiB)": 60.08,
634
+ "memory/max_active (GiB)": 46.9,
635
+ "memory/max_allocated (GiB)": 46.9,
636
+ "ppl": 1.01274,
637
+ "step": 45,
638
+ "tokens/total": 982727,
639
+ "tokens/train_per_sec_per_gpu": 4.43,
640
+ "tokens/trainable": 20560
641
+ },
642
+ {
643
+ "epoch": 0.1796875,
644
+ "grad_norm": 2.421875,
645
+ "learning_rate": 5e-06,
646
+ "loss": 0.00803878903388977,
647
+ "memory/device_reserved (GiB)": 60.08,
648
+ "memory/max_active (GiB)": 46.9,
649
+ "memory/max_allocated (GiB)": 46.9,
650
+ "ppl": 1.00807,
651
+ "step": 46,
652
+ "tokens/total": 1003446,
653
+ "tokens/train_per_sec_per_gpu": 5.17,
654
+ "tokens/trainable": 20992
655
+ },
656
+ {
657
+ "epoch": 0.18359375,
658
+ "grad_norm": 6.40625,
659
+ "learning_rate": 5e-06,
660
+ "loss": 0.027271926403045654,
661
+ "memory/device_reserved (GiB)": 60.08,
662
+ "memory/max_active (GiB)": 46.9,
663
+ "memory/max_allocated (GiB)": 46.9,
664
+ "ppl": 1.02765,
665
+ "step": 47,
666
+ "tokens/total": 1024518,
667
+ "tokens/train_per_sec_per_gpu": 3.84,
668
+ "tokens/trainable": 21441
669
+ },
670
+ {
671
+ "epoch": 0.1875,
672
+ "grad_norm": 6.90625,
673
+ "learning_rate": 5e-06,
674
+ "loss": 0.019454196095466614,
675
+ "memory/device_reserved (GiB)": 60.08,
676
+ "memory/max_active (GiB)": 46.9,
677
+ "memory/max_allocated (GiB)": 46.9,
678
+ "ppl": 1.01964,
679
+ "step": 48,
680
+ "tokens/total": 1046348,
681
+ "tokens/train_per_sec_per_gpu": 3.79,
682
+ "tokens/trainable": 21901
683
+ },
684
+ {
685
+ "epoch": 0.19140625,
686
+ "grad_norm": 14.1875,
687
+ "learning_rate": 5e-06,
688
+ "loss": 0.007123962044715881,
689
+ "memory/device_reserved (GiB)": 60.08,
690
+ "memory/max_active (GiB)": 46.9,
691
+ "memory/max_allocated (GiB)": 46.9,
692
+ "ppl": 1.00715,
693
+ "step": 49,
694
+ "tokens/total": 1067448,
695
+ "tokens/train_per_sec_per_gpu": 4.46,
696
+ "tokens/trainable": 22341
697
+ },
698
+ {
699
+ "epoch": 0.1953125,
700
+ "grad_norm": 6.65625,
701
+ "learning_rate": 5e-06,
702
+ "loss": 0.016036301851272583,
703
+ "memory/device_reserved (GiB)": 60.08,
704
+ "memory/max_active (GiB)": 46.9,
705
+ "memory/max_allocated (GiB)": 46.9,
706
+ "ppl": 1.01617,
707
+ "step": 50,
708
+ "tokens/total": 1091250,
709
+ "tokens/train_per_sec_per_gpu": 5.67,
710
+ "tokens/trainable": 22833
711
+ },
712
+ {
713
+ "epoch": 0.19921875,
714
+ "grad_norm": 6.625,
715
+ "learning_rate": 5e-06,
716
+ "loss": 0.006014183163642883,
717
+ "memory/device_reserved (GiB)": 60.08,
718
+ "memory/max_active (GiB)": 46.9,
719
+ "memory/max_allocated (GiB)": 46.9,
720
+ "ppl": 1.00603,
721
+ "step": 51,
722
+ "tokens/total": 1112130,
723
+ "tokens/train_per_sec_per_gpu": 3.9,
724
+ "tokens/trainable": 23277
725
+ },
726
+ {
727
+ "epoch": 0.203125,
728
+ "grad_norm": 8.5,
729
+ "learning_rate": 5e-06,
730
+ "loss": 0.01116333156824112,
731
+ "memory/device_reserved (GiB)": 60.08,
732
+ "memory/max_active (GiB)": 46.9,
733
+ "memory/max_allocated (GiB)": 46.9,
734
+ "ppl": 1.01123,
735
+ "step": 52,
736
+ "tokens/total": 1135386,
737
+ "tokens/train_per_sec_per_gpu": 4.38,
738
+ "tokens/trainable": 23759
739
+ },
740
+ {
741
+ "epoch": 0.20703125,
742
+ "grad_norm": 7.53125,
743
+ "learning_rate": 5e-06,
744
+ "loss": 0.012836582958698273,
745
+ "memory/device_reserved (GiB)": 60.08,
746
+ "memory/max_active (GiB)": 46.9,
747
+ "memory/max_allocated (GiB)": 46.9,
748
+ "ppl": 1.01292,
749
+ "step": 53,
750
+ "tokens/total": 1156121,
751
+ "tokens/train_per_sec_per_gpu": 3.88,
752
+ "tokens/trainable": 24193
753
+ },
754
+ {
755
+ "epoch": 0.2109375,
756
+ "grad_norm": 3.453125,
757
+ "learning_rate": 5e-06,
758
+ "loss": 0.005587518215179443,
759
+ "memory/device_reserved (GiB)": 60.08,
760
+ "memory/max_active (GiB)": 46.9,
761
+ "memory/max_allocated (GiB)": 46.9,
762
+ "ppl": 1.0056,
763
+ "step": 54,
764
+ "tokens/total": 1176702,
765
+ "tokens/train_per_sec_per_gpu": 3.68,
766
+ "tokens/trainable": 24619
767
+ },
768
+ {
769
+ "epoch": 0.21484375,
770
+ "grad_norm": 78.5,
771
+ "learning_rate": 5e-06,
772
+ "loss": 0.05362102389335632,
773
+ "memory/device_reserved (GiB)": 60.08,
774
+ "memory/max_active (GiB)": 46.9,
775
+ "memory/max_allocated (GiB)": 46.9,
776
+ "ppl": 1.05508,
777
+ "step": 55,
778
+ "tokens/total": 1198839,
779
+ "tokens/train_per_sec_per_gpu": 3.13,
780
+ "tokens/trainable": 25075
781
+ },
782
+ {
783
+ "epoch": 0.21875,
784
+ "grad_norm": 2.796875,
785
+ "learning_rate": 5e-06,
786
+ "loss": 0.005013041198253632,
787
+ "memory/device_reserved (GiB)": 60.08,
788
+ "memory/max_active (GiB)": 46.9,
789
+ "memory/max_allocated (GiB)": 46.9,
790
+ "ppl": 1.00503,
791
+ "step": 56,
792
+ "tokens/total": 1219997,
793
+ "tokens/train_per_sec_per_gpu": 4.51,
794
+ "tokens/trainable": 25519
795
+ },
796
+ {
797
+ "epoch": 0.22265625,
798
+ "grad_norm": 0.65234375,
799
+ "learning_rate": 5e-06,
800
+ "loss": 0.0010031796991825104,
801
+ "memory/device_reserved (GiB)": 60.08,
802
+ "memory/max_active (GiB)": 46.9,
803
+ "memory/max_allocated (GiB)": 46.9,
804
+ "ppl": 1.001,
805
+ "step": 57,
806
+ "tokens/total": 1241006,
807
+ "tokens/train_per_sec_per_gpu": 4.5,
808
+ "tokens/trainable": 25966
809
+ },
810
+ {
811
+ "epoch": 0.2265625,
812
+ "grad_norm": 2.1875,
813
+ "learning_rate": 5e-06,
814
+ "loss": 0.0032418183982372284,
815
+ "memory/device_reserved (GiB)": 60.14,
816
+ "memory/max_active (GiB)": 46.9,
817
+ "memory/max_allocated (GiB)": 46.9,
818
+ "ppl": 1.00325,
819
+ "step": 58,
820
+ "tokens/total": 1263530,
821
+ "tokens/train_per_sec_per_gpu": 4.29,
822
+ "tokens/trainable": 26423
823
+ },
824
+ {
825
+ "epoch": 0.23046875,
826
+ "grad_norm": 15.3125,
827
+ "learning_rate": 5e-06,
828
+ "loss": 0.007594488561153412,
829
+ "memory/device_reserved (GiB)": 60.14,
830
+ "memory/max_active (GiB)": 46.9,
831
+ "memory/max_allocated (GiB)": 46.9,
832
+ "ppl": 1.00762,
833
+ "step": 59,
834
+ "tokens/total": 1287978,
835
+ "tokens/train_per_sec_per_gpu": 4.36,
836
+ "tokens/trainable": 26934
837
+ },
838
+ {
839
+ "epoch": 0.234375,
840
+ "grad_norm": 10.375,
841
+ "learning_rate": 5e-06,
842
+ "loss": 0.009997941553592682,
843
+ "memory/device_reserved (GiB)": 60.14,
844
+ "memory/max_active (GiB)": 46.9,
845
+ "memory/max_allocated (GiB)": 46.9,
846
+ "ppl": 1.01005,
847
+ "step": 60,
848
+ "tokens/total": 1309463,
849
+ "tokens/train_per_sec_per_gpu": 3.83,
850
+ "tokens/trainable": 27393
851
+ },
852
+ {
853
+ "epoch": 0.23828125,
854
+ "grad_norm": 2.34375,
855
+ "learning_rate": 5e-06,
856
+ "loss": 0.0018807370215654373,
857
+ "memory/device_reserved (GiB)": 60.14,
858
+ "memory/max_active (GiB)": 46.9,
859
+ "memory/max_allocated (GiB)": 46.9,
860
+ "ppl": 1.00188,
861
+ "step": 61,
862
+ "tokens/total": 1331283,
863
+ "tokens/train_per_sec_per_gpu": 4.99,
864
+ "tokens/trainable": 27852
865
+ },
866
+ {
867
+ "epoch": 0.2421875,
868
+ "grad_norm": 11.6875,
869
+ "learning_rate": 5e-06,
870
+ "loss": 0.03929818421602249,
871
+ "memory/device_reserved (GiB)": 60.14,
872
+ "memory/max_active (GiB)": 46.9,
873
+ "memory/max_allocated (GiB)": 46.9,
874
+ "ppl": 1.04008,
875
+ "step": 62,
876
+ "tokens/total": 1352434,
877
+ "tokens/train_per_sec_per_gpu": 3.85,
878
+ "tokens/trainable": 28292
879
+ },
880
+ {
881
+ "epoch": 0.24609375,
882
+ "grad_norm": 12.3125,
883
+ "learning_rate": 5e-06,
884
+ "loss": 0.03207191079854965,
885
+ "memory/device_reserved (GiB)": 60.14,
886
+ "memory/max_active (GiB)": 46.9,
887
+ "memory/max_allocated (GiB)": 46.9,
888
+ "ppl": 1.03259,
889
+ "step": 63,
890
+ "tokens/total": 1376383,
891
+ "tokens/train_per_sec_per_gpu": 5.07,
892
+ "tokens/trainable": 28798
893
+ },
894
+ {
895
+ "epoch": 0.25,
896
+ "grad_norm": 27.125,
897
+ "learning_rate": 5e-06,
898
+ "loss": 0.03780709207057953,
899
+ "memory/device_reserved (GiB)": 60.14,
900
+ "memory/max_active (GiB)": 46.9,
901
+ "memory/max_allocated (GiB)": 46.9,
902
+ "ppl": 1.03853,
903
+ "step": 64,
904
+ "tokens/total": 1399140,
905
+ "tokens/train_per_sec_per_gpu": 5.15,
906
+ "tokens/trainable": 29283
907
+ },
908
+ {
909
+ "epoch": 0.25390625,
910
+ "grad_norm": 9.375,
911
+ "learning_rate": 5e-06,
912
+ "loss": 0.005781790241599083,
913
+ "memory/device_reserved (GiB)": 60.14,
914
+ "memory/max_active (GiB)": 46.9,
915
+ "memory/max_allocated (GiB)": 46.9,
916
+ "ppl": 1.0058,
917
+ "step": 65,
918
+ "tokens/total": 1421624,
919
+ "tokens/train_per_sec_per_gpu": 3.27,
920
+ "tokens/trainable": 29745
921
+ },
922
+ {
923
+ "epoch": 0.2578125,
924
+ "grad_norm": 5.78125,
925
+ "learning_rate": 5e-06,
926
+ "loss": 0.002019934356212616,
927
+ "memory/device_reserved (GiB)": 58.2,
928
+ "memory/max_active (GiB)": 46.9,
929
+ "memory/max_allocated (GiB)": 46.9,
930
+ "ppl": 1.00202,
931
+ "step": 66,
932
+ "tokens/total": 1442453,
933
+ "tokens/train_per_sec_per_gpu": 4.5,
934
+ "tokens/trainable": 30186
935
+ },
936
+ {
937
+ "epoch": 0.26171875,
938
+ "grad_norm": 22.25,
939
+ "learning_rate": 5e-06,
940
+ "loss": 0.005651544779539108,
941
+ "memory/device_reserved (GiB)": 58.2,
942
+ "memory/max_active (GiB)": 46.9,
943
+ "memory/max_allocated (GiB)": 46.9,
944
+ "ppl": 1.00567,
945
+ "step": 67,
946
+ "tokens/total": 1462799,
947
+ "tokens/train_per_sec_per_gpu": 5.22,
948
+ "tokens/trainable": 30622
949
+ },
950
+ {
951
+ "epoch": 0.265625,
952
+ "grad_norm": 14.8125,
953
+ "learning_rate": 5e-06,
954
+ "loss": 0.01753600873053074,
955
+ "memory/device_reserved (GiB)": 58.2,
956
+ "memory/max_active (GiB)": 46.9,
957
+ "memory/max_allocated (GiB)": 46.9,
958
+ "ppl": 1.01769,
959
+ "step": 68,
960
+ "tokens/total": 1485719,
961
+ "tokens/train_per_sec_per_gpu": 5.07,
962
+ "tokens/trainable": 31097
963
+ },
964
+ {
965
+ "epoch": 0.26953125,
966
+ "grad_norm": 10.25,
967
+ "learning_rate": 5e-06,
968
+ "loss": 0.011547420173883438,
969
+ "memory/device_reserved (GiB)": 58.2,
970
+ "memory/max_active (GiB)": 46.9,
971
+ "memory/max_allocated (GiB)": 46.9,
972
+ "ppl": 1.01161,
973
+ "step": 69,
974
+ "tokens/total": 1509089,
975
+ "tokens/train_per_sec_per_gpu": 4.59,
976
+ "tokens/trainable": 31595
977
+ },
978
+ {
979
+ "epoch": 0.2734375,
980
+ "grad_norm": 5.6875,
981
+ "learning_rate": 5e-06,
982
+ "loss": 0.010879918932914734,
983
+ "memory/device_reserved (GiB)": 58.2,
984
+ "memory/max_active (GiB)": 46.9,
985
+ "memory/max_allocated (GiB)": 46.9,
986
+ "ppl": 1.01094,
987
+ "step": 70,
988
+ "tokens/total": 1530488,
989
+ "tokens/train_per_sec_per_gpu": 5.17,
990
+ "tokens/trainable": 32036
991
+ },
992
+ {
993
+ "epoch": 0.27734375,
994
+ "grad_norm": 34.75,
995
+ "learning_rate": 5e-06,
996
+ "loss": 0.01058020070195198,
997
+ "memory/device_reserved (GiB)": 58.2,
998
+ "memory/max_active (GiB)": 46.9,
999
+ "memory/max_allocated (GiB)": 46.9,
1000
+ "ppl": 1.01064,
1001
+ "step": 71,
1002
+ "tokens/total": 1549610,
1003
+ "tokens/train_per_sec_per_gpu": 3.95,
1004
+ "tokens/trainable": 32428
1005
+ },
1006
+ {
1007
+ "epoch": 0.28125,
1008
+ "grad_norm": 7.40625,
1009
+ "learning_rate": 5e-06,
1010
+ "loss": 0.008541440591216087,
1011
+ "memory/device_reserved (GiB)": 58.2,
1012
+ "memory/max_active (GiB)": 46.9,
1013
+ "memory/max_allocated (GiB)": 46.9,
1014
+ "ppl": 1.00858,
1015
+ "step": 72,
1016
+ "tokens/total": 1571082,
1017
+ "tokens/train_per_sec_per_gpu": 3.96,
1018
+ "tokens/trainable": 32880
1019
+ },
1020
+ {
1021
+ "epoch": 0.28515625,
1022
+ "grad_norm": 4.375,
1023
+ "learning_rate": 5e-06,
1024
+ "loss": 0.011486759409308434,
1025
+ "memory/device_reserved (GiB)": 58.2,
1026
+ "memory/max_active (GiB)": 46.9,
1027
+ "memory/max_allocated (GiB)": 46.9,
1028
+ "ppl": 1.01155,
1029
+ "step": 73,
1030
+ "tokens/total": 1591765,
1031
+ "tokens/train_per_sec_per_gpu": 4.55,
1032
+ "tokens/trainable": 33323
1033
+ },
1034
+ {
1035
+ "epoch": 0.2890625,
1036
+ "grad_norm": 2.9375,
1037
+ "learning_rate": 5e-06,
1038
+ "loss": 0.0029051266610622406,
1039
+ "memory/device_reserved (GiB)": 58.2,
1040
+ "memory/max_active (GiB)": 46.9,
1041
+ "memory/max_allocated (GiB)": 46.9,
1042
+ "ppl": 1.00291,
1043
+ "step": 74,
1044
+ "tokens/total": 1613822,
1045
+ "tokens/train_per_sec_per_gpu": 4.52,
1046
+ "tokens/trainable": 33792
1047
+ },
1048
+ {
1049
+ "epoch": 0.29296875,
1050
+ "grad_norm": 6.375,
1051
+ "learning_rate": 5e-06,
1052
+ "loss": 0.007972124963998795,
1053
+ "memory/device_reserved (GiB)": 58.2,
1054
+ "memory/max_active (GiB)": 46.9,
1055
+ "memory/max_allocated (GiB)": 46.9,
1056
+ "ppl": 1.008,
1057
+ "step": 75,
1058
+ "tokens/total": 1634169,
1059
+ "tokens/train_per_sec_per_gpu": 4.61,
1060
+ "tokens/trainable": 34223
1061
+ },
1062
+ {
1063
+ "epoch": 0.296875,
1064
+ "grad_norm": 12.0625,
1065
+ "learning_rate": 5e-06,
1066
+ "loss": 0.016271991655230522,
1067
+ "memory/device_reserved (GiB)": 58.2,
1068
+ "memory/max_active (GiB)": 46.9,
1069
+ "memory/max_allocated (GiB)": 46.9,
1070
+ "ppl": 1.01641,
1071
+ "step": 76,
1072
+ "tokens/total": 1654482,
1073
+ "tokens/train_per_sec_per_gpu": 3.86,
1074
+ "tokens/trainable": 34649
1075
+ },
1076
+ {
1077
+ "epoch": 0.30078125,
1078
+ "grad_norm": 3.46875,
1079
+ "learning_rate": 5e-06,
1080
+ "loss": 0.006356913596391678,
1081
+ "memory/device_reserved (GiB)": 59.92,
1082
+ "memory/max_active (GiB)": 46.9,
1083
+ "memory/max_allocated (GiB)": 46.9,
1084
+ "ppl": 1.00638,
1085
+ "step": 77,
1086
+ "tokens/total": 1675738,
1087
+ "tokens/train_per_sec_per_gpu": 4.49,
1088
+ "tokens/trainable": 35094
1089
+ },
1090
+ {
1091
+ "epoch": 0.3046875,
1092
+ "grad_norm": 9.0625,
1093
+ "learning_rate": 5e-06,
1094
+ "loss": 0.026207175105810165,
1095
+ "memory/device_reserved (GiB)": 59.92,
1096
+ "memory/max_active (GiB)": 46.9,
1097
+ "memory/max_allocated (GiB)": 46.9,
1098
+ "ppl": 1.02655,
1099
+ "step": 78,
1100
+ "tokens/total": 1697159,
1101
+ "tokens/train_per_sec_per_gpu": 5.02,
1102
+ "tokens/trainable": 35534
1103
+ },
1104
+ {
1105
+ "epoch": 0.30859375,
1106
+ "grad_norm": 9.0,
1107
+ "learning_rate": 5e-06,
1108
+ "loss": 0.01054411381483078,
1109
+ "memory/device_reserved (GiB)": 59.92,
1110
+ "memory/max_active (GiB)": 46.9,
1111
+ "memory/max_allocated (GiB)": 46.9,
1112
+ "ppl": 1.0106,
1113
+ "step": 79,
1114
+ "tokens/total": 1720420,
1115
+ "tokens/train_per_sec_per_gpu": 4.31,
1116
+ "tokens/trainable": 36029
1117
+ },
1118
+ {
1119
+ "epoch": 0.3125,
1120
+ "grad_norm": 6.4375,
1121
+ "learning_rate": 5e-06,
1122
+ "loss": 0.003844674676656723,
1123
+ "memory/device_reserved (GiB)": 59.92,
1124
+ "memory/max_active (GiB)": 46.9,
1125
+ "memory/max_allocated (GiB)": 46.9,
1126
+ "ppl": 1.00385,
1127
+ "step": 80,
1128
+ "tokens/total": 1741815,
1129
+ "tokens/train_per_sec_per_gpu": 4.39,
1130
+ "tokens/trainable": 36469
1131
+ },
1132
+ {
1133
+ "epoch": 0.31640625,
1134
+ "grad_norm": 13.5,
1135
+ "learning_rate": 5e-06,
1136
+ "loss": 0.010209254920482635,
1137
+ "memory/device_reserved (GiB)": 59.92,
1138
+ "memory/max_active (GiB)": 46.9,
1139
+ "memory/max_allocated (GiB)": 46.9,
1140
+ "ppl": 1.01026,
1141
+ "step": 81,
1142
+ "tokens/total": 1762269,
1143
+ "tokens/train_per_sec_per_gpu": 3.91,
1144
+ "tokens/trainable": 36906
1145
+ },
1146
+ {
1147
+ "epoch": 0.3203125,
1148
+ "grad_norm": 3.8125,
1149
+ "learning_rate": 5e-06,
1150
+ "loss": 0.0027688145637512207,
1151
+ "memory/device_reserved (GiB)": 59.92,
1152
+ "memory/max_active (GiB)": 46.9,
1153
+ "memory/max_allocated (GiB)": 46.9,
1154
+ "ppl": 1.00277,
1155
+ "step": 82,
1156
+ "tokens/total": 1785551,
1157
+ "tokens/train_per_sec_per_gpu": 5.79,
1158
+ "tokens/trainable": 37390
1159
+ },
1160
+ {
1161
+ "epoch": 0.32421875,
1162
+ "grad_norm": 7.5,
1163
+ "learning_rate": 5e-06,
1164
+ "loss": 0.00830920785665512,
1165
+ "memory/device_reserved (GiB)": 60.1,
1166
+ "memory/max_active (GiB)": 46.9,
1167
+ "memory/max_allocated (GiB)": 46.9,
1168
+ "ppl": 1.00834,
1169
+ "step": 83,
1170
+ "tokens/total": 1807278,
1171
+ "tokens/train_per_sec_per_gpu": 4.46,
1172
+ "tokens/trainable": 37838
1173
+ },
1174
+ {
1175
+ "epoch": 0.328125,
1176
+ "grad_norm": 6.46875,
1177
+ "learning_rate": 5e-06,
1178
+ "loss": 0.002562195062637329,
1179
+ "memory/device_reserved (GiB)": 60.1,
1180
+ "memory/max_active (GiB)": 46.9,
1181
+ "memory/max_allocated (GiB)": 46.9,
1182
+ "ppl": 1.00257,
1183
+ "step": 84,
1184
+ "tokens/total": 1829716,
1185
+ "tokens/train_per_sec_per_gpu": 4.38,
1186
+ "tokens/trainable": 38310
1187
+ },
1188
+ {
1189
+ "epoch": 0.33203125,
1190
+ "grad_norm": 10.4375,
1191
+ "learning_rate": 5e-06,
1192
+ "loss": 0.029578130692243576,
1193
+ "memory/device_reserved (GiB)": 60.1,
1194
+ "memory/max_active (GiB)": 46.9,
1195
+ "memory/max_allocated (GiB)": 46.9,
1196
+ "ppl": 1.03002,
1197
+ "step": 85,
1198
+ "tokens/total": 1852531,
1199
+ "tokens/train_per_sec_per_gpu": 4.5,
1200
+ "tokens/trainable": 38789
1201
+ },
1202
+ {
1203
+ "epoch": 0.3359375,
1204
+ "grad_norm": 11.75,
1205
+ "learning_rate": 5e-06,
1206
+ "loss": 0.024770356714725494,
1207
+ "memory/device_reserved (GiB)": 60.1,
1208
+ "memory/max_active (GiB)": 46.9,
1209
+ "memory/max_allocated (GiB)": 46.9,
1210
+ "ppl": 1.02508,
1211
+ "step": 86,
1212
+ "tokens/total": 1875946,
1213
+ "tokens/train_per_sec_per_gpu": 4.5,
1214
+ "tokens/trainable": 39277
1215
+ },
1216
+ {
1217
+ "epoch": 0.33984375,
1218
+ "grad_norm": 7.0,
1219
+ "learning_rate": 5e-06,
1220
+ "loss": 0.003416106104850769,
1221
+ "memory/device_reserved (GiB)": 60.1,
1222
+ "memory/max_active (GiB)": 46.9,
1223
+ "memory/max_allocated (GiB)": 46.9,
1224
+ "ppl": 1.00342,
1225
+ "step": 87,
1226
+ "tokens/total": 1897851,
1227
+ "tokens/train_per_sec_per_gpu": 4.96,
1228
+ "tokens/trainable": 39737
1229
+ },
1230
+ {
1231
+ "epoch": 0.34375,
1232
+ "grad_norm": 49.75,
1233
+ "learning_rate": 5e-06,
1234
+ "loss": 0.0025530681014060974,
1235
+ "memory/device_reserved (GiB)": 60.1,
1236
+ "memory/max_active (GiB)": 46.9,
1237
+ "memory/max_allocated (GiB)": 46.9,
1238
+ "ppl": 1.00256,
1239
+ "step": 88,
1240
+ "tokens/total": 1918887,
1241
+ "tokens/train_per_sec_per_gpu": 3.94,
1242
+ "tokens/trainable": 40179
1243
+ },
1244
+ {
1245
+ "epoch": 0.34765625,
1246
+ "grad_norm": 4.71875,
1247
+ "learning_rate": 5e-06,
1248
+ "loss": 0.0036525018513202667,
1249
+ "memory/device_reserved (GiB)": 60.1,
1250
+ "memory/max_active (GiB)": 46.9,
1251
+ "memory/max_allocated (GiB)": 46.9,
1252
+ "ppl": 1.00366,
1253
+ "step": 89,
1254
+ "tokens/total": 1941639,
1255
+ "tokens/train_per_sec_per_gpu": 4.45,
1256
+ "tokens/trainable": 40649
1257
+ },
1258
+ {
1259
+ "epoch": 0.3515625,
1260
+ "grad_norm": 22.75,
1261
+ "learning_rate": 5e-06,
1262
+ "loss": 0.003507021814584732,
1263
+ "memory/device_reserved (GiB)": 60.1,
1264
+ "memory/max_active (GiB)": 46.9,
1265
+ "memory/max_allocated (GiB)": 46.9,
1266
+ "ppl": 1.00351,
1267
+ "step": 90,
1268
+ "tokens/total": 1961837,
1269
+ "tokens/train_per_sec_per_gpu": 3.85,
1270
+ "tokens/trainable": 41080
1271
+ },
1272
+ {
1273
+ "epoch": 0.35546875,
1274
+ "grad_norm": 0.443359375,
1275
+ "learning_rate": 5e-06,
1276
+ "loss": 0.0008387751877307892,
1277
+ "memory/device_reserved (GiB)": 60.1,
1278
+ "memory/max_active (GiB)": 46.9,
1279
+ "memory/max_allocated (GiB)": 46.9,
1280
+ "ppl": 1.00084,
1281
+ "step": 91,
1282
+ "tokens/total": 1986102,
1283
+ "tokens/train_per_sec_per_gpu": 4.95,
1284
+ "tokens/trainable": 41599
1285
+ },
1286
+ {
1287
+ "epoch": 0.359375,
1288
+ "grad_norm": 2.75,
1289
+ "learning_rate": 5e-06,
1290
+ "loss": 0.0063427165150642395,
1291
+ "memory/device_reserved (GiB)": 60.1,
1292
+ "memory/max_active (GiB)": 46.9,
1293
+ "memory/max_allocated (GiB)": 46.9,
1294
+ "ppl": 1.00636,
1295
+ "step": 92,
1296
+ "tokens/total": 2008498,
1297
+ "tokens/train_per_sec_per_gpu": 3.79,
1298
+ "tokens/trainable": 42076
1299
+ },
1300
+ {
1301
+ "epoch": 0.36328125,
1302
+ "grad_norm": 8.375,
1303
+ "learning_rate": 5e-06,
1304
+ "loss": 0.0007749572396278381,
1305
+ "memory/device_reserved (GiB)": 60.1,
1306
+ "memory/max_active (GiB)": 46.9,
1307
+ "memory/max_allocated (GiB)": 46.9,
1308
+ "ppl": 1.00078,
1309
+ "step": 93,
1310
+ "tokens/total": 2030279,
1311
+ "tokens/train_per_sec_per_gpu": 5.07,
1312
+ "tokens/trainable": 42522
1313
+ },
1314
+ {
1315
+ "epoch": 0.3671875,
1316
+ "grad_norm": 4.90625,
1317
+ "learning_rate": 5e-06,
1318
+ "loss": 0.02601948380470276,
1319
+ "memory/device_reserved (GiB)": 60.1,
1320
+ "memory/max_active (GiB)": 46.9,
1321
+ "memory/max_allocated (GiB)": 46.9,
1322
+ "ppl": 1.02636,
1323
+ "step": 94,
1324
+ "tokens/total": 2051914,
1325
+ "tokens/train_per_sec_per_gpu": 5.33,
1326
+ "tokens/trainable": 42974
1327
+ },
1328
+ {
1329
+ "epoch": 0.37109375,
1330
+ "grad_norm": 4.71875,
1331
+ "learning_rate": 5e-06,
1332
+ "loss": 0.015122350305318832,
1333
+ "memory/device_reserved (GiB)": 60.1,
1334
+ "memory/max_active (GiB)": 46.9,
1335
+ "memory/max_allocated (GiB)": 46.9,
1336
+ "ppl": 1.01524,
1337
+ "step": 95,
1338
+ "tokens/total": 2071768,
1339
+ "tokens/train_per_sec_per_gpu": 3.93,
1340
+ "tokens/trainable": 43380
1341
+ },
1342
+ {
1343
+ "epoch": 0.375,
1344
+ "grad_norm": 11.0625,
1345
+ "learning_rate": 5e-06,
1346
+ "loss": 0.01082584261894226,
1347
+ "memory/device_reserved (GiB)": 60.1,
1348
+ "memory/max_active (GiB)": 46.9,
1349
+ "memory/max_allocated (GiB)": 46.9,
1350
+ "ppl": 1.01088,
1351
+ "step": 96,
1352
+ "tokens/total": 2094957,
1353
+ "tokens/train_per_sec_per_gpu": 3.8,
1354
+ "tokens/trainable": 43865
1355
+ },
1356
+ {
1357
+ "epoch": 0.37890625,
1358
+ "grad_norm": 3.953125,
1359
+ "learning_rate": 5e-06,
1360
+ "loss": 0.008022107183933258,
1361
+ "memory/device_reserved (GiB)": 60.1,
1362
+ "memory/max_active (GiB)": 46.9,
1363
+ "memory/max_allocated (GiB)": 46.9,
1364
+ "ppl": 1.00805,
1365
+ "step": 97,
1366
+ "tokens/total": 2117603,
1367
+ "tokens/train_per_sec_per_gpu": 2.97,
1368
+ "tokens/trainable": 44328
1369
+ },
1370
+ {
1371
+ "epoch": 0.3828125,
1372
+ "grad_norm": 8.875,
1373
+ "learning_rate": 5e-06,
1374
+ "loss": 0.010948814451694489,
1375
+ "memory/device_reserved (GiB)": 60.1,
1376
+ "memory/max_active (GiB)": 46.9,
1377
+ "memory/max_allocated (GiB)": 46.9,
1378
+ "ppl": 1.01101,
1379
+ "step": 98,
1380
+ "tokens/total": 2138416,
1381
+ "tokens/train_per_sec_per_gpu": 4.0,
1382
+ "tokens/trainable": 44778
1383
+ },
1384
+ {
1385
+ "epoch": 0.38671875,
1386
+ "grad_norm": 7.5,
1387
+ "learning_rate": 5e-06,
1388
+ "loss": 0.01736413687467575,
1389
+ "memory/device_reserved (GiB)": 60.1,
1390
+ "memory/max_active (GiB)": 46.9,
1391
+ "memory/max_allocated (GiB)": 46.9,
1392
+ "ppl": 1.01752,
1393
+ "step": 99,
1394
+ "tokens/total": 2160046,
1395
+ "tokens/train_per_sec_per_gpu": 5.63,
1396
+ "tokens/trainable": 45225
1397
+ },
1398
+ {
1399
+ "epoch": 0.390625,
1400
+ "grad_norm": 1.828125,
1401
+ "learning_rate": 5e-06,
1402
+ "loss": 0.002016715705394745,
1403
+ "memory/device_reserved (GiB)": 60.1,
1404
+ "memory/max_active (GiB)": 46.9,
1405
+ "memory/max_allocated (GiB)": 46.9,
1406
+ "ppl": 1.00202,
1407
+ "step": 100,
1408
+ "tokens/total": 2181923,
1409
+ "tokens/train_per_sec_per_gpu": 4.5,
1410
+ "tokens/trainable": 45678
1411
+ },
1412
+ {
1413
+ "epoch": 0.39453125,
1414
+ "grad_norm": 1.25,
1415
+ "learning_rate": 5e-06,
1416
+ "loss": 0.0019299909472465515,
1417
+ "memory/device_reserved (GiB)": 60.1,
1418
+ "memory/max_active (GiB)": 46.9,
1419
+ "memory/max_allocated (GiB)": 46.9,
1420
+ "ppl": 1.00193,
1421
+ "step": 101,
1422
+ "tokens/total": 2203558,
1423
+ "tokens/train_per_sec_per_gpu": 4.39,
1424
+ "tokens/trainable": 46125
1425
+ },
1426
+ {
1427
+ "epoch": 0.3984375,
1428
+ "grad_norm": 3.859375,
1429
+ "learning_rate": 5e-06,
1430
+ "loss": 0.004272259771823883,
1431
+ "memory/device_reserved (GiB)": 60.1,
1432
+ "memory/max_active (GiB)": 46.9,
1433
+ "memory/max_allocated (GiB)": 46.9,
1434
+ "ppl": 1.00428,
1435
+ "step": 102,
1436
+ "tokens/total": 2226903,
1437
+ "tokens/train_per_sec_per_gpu": 4.99,
1438
+ "tokens/trainable": 46612
1439
+ },
1440
+ {
1441
+ "epoch": 0.40234375,
1442
+ "grad_norm": 8.5625,
1443
+ "learning_rate": 5e-06,
1444
+ "loss": 0.005463987588882446,
1445
+ "memory/device_reserved (GiB)": 60.1,
1446
+ "memory/max_active (GiB)": 46.9,
1447
+ "memory/max_allocated (GiB)": 46.9,
1448
+ "ppl": 1.00548,
1449
+ "step": 103,
1450
+ "tokens/total": 2248454,
1451
+ "tokens/train_per_sec_per_gpu": 4.56,
1452
+ "tokens/trainable": 47059
1453
+ },
1454
+ {
1455
+ "epoch": 0.40625,
1456
+ "grad_norm": 28.25,
1457
+ "learning_rate": 5e-06,
1458
+ "loss": 0.029229726642370224,
1459
+ "memory/device_reserved (GiB)": 60.1,
1460
+ "memory/max_active (GiB)": 46.9,
1461
+ "memory/max_allocated (GiB)": 46.9,
1462
+ "ppl": 1.02966,
1463
+ "step": 104,
1464
+ "tokens/total": 2271600,
1465
+ "tokens/train_per_sec_per_gpu": 4.31,
1466
+ "tokens/trainable": 47536
1467
+ },
1468
+ {
1469
+ "epoch": 0.41015625,
1470
+ "grad_norm": 3.8125,
1471
+ "learning_rate": 5e-06,
1472
+ "loss": 0.002427205443382263,
1473
+ "memory/device_reserved (GiB)": 60.1,
1474
+ "memory/max_active (GiB)": 46.9,
1475
+ "memory/max_allocated (GiB)": 46.9,
1476
+ "ppl": 1.00243,
1477
+ "step": 105,
1478
+ "tokens/total": 2295278,
1479
+ "tokens/train_per_sec_per_gpu": 3.72,
1480
+ "tokens/trainable": 48037
1481
+ },
1482
+ {
1483
+ "epoch": 0.4140625,
1484
+ "grad_norm": 3.734375,
1485
+ "learning_rate": 5e-06,
1486
+ "loss": 0.0015581585466861725,
1487
+ "memory/device_reserved (GiB)": 60.1,
1488
+ "memory/max_active (GiB)": 46.9,
1489
+ "memory/max_allocated (GiB)": 46.9,
1490
+ "ppl": 1.00156,
1491
+ "step": 106,
1492
+ "tokens/total": 2317645,
1493
+ "tokens/train_per_sec_per_gpu": 4.93,
1494
+ "tokens/trainable": 48491
1495
+ },
1496
+ {
1497
+ "epoch": 0.41796875,
1498
+ "grad_norm": 17.125,
1499
+ "learning_rate": 5e-06,
1500
+ "loss": 0.023405246436595917,
1501
+ "memory/device_reserved (GiB)": 60.1,
1502
+ "memory/max_active (GiB)": 46.9,
1503
+ "memory/max_allocated (GiB)": 46.9,
1504
+ "ppl": 1.02368,
1505
+ "step": 107,
1506
+ "tokens/total": 2339279,
1507
+ "tokens/train_per_sec_per_gpu": 4.48,
1508
+ "tokens/trainable": 48940
1509
+ },
1510
+ {
1511
+ "epoch": 0.421875,
1512
+ "grad_norm": 2.53125,
1513
+ "learning_rate": 5e-06,
1514
+ "loss": 0.0015513133257627487,
1515
+ "memory/device_reserved (GiB)": 60.1,
1516
+ "memory/max_active (GiB)": 46.9,
1517
+ "memory/max_allocated (GiB)": 46.9,
1518
+ "ppl": 1.00155,
1519
+ "step": 108,
1520
+ "tokens/total": 2360540,
1521
+ "tokens/train_per_sec_per_gpu": 3.87,
1522
+ "tokens/trainable": 49389
1523
+ },
1524
+ {
1525
+ "epoch": 0.42578125,
1526
+ "grad_norm": 12.875,
1527
+ "learning_rate": 5e-06,
1528
+ "loss": 0.031832728534936905,
1529
+ "memory/device_reserved (GiB)": 60.1,
1530
+ "memory/max_active (GiB)": 46.9,
1531
+ "memory/max_allocated (GiB)": 46.9,
1532
+ "ppl": 1.03234,
1533
+ "step": 109,
1534
+ "tokens/total": 2381374,
1535
+ "tokens/train_per_sec_per_gpu": 4.73,
1536
+ "tokens/trainable": 49838
1537
+ },
1538
+ {
1539
+ "epoch": 0.4296875,
1540
+ "grad_norm": 6.71875,
1541
+ "learning_rate": 5e-06,
1542
+ "loss": 0.002377200871706009,
1543
+ "memory/device_reserved (GiB)": 60.1,
1544
+ "memory/max_active (GiB)": 46.9,
1545
+ "memory/max_allocated (GiB)": 46.9,
1546
+ "ppl": 1.00238,
1547
+ "step": 110,
1548
+ "tokens/total": 2403924,
1549
+ "tokens/train_per_sec_per_gpu": 4.38,
1550
+ "tokens/trainable": 50310
1551
+ },
1552
+ {
1553
+ "epoch": 0.43359375,
1554
+ "grad_norm": 1.1875,
1555
+ "learning_rate": 5e-06,
1556
+ "loss": 0.0006611049175262451,
1557
+ "memory/device_reserved (GiB)": 60.1,
1558
+ "memory/max_active (GiB)": 46.9,
1559
+ "memory/max_allocated (GiB)": 46.9,
1560
+ "ppl": 1.00066,
1561
+ "step": 111,
1562
+ "tokens/total": 2424591,
1563
+ "tokens/train_per_sec_per_gpu": 3.89,
1564
+ "tokens/trainable": 50755
1565
+ },
1566
+ {
1567
+ "epoch": 0.4375,
1568
+ "grad_norm": 1.359375,
1569
+ "learning_rate": 5e-06,
1570
+ "loss": 0.0006331540644168854,
1571
+ "memory/device_reserved (GiB)": 60.1,
1572
+ "memory/max_active (GiB)": 46.9,
1573
+ "memory/max_allocated (GiB)": 46.9,
1574
+ "ppl": 1.00063,
1575
+ "step": 112,
1576
+ "tokens/total": 2446829,
1577
+ "tokens/train_per_sec_per_gpu": 4.36,
1578
+ "tokens/trainable": 51213
1579
+ },
1580
+ {
1581
+ "epoch": 0.44140625,
1582
+ "grad_norm": 0.1259765625,
1583
+ "learning_rate": 5e-06,
1584
+ "loss": 0.0002308972179889679,
1585
+ "memory/device_reserved (GiB)": 60.1,
1586
+ "memory/max_active (GiB)": 46.9,
1587
+ "memory/max_allocated (GiB)": 46.9,
1588
+ "ppl": 1.00023,
1589
+ "step": 113,
1590
+ "tokens/total": 2467554,
1591
+ "tokens/train_per_sec_per_gpu": 5.3,
1592
+ "tokens/trainable": 51660
1593
+ },
1594
+ {
1595
+ "epoch": 0.4453125,
1596
+ "grad_norm": 0.53515625,
1597
+ "learning_rate": 5e-06,
1598
+ "loss": 0.00044569000601768494,
1599
+ "memory/device_reserved (GiB)": 60.1,
1600
+ "memory/max_active (GiB)": 46.9,
1601
+ "memory/max_allocated (GiB)": 46.9,
1602
+ "ppl": 1.00045,
1603
+ "step": 114,
1604
+ "tokens/total": 2490093,
1605
+ "tokens/train_per_sec_per_gpu": 3.86,
1606
+ "tokens/trainable": 52133
1607
+ },
1608
+ {
1609
+ "epoch": 0.44921875,
1610
+ "grad_norm": 0.640625,
1611
+ "learning_rate": 5e-06,
1612
+ "loss": 0.000523759052157402,
1613
+ "memory/device_reserved (GiB)": 60.1,
1614
+ "memory/max_active (GiB)": 46.9,
1615
+ "memory/max_allocated (GiB)": 46.9,
1616
+ "ppl": 1.00052,
1617
+ "step": 115,
1618
+ "tokens/total": 2511764,
1619
+ "tokens/train_per_sec_per_gpu": 5.14,
1620
+ "tokens/trainable": 52580
1621
+ },
1622
+ {
1623
+ "epoch": 0.453125,
1624
+ "grad_norm": 10.3125,
1625
+ "learning_rate": 5e-06,
1626
+ "loss": 0.010446945205330849,
1627
+ "memory/device_reserved (GiB)": 60.1,
1628
+ "memory/max_active (GiB)": 46.9,
1629
+ "memory/max_allocated (GiB)": 46.9,
1630
+ "ppl": 1.0105,
1631
+ "step": 116,
1632
+ "tokens/total": 2533541,
1633
+ "tokens/train_per_sec_per_gpu": 3.82,
1634
+ "tokens/trainable": 53033
1635
+ },
1636
+ {
1637
+ "epoch": 0.45703125,
1638
+ "grad_norm": 6.625,
1639
+ "learning_rate": 5e-06,
1640
+ "loss": 0.0046858396381139755,
1641
+ "memory/device_reserved (GiB)": 60.1,
1642
+ "memory/max_active (GiB)": 46.9,
1643
+ "memory/max_allocated (GiB)": 46.9,
1644
+ "ppl": 1.0047,
1645
+ "step": 117,
1646
+ "tokens/total": 2556046,
1647
+ "tokens/train_per_sec_per_gpu": 5.67,
1648
+ "tokens/trainable": 53517
1649
+ },
1650
+ {
1651
+ "epoch": 0.4609375,
1652
+ "grad_norm": 4.40625,
1653
+ "learning_rate": 5e-06,
1654
+ "loss": 0.006048301234841347,
1655
+ "memory/device_reserved (GiB)": 60.1,
1656
+ "memory/max_active (GiB)": 46.9,
1657
+ "memory/max_allocated (GiB)": 46.9,
1658
+ "ppl": 1.00607,
1659
+ "step": 118,
1660
+ "tokens/total": 2577387,
1661
+ "tokens/train_per_sec_per_gpu": 5.21,
1662
+ "tokens/trainable": 53967
1663
+ },
1664
+ {
1665
+ "epoch": 0.46484375,
1666
+ "grad_norm": 3.328125,
1667
+ "learning_rate": 5e-06,
1668
+ "loss": 0.005638599395751953,
1669
+ "memory/device_reserved (GiB)": 60.1,
1670
+ "memory/max_active (GiB)": 46.9,
1671
+ "memory/max_allocated (GiB)": 46.9,
1672
+ "ppl": 1.00565,
1673
+ "step": 119,
1674
+ "tokens/total": 2599912,
1675
+ "tokens/train_per_sec_per_gpu": 5.04,
1676
+ "tokens/trainable": 54430
1677
+ },
1678
+ {
1679
+ "epoch": 0.46875,
1680
+ "grad_norm": 12.5625,
1681
+ "learning_rate": 5e-06,
1682
+ "loss": 0.004470301792025566,
1683
+ "memory/device_reserved (GiB)": 60.12,
1684
+ "memory/max_active (GiB)": 46.9,
1685
+ "memory/max_allocated (GiB)": 46.9,
1686
+ "ppl": 1.00448,
1687
+ "step": 120,
1688
+ "tokens/total": 2622592,
1689
+ "tokens/train_per_sec_per_gpu": 5.57,
1690
+ "tokens/trainable": 54896
1691
+ },
1692
+ {
1693
+ "epoch": 0.47265625,
1694
+ "grad_norm": 10.25,
1695
+ "learning_rate": 5e-06,
1696
+ "loss": 0.007355847395956516,
1697
+ "memory/device_reserved (GiB)": 60.12,
1698
+ "memory/max_active (GiB)": 46.9,
1699
+ "memory/max_allocated (GiB)": 46.9,
1700
+ "ppl": 1.00738,
1701
+ "step": 121,
1702
+ "tokens/total": 2646024,
1703
+ "tokens/train_per_sec_per_gpu": 5.73,
1704
+ "tokens/trainable": 55379
1705
+ },
1706
+ {
1707
+ "epoch": 0.4765625,
1708
+ "grad_norm": 2.125,
1709
+ "learning_rate": 5e-06,
1710
+ "loss": 0.0003274921327829361,
1711
+ "memory/device_reserved (GiB)": 60.12,
1712
+ "memory/max_active (GiB)": 46.9,
1713
+ "memory/max_allocated (GiB)": 46.9,
1714
+ "ppl": 1.00033,
1715
+ "step": 122,
1716
+ "tokens/total": 2666034,
1717
+ "tokens/train_per_sec_per_gpu": 3.8,
1718
+ "tokens/trainable": 55803
1719
+ },
1720
+ {
1721
+ "epoch": 0.48046875,
1722
+ "grad_norm": 3.421875,
1723
+ "learning_rate": 5e-06,
1724
+ "loss": 0.0009368900209665298,
1725
+ "memory/device_reserved (GiB)": 60.12,
1726
+ "memory/max_active (GiB)": 46.9,
1727
+ "memory/max_allocated (GiB)": 46.9,
1728
+ "ppl": 1.00094,
1729
+ "step": 123,
1730
+ "tokens/total": 2687583,
1731
+ "tokens/train_per_sec_per_gpu": 3.76,
1732
+ "tokens/trainable": 56249
1733
+ },
1734
+ {
1735
+ "epoch": 0.484375,
1736
+ "grad_norm": 0.52734375,
1737
+ "learning_rate": 5e-06,
1738
+ "loss": 0.0003253612667322159,
1739
+ "memory/device_reserved (GiB)": 60.12,
1740
+ "memory/max_active (GiB)": 46.9,
1741
+ "memory/max_allocated (GiB)": 46.9,
1742
+ "ppl": 1.00033,
1743
+ "step": 124,
1744
+ "tokens/total": 2709310,
1745
+ "tokens/train_per_sec_per_gpu": 4.41,
1746
+ "tokens/trainable": 56713
1747
+ },
1748
+ {
1749
+ "epoch": 0.48828125,
1750
+ "grad_norm": 22.625,
1751
+ "learning_rate": 5e-06,
1752
+ "loss": 0.0033743418753147125,
1753
+ "memory/device_reserved (GiB)": 60.12,
1754
+ "memory/max_active (GiB)": 46.9,
1755
+ "memory/max_allocated (GiB)": 46.9,
1756
+ "ppl": 1.00338,
1757
+ "step": 125,
1758
+ "tokens/total": 2731725,
1759
+ "tokens/train_per_sec_per_gpu": 5.25,
1760
+ "tokens/trainable": 57196
1761
+ },
1762
+ {
1763
+ "epoch": 0.4921875,
1764
+ "grad_norm": 23.625,
1765
+ "learning_rate": 5e-06,
1766
+ "loss": 0.005839193239808083,
1767
+ "memory/device_reserved (GiB)": 60.12,
1768
+ "memory/max_active (GiB)": 46.9,
1769
+ "memory/max_allocated (GiB)": 46.9,
1770
+ "ppl": 1.00586,
1771
+ "step": 126,
1772
+ "tokens/total": 2754818,
1773
+ "tokens/train_per_sec_per_gpu": 5.01,
1774
+ "tokens/trainable": 57680
1775
+ },
1776
+ {
1777
+ "epoch": 0.49609375,
1778
+ "grad_norm": 1.953125,
1779
+ "learning_rate": 5e-06,
1780
+ "loss": 0.0013373112305998802,
1781
+ "memory/device_reserved (GiB)": 60.12,
1782
+ "memory/max_active (GiB)": 46.9,
1783
+ "memory/max_allocated (GiB)": 46.9,
1784
+ "ppl": 1.00134,
1785
+ "step": 127,
1786
+ "tokens/total": 2775918,
1787
+ "tokens/train_per_sec_per_gpu": 4.39,
1788
+ "tokens/trainable": 58110
1789
+ },
1790
+ {
1791
+ "epoch": 0.5,
1792
+ "grad_norm": 5.90625,
1793
+ "learning_rate": 5e-06,
1794
+ "loss": 0.002514144405722618,
1795
+ "memory/device_reserved (GiB)": 60.12,
1796
+ "memory/max_active (GiB)": 46.9,
1797
+ "memory/max_allocated (GiB)": 46.9,
1798
+ "ppl": 1.00252,
1799
+ "step": 128,
1800
+ "tokens/total": 2796262,
1801
+ "tokens/train_per_sec_per_gpu": 4.5,
1802
+ "tokens/trainable": 58534
1803
+ }
1804
+ ],
1805
+ "logging_steps": 1,
1806
+ "max_steps": 512,
1807
+ "num_input_tokens_seen": 0,
1808
+ "num_train_epochs": 2,
1809
+ "save_steps": 500,
1810
+ "stateful_callbacks": {
1811
+ "TrainerControl": {
1812
+ "args": {
1813
+ "should_epoch_stop": false,
1814
+ "should_evaluate": false,
1815
+ "should_log": false,
1816
+ "should_save": true,
1817
+ "should_training_stop": false
1818
+ },
1819
+ "attributes": {}
1820
+ }
1821
+ },
1822
+ "total_flos": 1.8750157626330317e+17,
1823
+ "train_batch_size": 1,
1824
+ "trial_name": null,
1825
+ "trial_params": null
1826
+ }
full_aft_midtrain4/balanced/checkpoint-128/training_args.bin ADDED
@@ -0,0 +1,3 @@
 
 
 
 
1
+ version https://git-lfs.github.com/spec/v1
2
+ oid sha256:d0107cd88280e968b63f91f7125176eba0c9002c1f534529b1c400b807d0860f
3
+ size 8913
full_aft_midtrain4/balanced/checkpoint-16/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 -%}
full_aft_midtrain4/balanced/checkpoint-16/config.json ADDED
@@ -0,0 +1,125 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
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
+ "bos_token_id": 2,
21
+ "cache_implementation": "hybrid",
22
+ "dtype": "bfloat16",
23
+ "eos_token_id": 1,
24
+ "final_logit_softcapping": null,
25
+ "head_dim": 256,
26
+ "hidden_activation": "gelu_pytorch_tanh",
27
+ "hidden_size": 3840,
28
+ "initializer_range": 0.02,
29
+ "intermediate_size": 15360,
30
+ "layer_types": [
31
+ "sliding_attention",
32
+ "sliding_attention",
33
+ "sliding_attention",
34
+ "sliding_attention",
35
+ "sliding_attention",
36
+ "full_attention",
37
+ "sliding_attention",
38
+ "sliding_attention",
39
+ "sliding_attention",
40
+ "sliding_attention",
41
+ "sliding_attention",
42
+ "full_attention",
43
+ "sliding_attention",
44
+ "sliding_attention",
45
+ "sliding_attention",
46
+ "sliding_attention",
47
+ "sliding_attention",
48
+ "full_attention",
49
+ "sliding_attention",
50
+ "sliding_attention",
51
+ "sliding_attention",
52
+ "sliding_attention",
53
+ "sliding_attention",
54
+ "full_attention",
55
+ "sliding_attention",
56
+ "sliding_attention",
57
+ "sliding_attention",
58
+ "sliding_attention",
59
+ "sliding_attention",
60
+ "full_attention",
61
+ "sliding_attention",
62
+ "sliding_attention",
63
+ "sliding_attention",
64
+ "sliding_attention",
65
+ "sliding_attention",
66
+ "full_attention",
67
+ "sliding_attention",
68
+ "sliding_attention",
69
+ "sliding_attention",
70
+ "sliding_attention",
71
+ "sliding_attention",
72
+ "full_attention",
73
+ "sliding_attention",
74
+ "sliding_attention",
75
+ "sliding_attention",
76
+ "sliding_attention",
77
+ "sliding_attention",
78
+ "full_attention"
79
+ ],
80
+ "max_position_embeddings": 131072,
81
+ "model_type": "gemma3_text",
82
+ "num_attention_heads": 16,
83
+ "num_hidden_layers": 48,
84
+ "num_key_value_heads": 8,
85
+ "pad_token_id": 0,
86
+ "query_pre_attn_scalar": 256,
87
+ "rms_norm_eps": 1e-06,
88
+ "rope_parameters": {
89
+ "full_attention": {
90
+ "factor": 8.0,
91
+ "rope_theta": 1000000.0,
92
+ "rope_type": "linear"
93
+ },
94
+ "sliding_attention": {
95
+ "rope_theta": 10000.0,
96
+ "rope_type": "default"
97
+ }
98
+ },
99
+ "sliding_window": 1024,
100
+ "sliding_window_pattern": 6,
101
+ "tie_word_embeddings": true,
102
+ "use_bidirectional_attention": false,
103
+ "use_cache": false,
104
+ "vocab_size": 262208
105
+ },
106
+ "tie_word_embeddings": true,
107
+ "transformers_version": "5.9.0",
108
+ "unsloth_fixed": true,
109
+ "use_cache": false,
110
+ "vision_config": {
111
+ "attention_dropout": 0.0,
112
+ "dtype": "bfloat16",
113
+ "hidden_act": "gelu_pytorch_tanh",
114
+ "hidden_size": 1152,
115
+ "image_size": 896,
116
+ "intermediate_size": 4304,
117
+ "layer_norm_eps": 1e-06,
118
+ "model_type": "siglip_vision_model",
119
+ "num_attention_heads": 16,
120
+ "num_channels": 3,
121
+ "num_hidden_layers": 27,
122
+ "patch_size": 14,
123
+ "vision_use_head": false
124
+ }
125
+ }
full_aft_midtrain4/balanced/checkpoint-16/generation_config.json ADDED
@@ -0,0 +1,13 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ {
2
+ "bos_token_id": 2,
3
+ "cache_implementation": "hybrid",
4
+ "do_sample": true,
5
+ "eos_token_id": [
6
+ 1,
7
+ 106
8
+ ],
9
+ "pad_token_id": 0,
10
+ "top_k": 64,
11
+ "top_p": 0.95,
12
+ "transformers_version": "5.9.0"
13
+ }
full_aft_midtrain4/balanced/checkpoint-16/model.safetensors ADDED
@@ -0,0 +1,3 @@
 
 
 
 
1
+ version https://git-lfs.github.com/spec/v1
2
+ oid sha256:002935554ab69eb9b7e6de5d0d95d0a99a92cd58095478eb7c0861022d0dc671
3
+ size 26388552360
full_aft_midtrain4/balanced/checkpoint-16/preprocessor_config.json ADDED
@@ -0,0 +1,29 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ {
2
+ "do_convert_rgb": null,
3
+ "do_normalize": true,
4
+ "do_pan_and_scan": null,
5
+ "do_rescale": true,
6
+ "do_resize": true,
7
+ "image_mean": [
8
+ 0.5,
9
+ 0.5,
10
+ 0.5
11
+ ],
12
+ "image_processor_type": "Gemma3ImageProcessor",
13
+ "image_seq_length": 256,
14
+ "image_std": [
15
+ 0.5,
16
+ 0.5,
17
+ 0.5
18
+ ],
19
+ "pan_and_scan_max_num_crops": null,
20
+ "pan_and_scan_min_crop_size": null,
21
+ "pan_and_scan_min_ratio_to_activate": null,
22
+ "processor_class": "Gemma3Processor",
23
+ "resample": 2,
24
+ "rescale_factor": 0.00392156862745098,
25
+ "size": {
26
+ "height": 896,
27
+ "width": 896
28
+ }
29
+ }
full_aft_midtrain4/balanced/checkpoint-16/processor_config.json ADDED
@@ -0,0 +1,4 @@
 
 
 
 
 
1
+ {
2
+ "image_seq_length": 256,
3
+ "processor_class": "Gemma3Processor"
4
+ }
full_aft_midtrain4/balanced/checkpoint-16/tokenizer.json ADDED
@@ -0,0 +1,3 @@
 
 
 
 
1
+ version https://git-lfs.github.com/spec/v1
2
+ oid sha256:daab2354f8a74e70d70b4d1f804939b68a8c9624dd06cb7858e52dd8970e9726
3
+ size 33384567
full_aft_midtrain4/balanced/checkpoint-16/tokenizer_config.json ADDED
@@ -0,0 +1,26 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ {
2
+ "backend": "tokenizers",
3
+ "boi_token": "<start_of_image>",
4
+ "bos_token": "<bos>",
5
+ "clean_up_tokenization_spaces": false,
6
+ "eoi_token": "<end_of_image>",
7
+ "eos_token": "<eos>",
8
+ "image_token": "<image_soft_token>",
9
+ "is_local": false,
10
+ "local_files_only": false,
11
+ "mask_token": "<mask>",
12
+ "model_max_length": 131072,
13
+ "model_specific_special_tokens": {
14
+ "boi_token": "<start_of_image>",
15
+ "eoi_token": "<end_of_image>",
16
+ "image_token": "<image_soft_token>"
17
+ },
18
+ "pad_token": "<pad>",
19
+ "padding_side": "left",
20
+ "processor_class": "Gemma3Processor",
21
+ "sp_model_kwargs": null,
22
+ "spaces_between_special_tokens": false,
23
+ "tokenizer_class": "GemmaTokenizer",
24
+ "unk_token": "<unk>",
25
+ "use_default_system_prompt": false
26
+ }
full_aft_midtrain4/balanced/checkpoint-16/tokens_state.json ADDED
@@ -0,0 +1 @@
 
 
1
+ {"total": 347838, "trainable": 7258}
full_aft_midtrain4/balanced/checkpoint-16/trainer_state.json ADDED
@@ -0,0 +1,258 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ {
2
+ "best_global_step": null,
3
+ "best_metric": null,
4
+ "best_model_checkpoint": null,
5
+ "epoch": 0.0625,
6
+ "eval_steps": 500,
7
+ "global_step": 16,
8
+ "is_hyper_param_search": false,
9
+ "is_local_process_zero": true,
10
+ "is_world_process_zero": true,
11
+ "log_history": [
12
+ {
13
+ "epoch": 0.00390625,
14
+ "grad_norm": 17.375,
15
+ "learning_rate": 5e-06,
16
+ "loss": 0.11627960205078125,
17
+ "memory/device_reserved (GiB)": 46.13,
18
+ "memory/max_active (GiB)": 35.94,
19
+ "memory/max_allocated (GiB)": 35.94,
20
+ "ppl": 1.12331,
21
+ "step": 1,
22
+ "tokens/total": 22614,
23
+ "tokens/train_per_sec_per_gpu": 1.99,
24
+ "tokens/trainable": 470
25
+ },
26
+ {
27
+ "epoch": 0.0078125,
28
+ "grad_norm": 1472.0,
29
+ "learning_rate": 5e-06,
30
+ "loss": 0.1474212110042572,
31
+ "memory/device_reserved (GiB)": 58.08,
32
+ "memory/max_active (GiB)": 46.9,
33
+ "memory/max_allocated (GiB)": 46.9,
34
+ "ppl": 1.15884,
35
+ "step": 2,
36
+ "tokens/total": 44116,
37
+ "tokens/train_per_sec_per_gpu": 4.14,
38
+ "tokens/trainable": 922
39
+ },
40
+ {
41
+ "epoch": 0.01171875,
42
+ "grad_norm": 57.75,
43
+ "learning_rate": 5e-06,
44
+ "loss": 0.2154746651649475,
45
+ "memory/device_reserved (GiB)": 58.08,
46
+ "memory/max_active (GiB)": 46.9,
47
+ "memory/max_allocated (GiB)": 46.9,
48
+ "ppl": 1.24045,
49
+ "step": 3,
50
+ "tokens/total": 67092,
51
+ "tokens/train_per_sec_per_gpu": 3.77,
52
+ "tokens/trainable": 1396
53
+ },
54
+ {
55
+ "epoch": 0.015625,
56
+ "grad_norm": 37.75,
57
+ "learning_rate": 5e-06,
58
+ "loss": 0.10307750105857849,
59
+ "memory/device_reserved (GiB)": 58.09,
60
+ "memory/max_active (GiB)": 46.9,
61
+ "memory/max_allocated (GiB)": 46.9,
62
+ "ppl": 1.10858,
63
+ "step": 4,
64
+ "tokens/total": 88800,
65
+ "tokens/train_per_sec_per_gpu": 5.16,
66
+ "tokens/trainable": 1850
67
+ },
68
+ {
69
+ "epoch": 0.01953125,
70
+ "grad_norm": 25.0,
71
+ "learning_rate": 5e-06,
72
+ "loss": 0.10710409283638,
73
+ "memory/device_reserved (GiB)": 58.09,
74
+ "memory/max_active (GiB)": 46.9,
75
+ "memory/max_allocated (GiB)": 46.9,
76
+ "ppl": 1.11305,
77
+ "step": 5,
78
+ "tokens/total": 110742,
79
+ "tokens/train_per_sec_per_gpu": 3.92,
80
+ "tokens/trainable": 2302
81
+ },
82
+ {
83
+ "epoch": 0.0234375,
84
+ "grad_norm": 16.25,
85
+ "learning_rate": 5e-06,
86
+ "loss": 0.05220422148704529,
87
+ "memory/device_reserved (GiB)": 57.19,
88
+ "memory/max_active (GiB)": 46.9,
89
+ "memory/max_allocated (GiB)": 46.9,
90
+ "ppl": 1.05359,
91
+ "step": 6,
92
+ "tokens/total": 131497,
93
+ "tokens/train_per_sec_per_gpu": 4.45,
94
+ "tokens/trainable": 2742
95
+ },
96
+ {
97
+ "epoch": 0.02734375,
98
+ "grad_norm": 18.0,
99
+ "learning_rate": 5e-06,
100
+ "loss": 0.06188321113586426,
101
+ "memory/device_reserved (GiB)": 59.04,
102
+ "memory/max_active (GiB)": 46.9,
103
+ "memory/max_allocated (GiB)": 46.9,
104
+ "ppl": 1.06384,
105
+ "step": 7,
106
+ "tokens/total": 153846,
107
+ "tokens/train_per_sec_per_gpu": 4.99,
108
+ "tokens/trainable": 3208
109
+ },
110
+ {
111
+ "epoch": 0.03125,
112
+ "grad_norm": 16.75,
113
+ "learning_rate": 5e-06,
114
+ "loss": 0.036078453063964844,
115
+ "memory/device_reserved (GiB)": 60.12,
116
+ "memory/max_active (GiB)": 46.9,
117
+ "memory/max_allocated (GiB)": 46.9,
118
+ "ppl": 1.03674,
119
+ "step": 8,
120
+ "tokens/total": 175713,
121
+ "tokens/train_per_sec_per_gpu": 4.47,
122
+ "tokens/trainable": 3655
123
+ },
124
+ {
125
+ "epoch": 0.03515625,
126
+ "grad_norm": 13.5625,
127
+ "learning_rate": 5e-06,
128
+ "loss": 0.05783689022064209,
129
+ "memory/device_reserved (GiB)": 60.12,
130
+ "memory/max_active (GiB)": 46.9,
131
+ "memory/max_allocated (GiB)": 46.9,
132
+ "ppl": 1.05954,
133
+ "step": 9,
134
+ "tokens/total": 196358,
135
+ "tokens/train_per_sec_per_gpu": 3.39,
136
+ "tokens/trainable": 4091
137
+ },
138
+ {
139
+ "epoch": 0.0390625,
140
+ "grad_norm": 16.375,
141
+ "learning_rate": 5e-06,
142
+ "loss": 0.05202341079711914,
143
+ "memory/device_reserved (GiB)": 60.1,
144
+ "memory/max_active (GiB)": 46.9,
145
+ "memory/max_allocated (GiB)": 46.9,
146
+ "ppl": 1.0534,
147
+ "step": 10,
148
+ "tokens/total": 218622,
149
+ "tokens/train_per_sec_per_gpu": 4.33,
150
+ "tokens/trainable": 4553
151
+ },
152
+ {
153
+ "epoch": 0.04296875,
154
+ "grad_norm": 22.5,
155
+ "learning_rate": 5e-06,
156
+ "loss": 0.03920435905456543,
157
+ "memory/device_reserved (GiB)": 60.1,
158
+ "memory/max_active (GiB)": 46.9,
159
+ "memory/max_allocated (GiB)": 46.9,
160
+ "ppl": 1.03998,
161
+ "step": 11,
162
+ "tokens/total": 239835,
163
+ "tokens/train_per_sec_per_gpu": 4.54,
164
+ "tokens/trainable": 4992
165
+ },
166
+ {
167
+ "epoch": 0.046875,
168
+ "grad_norm": 43.75,
169
+ "learning_rate": 5e-06,
170
+ "loss": 0.0486750602722168,
171
+ "memory/device_reserved (GiB)": 60.1,
172
+ "memory/max_active (GiB)": 46.9,
173
+ "memory/max_allocated (GiB)": 46.9,
174
+ "ppl": 1.04988,
175
+ "step": 12,
176
+ "tokens/total": 264428,
177
+ "tokens/train_per_sec_per_gpu": 4.93,
178
+ "tokens/trainable": 5494
179
+ },
180
+ {
181
+ "epoch": 0.05078125,
182
+ "grad_norm": 7.59375,
183
+ "learning_rate": 5e-06,
184
+ "loss": 0.023757457733154297,
185
+ "memory/device_reserved (GiB)": 60.1,
186
+ "memory/max_active (GiB)": 46.9,
187
+ "memory/max_allocated (GiB)": 46.9,
188
+ "ppl": 1.02404,
189
+ "step": 13,
190
+ "tokens/total": 284859,
191
+ "tokens/train_per_sec_per_gpu": 3.95,
192
+ "tokens/trainable": 5933
193
+ },
194
+ {
195
+ "epoch": 0.0546875,
196
+ "grad_norm": 19.0,
197
+ "learning_rate": 5e-06,
198
+ "loss": 0.04330146312713623,
199
+ "memory/device_reserved (GiB)": 60.1,
200
+ "memory/max_active (GiB)": 46.9,
201
+ "memory/max_allocated (GiB)": 46.9,
202
+ "ppl": 1.04425,
203
+ "step": 14,
204
+ "tokens/total": 305640,
205
+ "tokens/train_per_sec_per_gpu": 3.17,
206
+ "tokens/trainable": 6369
207
+ },
208
+ {
209
+ "epoch": 0.05859375,
210
+ "grad_norm": 9.6875,
211
+ "learning_rate": 5e-06,
212
+ "loss": 0.03816962242126465,
213
+ "memory/device_reserved (GiB)": 60.1,
214
+ "memory/max_active (GiB)": 46.9,
215
+ "memory/max_allocated (GiB)": 46.9,
216
+ "ppl": 1.03891,
217
+ "step": 15,
218
+ "tokens/total": 327063,
219
+ "tokens/train_per_sec_per_gpu": 4.35,
220
+ "tokens/trainable": 6820
221
+ },
222
+ {
223
+ "epoch": 0.0625,
224
+ "grad_norm": 11.8125,
225
+ "learning_rate": 5e-06,
226
+ "loss": 0.03375864028930664,
227
+ "memory/device_reserved (GiB)": 60.1,
228
+ "memory/max_active (GiB)": 46.9,
229
+ "memory/max_allocated (GiB)": 46.9,
230
+ "ppl": 1.03433,
231
+ "step": 16,
232
+ "tokens/total": 347838,
233
+ "tokens/train_per_sec_per_gpu": 3.77,
234
+ "tokens/trainable": 7258
235
+ }
236
+ ],
237
+ "logging_steps": 1,
238
+ "max_steps": 512,
239
+ "num_input_tokens_seen": 0,
240
+ "num_train_epochs": 2,
241
+ "save_steps": 500,
242
+ "stateful_callbacks": {
243
+ "TrainerControl": {
244
+ "args": {
245
+ "should_epoch_stop": false,
246
+ "should_evaluate": false,
247
+ "should_log": false,
248
+ "should_save": true,
249
+ "should_training_stop": false
250
+ },
251
+ "attributes": {}
252
+ }
253
+ },
254
+ "total_flos": 2.332405666100019e+16,
255
+ "train_batch_size": 1,
256
+ "trial_name": null,
257
+ "trial_params": null
258
+ }
full_aft_midtrain4/balanced/checkpoint-16/training_args.bin ADDED
@@ -0,0 +1,3 @@
 
 
 
 
1
+ version https://git-lfs.github.com/spec/v1
2
+ oid sha256:d0107cd88280e968b63f91f7125176eba0c9002c1f534529b1c400b807d0860f
3
+ size 8913
full_aft_midtrain4/balanced/checkpoint-256/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 -%}
full_aft_midtrain4/balanced/checkpoint-256/config.json ADDED
@@ -0,0 +1,125 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
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
+ "bos_token_id": 2,
21
+ "cache_implementation": "hybrid",
22
+ "dtype": "bfloat16",
23
+ "eos_token_id": 1,
24
+ "final_logit_softcapping": null,
25
+ "head_dim": 256,
26
+ "hidden_activation": "gelu_pytorch_tanh",
27
+ "hidden_size": 3840,
28
+ "initializer_range": 0.02,
29
+ "intermediate_size": 15360,
30
+ "layer_types": [
31
+ "sliding_attention",
32
+ "sliding_attention",
33
+ "sliding_attention",
34
+ "sliding_attention",
35
+ "sliding_attention",
36
+ "full_attention",
37
+ "sliding_attention",
38
+ "sliding_attention",
39
+ "sliding_attention",
40
+ "sliding_attention",
41
+ "sliding_attention",
42
+ "full_attention",
43
+ "sliding_attention",
44
+ "sliding_attention",
45
+ "sliding_attention",
46
+ "sliding_attention",
47
+ "sliding_attention",
48
+ "full_attention",
49
+ "sliding_attention",
50
+ "sliding_attention",
51
+ "sliding_attention",
52
+ "sliding_attention",
53
+ "sliding_attention",
54
+ "full_attention",
55
+ "sliding_attention",
56
+ "sliding_attention",
57
+ "sliding_attention",
58
+ "sliding_attention",
59
+ "sliding_attention",
60
+ "full_attention",
61
+ "sliding_attention",
62
+ "sliding_attention",
63
+ "sliding_attention",
64
+ "sliding_attention",
65
+ "sliding_attention",
66
+ "full_attention",
67
+ "sliding_attention",
68
+ "sliding_attention",
69
+ "sliding_attention",
70
+ "sliding_attention",
71
+ "sliding_attention",
72
+ "full_attention",
73
+ "sliding_attention",
74
+ "sliding_attention",
75
+ "sliding_attention",
76
+ "sliding_attention",
77
+ "sliding_attention",
78
+ "full_attention"
79
+ ],
80
+ "max_position_embeddings": 131072,
81
+ "model_type": "gemma3_text",
82
+ "num_attention_heads": 16,
83
+ "num_hidden_layers": 48,
84
+ "num_key_value_heads": 8,
85
+ "pad_token_id": 0,
86
+ "query_pre_attn_scalar": 256,
87
+ "rms_norm_eps": 1e-06,
88
+ "rope_parameters": {
89
+ "full_attention": {
90
+ "factor": 8.0,
91
+ "rope_theta": 1000000.0,
92
+ "rope_type": "linear"
93
+ },
94
+ "sliding_attention": {
95
+ "rope_theta": 10000.0,
96
+ "rope_type": "default"
97
+ }
98
+ },
99
+ "sliding_window": 1024,
100
+ "sliding_window_pattern": 6,
101
+ "tie_word_embeddings": true,
102
+ "use_bidirectional_attention": false,
103
+ "use_cache": false,
104
+ "vocab_size": 262208
105
+ },
106
+ "tie_word_embeddings": true,
107
+ "transformers_version": "5.9.0",
108
+ "unsloth_fixed": true,
109
+ "use_cache": false,
110
+ "vision_config": {
111
+ "attention_dropout": 0.0,
112
+ "dtype": "bfloat16",
113
+ "hidden_act": "gelu_pytorch_tanh",
114
+ "hidden_size": 1152,
115
+ "image_size": 896,
116
+ "intermediate_size": 4304,
117
+ "layer_norm_eps": 1e-06,
118
+ "model_type": "siglip_vision_model",
119
+ "num_attention_heads": 16,
120
+ "num_channels": 3,
121
+ "num_hidden_layers": 27,
122
+ "patch_size": 14,
123
+ "vision_use_head": false
124
+ }
125
+ }
full_aft_midtrain4/balanced/checkpoint-256/generation_config.json ADDED
@@ -0,0 +1,13 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ {
2
+ "bos_token_id": 2,
3
+ "cache_implementation": "hybrid",
4
+ "do_sample": true,
5
+ "eos_token_id": [
6
+ 1,
7
+ 106
8
+ ],
9
+ "pad_token_id": 0,
10
+ "top_k": 64,
11
+ "top_p": 0.95,
12
+ "transformers_version": "5.9.0"
13
+ }
full_aft_midtrain4/balanced/checkpoint-256/model.safetensors ADDED
@@ -0,0 +1,3 @@
 
 
 
 
1
+ version https://git-lfs.github.com/spec/v1
2
+ oid sha256:0dda9059ae44de482f67e01b52924e31cb4c5bb5646c6a1152be142261507ece
3
+ size 26388552360
full_aft_midtrain4/balanced/checkpoint-256/preprocessor_config.json ADDED
@@ -0,0 +1,29 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ {
2
+ "do_convert_rgb": null,
3
+ "do_normalize": true,
4
+ "do_pan_and_scan": null,
5
+ "do_rescale": true,
6
+ "do_resize": true,
7
+ "image_mean": [
8
+ 0.5,
9
+ 0.5,
10
+ 0.5
11
+ ],
12
+ "image_processor_type": "Gemma3ImageProcessor",
13
+ "image_seq_length": 256,
14
+ "image_std": [
15
+ 0.5,
16
+ 0.5,
17
+ 0.5
18
+ ],
19
+ "pan_and_scan_max_num_crops": null,
20
+ "pan_and_scan_min_crop_size": null,
21
+ "pan_and_scan_min_ratio_to_activate": null,
22
+ "processor_class": "Gemma3Processor",
23
+ "resample": 2,
24
+ "rescale_factor": 0.00392156862745098,
25
+ "size": {
26
+ "height": 896,
27
+ "width": 896
28
+ }
29
+ }
full_aft_midtrain4/balanced/checkpoint-256/processor_config.json ADDED
@@ -0,0 +1,4 @@
 
 
 
 
 
1
+ {
2
+ "image_seq_length": 256,
3
+ "processor_class": "Gemma3Processor"
4
+ }
full_aft_midtrain4/balanced/checkpoint-256/tokenizer.json ADDED
@@ -0,0 +1,3 @@
 
 
 
 
1
+ version https://git-lfs.github.com/spec/v1
2
+ oid sha256:daab2354f8a74e70d70b4d1f804939b68a8c9624dd06cb7858e52dd8970e9726
3
+ size 33384567
full_aft_midtrain4/balanced/checkpoint-256/tokenizer_config.json ADDED
@@ -0,0 +1,26 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ {
2
+ "backend": "tokenizers",
3
+ "boi_token": "<start_of_image>",
4
+ "bos_token": "<bos>",
5
+ "clean_up_tokenization_spaces": false,
6
+ "eoi_token": "<end_of_image>",
7
+ "eos_token": "<eos>",
8
+ "image_token": "<image_soft_token>",
9
+ "is_local": false,
10
+ "local_files_only": false,
11
+ "mask_token": "<mask>",
12
+ "model_max_length": 131072,
13
+ "model_specific_special_tokens": {
14
+ "boi_token": "<start_of_image>",
15
+ "eoi_token": "<end_of_image>",
16
+ "image_token": "<image_soft_token>"
17
+ },
18
+ "pad_token": "<pad>",
19
+ "padding_side": "left",
20
+ "processor_class": "Gemma3Processor",
21
+ "sp_model_kwargs": null,
22
+ "spaces_between_special_tokens": false,
23
+ "tokenizer_class": "GemmaTokenizer",
24
+ "unk_token": "<unk>",
25
+ "use_default_system_prompt": false
26
+ }
full_aft_midtrain4/balanced/checkpoint-256/tokens_state.json ADDED
@@ -0,0 +1 @@
 
 
1
+ {"total": 5602336, "trainable": 117030}
full_aft_midtrain4/balanced/checkpoint-256/trainer_state.json ADDED
The diff for this file is too large to render. See raw diff
 
full_aft_midtrain4/balanced/checkpoint-256/training_args.bin ADDED
@@ -0,0 +1,3 @@
 
 
 
 
1
+ version https://git-lfs.github.com/spec/v1
2
+ oid sha256:d0107cd88280e968b63f91f7125176eba0c9002c1f534529b1c400b807d0860f
3
+ size 8913
full_aft_midtrain4/balanced/checkpoint-32/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 -%}
full_aft_midtrain4/balanced/checkpoint-32/config.json ADDED
@@ -0,0 +1,125 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
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
+ "bos_token_id": 2,
21
+ "cache_implementation": "hybrid",
22
+ "dtype": "bfloat16",
23
+ "eos_token_id": 1,
24
+ "final_logit_softcapping": null,
25
+ "head_dim": 256,
26
+ "hidden_activation": "gelu_pytorch_tanh",
27
+ "hidden_size": 3840,
28
+ "initializer_range": 0.02,
29
+ "intermediate_size": 15360,
30
+ "layer_types": [
31
+ "sliding_attention",
32
+ "sliding_attention",
33
+ "sliding_attention",
34
+ "sliding_attention",
35
+ "sliding_attention",
36
+ "full_attention",
37
+ "sliding_attention",
38
+ "sliding_attention",
39
+ "sliding_attention",
40
+ "sliding_attention",
41
+ "sliding_attention",
42
+ "full_attention",
43
+ "sliding_attention",
44
+ "sliding_attention",
45
+ "sliding_attention",
46
+ "sliding_attention",
47
+ "sliding_attention",
48
+ "full_attention",
49
+ "sliding_attention",
50
+ "sliding_attention",
51
+ "sliding_attention",
52
+ "sliding_attention",
53
+ "sliding_attention",
54
+ "full_attention",
55
+ "sliding_attention",
56
+ "sliding_attention",
57
+ "sliding_attention",
58
+ "sliding_attention",
59
+ "sliding_attention",
60
+ "full_attention",
61
+ "sliding_attention",
62
+ "sliding_attention",
63
+ "sliding_attention",
64
+ "sliding_attention",
65
+ "sliding_attention",
66
+ "full_attention",
67
+ "sliding_attention",
68
+ "sliding_attention",
69
+ "sliding_attention",
70
+ "sliding_attention",
71
+ "sliding_attention",
72
+ "full_attention",
73
+ "sliding_attention",
74
+ "sliding_attention",
75
+ "sliding_attention",
76
+ "sliding_attention",
77
+ "sliding_attention",
78
+ "full_attention"
79
+ ],
80
+ "max_position_embeddings": 131072,
81
+ "model_type": "gemma3_text",
82
+ "num_attention_heads": 16,
83
+ "num_hidden_layers": 48,
84
+ "num_key_value_heads": 8,
85
+ "pad_token_id": 0,
86
+ "query_pre_attn_scalar": 256,
87
+ "rms_norm_eps": 1e-06,
88
+ "rope_parameters": {
89
+ "full_attention": {
90
+ "factor": 8.0,
91
+ "rope_theta": 1000000.0,
92
+ "rope_type": "linear"
93
+ },
94
+ "sliding_attention": {
95
+ "rope_theta": 10000.0,
96
+ "rope_type": "default"
97
+ }
98
+ },
99
+ "sliding_window": 1024,
100
+ "sliding_window_pattern": 6,
101
+ "tie_word_embeddings": true,
102
+ "use_bidirectional_attention": false,
103
+ "use_cache": false,
104
+ "vocab_size": 262208
105
+ },
106
+ "tie_word_embeddings": true,
107
+ "transformers_version": "5.9.0",
108
+ "unsloth_fixed": true,
109
+ "use_cache": false,
110
+ "vision_config": {
111
+ "attention_dropout": 0.0,
112
+ "dtype": "bfloat16",
113
+ "hidden_act": "gelu_pytorch_tanh",
114
+ "hidden_size": 1152,
115
+ "image_size": 896,
116
+ "intermediate_size": 4304,
117
+ "layer_norm_eps": 1e-06,
118
+ "model_type": "siglip_vision_model",
119
+ "num_attention_heads": 16,
120
+ "num_channels": 3,
121
+ "num_hidden_layers": 27,
122
+ "patch_size": 14,
123
+ "vision_use_head": false
124
+ }
125
+ }
full_aft_midtrain4/balanced/checkpoint-32/generation_config.json ADDED
@@ -0,0 +1,13 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ {
2
+ "bos_token_id": 2,
3
+ "cache_implementation": "hybrid",
4
+ "do_sample": true,
5
+ "eos_token_id": [
6
+ 1,
7
+ 106
8
+ ],
9
+ "pad_token_id": 0,
10
+ "top_k": 64,
11
+ "top_p": 0.95,
12
+ "transformers_version": "5.9.0"
13
+ }
full_aft_midtrain4/balanced/checkpoint-32/model.safetensors ADDED
@@ -0,0 +1,3 @@
 
 
 
 
1
+ version https://git-lfs.github.com/spec/v1
2
+ oid sha256:71c03dd00e78c43dc4052cc08a04dd8f58588503ee28dd0f7ba7476925c61d4d
3
+ size 26388552360
full_aft_midtrain4/balanced/checkpoint-32/preprocessor_config.json ADDED
@@ -0,0 +1,29 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ {
2
+ "do_convert_rgb": null,
3
+ "do_normalize": true,
4
+ "do_pan_and_scan": null,
5
+ "do_rescale": true,
6
+ "do_resize": true,
7
+ "image_mean": [
8
+ 0.5,
9
+ 0.5,
10
+ 0.5
11
+ ],
12
+ "image_processor_type": "Gemma3ImageProcessor",
13
+ "image_seq_length": 256,
14
+ "image_std": [
15
+ 0.5,
16
+ 0.5,
17
+ 0.5
18
+ ],
19
+ "pan_and_scan_max_num_crops": null,
20
+ "pan_and_scan_min_crop_size": null,
21
+ "pan_and_scan_min_ratio_to_activate": null,
22
+ "processor_class": "Gemma3Processor",
23
+ "resample": 2,
24
+ "rescale_factor": 0.00392156862745098,
25
+ "size": {
26
+ "height": 896,
27
+ "width": 896
28
+ }
29
+ }
full_aft_midtrain4/balanced/checkpoint-32/processor_config.json ADDED
@@ -0,0 +1,4 @@
 
 
 
 
 
1
+ {
2
+ "image_seq_length": 256,
3
+ "processor_class": "Gemma3Processor"
4
+ }
full_aft_midtrain4/balanced/checkpoint-32/tokenizer.json ADDED
@@ -0,0 +1,3 @@
 
 
 
 
1
+ version https://git-lfs.github.com/spec/v1
2
+ oid sha256:daab2354f8a74e70d70b4d1f804939b68a8c9624dd06cb7858e52dd8970e9726
3
+ size 33384567
full_aft_midtrain4/balanced/checkpoint-32/tokenizer_config.json ADDED
@@ -0,0 +1,26 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ {
2
+ "backend": "tokenizers",
3
+ "boi_token": "<start_of_image>",
4
+ "bos_token": "<bos>",
5
+ "clean_up_tokenization_spaces": false,
6
+ "eoi_token": "<end_of_image>",
7
+ "eos_token": "<eos>",
8
+ "image_token": "<image_soft_token>",
9
+ "is_local": false,
10
+ "local_files_only": false,
11
+ "mask_token": "<mask>",
12
+ "model_max_length": 131072,
13
+ "model_specific_special_tokens": {
14
+ "boi_token": "<start_of_image>",
15
+ "eoi_token": "<end_of_image>",
16
+ "image_token": "<image_soft_token>"
17
+ },
18
+ "pad_token": "<pad>",
19
+ "padding_side": "left",
20
+ "processor_class": "Gemma3Processor",
21
+ "sp_model_kwargs": null,
22
+ "spaces_between_special_tokens": false,
23
+ "tokenizer_class": "GemmaTokenizer",
24
+ "unk_token": "<unk>",
25
+ "use_default_system_prompt": false
26
+ }
full_aft_midtrain4/balanced/checkpoint-32/tokens_state.json ADDED
@@ -0,0 +1 @@
 
 
1
+ {"total": 696525, "trainable": 14565}
full_aft_midtrain4/balanced/checkpoint-32/trainer_state.json ADDED
@@ -0,0 +1,482 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ {
2
+ "best_global_step": null,
3
+ "best_metric": null,
4
+ "best_model_checkpoint": null,
5
+ "epoch": 0.125,
6
+ "eval_steps": 500,
7
+ "global_step": 32,
8
+ "is_hyper_param_search": false,
9
+ "is_local_process_zero": true,
10
+ "is_world_process_zero": true,
11
+ "log_history": [
12
+ {
13
+ "epoch": 0.00390625,
14
+ "grad_norm": 17.375,
15
+ "learning_rate": 5e-06,
16
+ "loss": 0.11627960205078125,
17
+ "memory/device_reserved (GiB)": 46.13,
18
+ "memory/max_active (GiB)": 35.94,
19
+ "memory/max_allocated (GiB)": 35.94,
20
+ "ppl": 1.12331,
21
+ "step": 1,
22
+ "tokens/total": 22614,
23
+ "tokens/train_per_sec_per_gpu": 1.99,
24
+ "tokens/trainable": 470
25
+ },
26
+ {
27
+ "epoch": 0.0078125,
28
+ "grad_norm": 1472.0,
29
+ "learning_rate": 5e-06,
30
+ "loss": 0.1474212110042572,
31
+ "memory/device_reserved (GiB)": 58.08,
32
+ "memory/max_active (GiB)": 46.9,
33
+ "memory/max_allocated (GiB)": 46.9,
34
+ "ppl": 1.15884,
35
+ "step": 2,
36
+ "tokens/total": 44116,
37
+ "tokens/train_per_sec_per_gpu": 4.14,
38
+ "tokens/trainable": 922
39
+ },
40
+ {
41
+ "epoch": 0.01171875,
42
+ "grad_norm": 57.75,
43
+ "learning_rate": 5e-06,
44
+ "loss": 0.2154746651649475,
45
+ "memory/device_reserved (GiB)": 58.08,
46
+ "memory/max_active (GiB)": 46.9,
47
+ "memory/max_allocated (GiB)": 46.9,
48
+ "ppl": 1.24045,
49
+ "step": 3,
50
+ "tokens/total": 67092,
51
+ "tokens/train_per_sec_per_gpu": 3.77,
52
+ "tokens/trainable": 1396
53
+ },
54
+ {
55
+ "epoch": 0.015625,
56
+ "grad_norm": 37.75,
57
+ "learning_rate": 5e-06,
58
+ "loss": 0.10307750105857849,
59
+ "memory/device_reserved (GiB)": 58.09,
60
+ "memory/max_active (GiB)": 46.9,
61
+ "memory/max_allocated (GiB)": 46.9,
62
+ "ppl": 1.10858,
63
+ "step": 4,
64
+ "tokens/total": 88800,
65
+ "tokens/train_per_sec_per_gpu": 5.16,
66
+ "tokens/trainable": 1850
67
+ },
68
+ {
69
+ "epoch": 0.01953125,
70
+ "grad_norm": 25.0,
71
+ "learning_rate": 5e-06,
72
+ "loss": 0.10710409283638,
73
+ "memory/device_reserved (GiB)": 58.09,
74
+ "memory/max_active (GiB)": 46.9,
75
+ "memory/max_allocated (GiB)": 46.9,
76
+ "ppl": 1.11305,
77
+ "step": 5,
78
+ "tokens/total": 110742,
79
+ "tokens/train_per_sec_per_gpu": 3.92,
80
+ "tokens/trainable": 2302
81
+ },
82
+ {
83
+ "epoch": 0.0234375,
84
+ "grad_norm": 16.25,
85
+ "learning_rate": 5e-06,
86
+ "loss": 0.05220422148704529,
87
+ "memory/device_reserved (GiB)": 57.19,
88
+ "memory/max_active (GiB)": 46.9,
89
+ "memory/max_allocated (GiB)": 46.9,
90
+ "ppl": 1.05359,
91
+ "step": 6,
92
+ "tokens/total": 131497,
93
+ "tokens/train_per_sec_per_gpu": 4.45,
94
+ "tokens/trainable": 2742
95
+ },
96
+ {
97
+ "epoch": 0.02734375,
98
+ "grad_norm": 18.0,
99
+ "learning_rate": 5e-06,
100
+ "loss": 0.06188321113586426,
101
+ "memory/device_reserved (GiB)": 59.04,
102
+ "memory/max_active (GiB)": 46.9,
103
+ "memory/max_allocated (GiB)": 46.9,
104
+ "ppl": 1.06384,
105
+ "step": 7,
106
+ "tokens/total": 153846,
107
+ "tokens/train_per_sec_per_gpu": 4.99,
108
+ "tokens/trainable": 3208
109
+ },
110
+ {
111
+ "epoch": 0.03125,
112
+ "grad_norm": 16.75,
113
+ "learning_rate": 5e-06,
114
+ "loss": 0.036078453063964844,
115
+ "memory/device_reserved (GiB)": 60.12,
116
+ "memory/max_active (GiB)": 46.9,
117
+ "memory/max_allocated (GiB)": 46.9,
118
+ "ppl": 1.03674,
119
+ "step": 8,
120
+ "tokens/total": 175713,
121
+ "tokens/train_per_sec_per_gpu": 4.47,
122
+ "tokens/trainable": 3655
123
+ },
124
+ {
125
+ "epoch": 0.03515625,
126
+ "grad_norm": 13.5625,
127
+ "learning_rate": 5e-06,
128
+ "loss": 0.05783689022064209,
129
+ "memory/device_reserved (GiB)": 60.12,
130
+ "memory/max_active (GiB)": 46.9,
131
+ "memory/max_allocated (GiB)": 46.9,
132
+ "ppl": 1.05954,
133
+ "step": 9,
134
+ "tokens/total": 196358,
135
+ "tokens/train_per_sec_per_gpu": 3.39,
136
+ "tokens/trainable": 4091
137
+ },
138
+ {
139
+ "epoch": 0.0390625,
140
+ "grad_norm": 16.375,
141
+ "learning_rate": 5e-06,
142
+ "loss": 0.05202341079711914,
143
+ "memory/device_reserved (GiB)": 60.1,
144
+ "memory/max_active (GiB)": 46.9,
145
+ "memory/max_allocated (GiB)": 46.9,
146
+ "ppl": 1.0534,
147
+ "step": 10,
148
+ "tokens/total": 218622,
149
+ "tokens/train_per_sec_per_gpu": 4.33,
150
+ "tokens/trainable": 4553
151
+ },
152
+ {
153
+ "epoch": 0.04296875,
154
+ "grad_norm": 22.5,
155
+ "learning_rate": 5e-06,
156
+ "loss": 0.03920435905456543,
157
+ "memory/device_reserved (GiB)": 60.1,
158
+ "memory/max_active (GiB)": 46.9,
159
+ "memory/max_allocated (GiB)": 46.9,
160
+ "ppl": 1.03998,
161
+ "step": 11,
162
+ "tokens/total": 239835,
163
+ "tokens/train_per_sec_per_gpu": 4.54,
164
+ "tokens/trainable": 4992
165
+ },
166
+ {
167
+ "epoch": 0.046875,
168
+ "grad_norm": 43.75,
169
+ "learning_rate": 5e-06,
170
+ "loss": 0.0486750602722168,
171
+ "memory/device_reserved (GiB)": 60.1,
172
+ "memory/max_active (GiB)": 46.9,
173
+ "memory/max_allocated (GiB)": 46.9,
174
+ "ppl": 1.04988,
175
+ "step": 12,
176
+ "tokens/total": 264428,
177
+ "tokens/train_per_sec_per_gpu": 4.93,
178
+ "tokens/trainable": 5494
179
+ },
180
+ {
181
+ "epoch": 0.05078125,
182
+ "grad_norm": 7.59375,
183
+ "learning_rate": 5e-06,
184
+ "loss": 0.023757457733154297,
185
+ "memory/device_reserved (GiB)": 60.1,
186
+ "memory/max_active (GiB)": 46.9,
187
+ "memory/max_allocated (GiB)": 46.9,
188
+ "ppl": 1.02404,
189
+ "step": 13,
190
+ "tokens/total": 284859,
191
+ "tokens/train_per_sec_per_gpu": 3.95,
192
+ "tokens/trainable": 5933
193
+ },
194
+ {
195
+ "epoch": 0.0546875,
196
+ "grad_norm": 19.0,
197
+ "learning_rate": 5e-06,
198
+ "loss": 0.04330146312713623,
199
+ "memory/device_reserved (GiB)": 60.1,
200
+ "memory/max_active (GiB)": 46.9,
201
+ "memory/max_allocated (GiB)": 46.9,
202
+ "ppl": 1.04425,
203
+ "step": 14,
204
+ "tokens/total": 305640,
205
+ "tokens/train_per_sec_per_gpu": 3.17,
206
+ "tokens/trainable": 6369
207
+ },
208
+ {
209
+ "epoch": 0.05859375,
210
+ "grad_norm": 9.6875,
211
+ "learning_rate": 5e-06,
212
+ "loss": 0.03816962242126465,
213
+ "memory/device_reserved (GiB)": 60.1,
214
+ "memory/max_active (GiB)": 46.9,
215
+ "memory/max_allocated (GiB)": 46.9,
216
+ "ppl": 1.03891,
217
+ "step": 15,
218
+ "tokens/total": 327063,
219
+ "tokens/train_per_sec_per_gpu": 4.35,
220
+ "tokens/trainable": 6820
221
+ },
222
+ {
223
+ "epoch": 0.0625,
224
+ "grad_norm": 11.8125,
225
+ "learning_rate": 5e-06,
226
+ "loss": 0.03375864028930664,
227
+ "memory/device_reserved (GiB)": 60.1,
228
+ "memory/max_active (GiB)": 46.9,
229
+ "memory/max_allocated (GiB)": 46.9,
230
+ "ppl": 1.03433,
231
+ "step": 16,
232
+ "tokens/total": 347838,
233
+ "tokens/train_per_sec_per_gpu": 3.77,
234
+ "tokens/trainable": 7258
235
+ },
236
+ {
237
+ "epoch": 0.06640625,
238
+ "grad_norm": 12.4375,
239
+ "learning_rate": 5e-06,
240
+ "loss": 0.023131251335144043,
241
+ "memory/device_reserved (GiB)": 60.1,
242
+ "memory/max_active (GiB)": 46.9,
243
+ "memory/max_allocated (GiB)": 46.9,
244
+ "ppl": 1.0234,
245
+ "step": 17,
246
+ "tokens/total": 369888,
247
+ "tokens/train_per_sec_per_gpu": 4.4,
248
+ "tokens/trainable": 7710
249
+ },
250
+ {
251
+ "epoch": 0.0703125,
252
+ "grad_norm": 10.75,
253
+ "learning_rate": 5e-06,
254
+ "loss": 0.02522754669189453,
255
+ "memory/device_reserved (GiB)": 57.25,
256
+ "memory/max_active (GiB)": 46.9,
257
+ "memory/max_allocated (GiB)": 46.9,
258
+ "ppl": 1.02555,
259
+ "step": 18,
260
+ "tokens/total": 392455,
261
+ "tokens/train_per_sec_per_gpu": 3.72,
262
+ "tokens/trainable": 8174
263
+ },
264
+ {
265
+ "epoch": 0.07421875,
266
+ "grad_norm": 8.4375,
267
+ "learning_rate": 5e-06,
268
+ "loss": 0.028311967849731445,
269
+ "memory/device_reserved (GiB)": 58.4,
270
+ "memory/max_active (GiB)": 46.9,
271
+ "memory/max_allocated (GiB)": 46.9,
272
+ "ppl": 1.02872,
273
+ "step": 19,
274
+ "tokens/total": 413549,
275
+ "tokens/train_per_sec_per_gpu": 4.51,
276
+ "tokens/trainable": 8617
277
+ },
278
+ {
279
+ "epoch": 0.078125,
280
+ "grad_norm": 4.34375,
281
+ "learning_rate": 5e-06,
282
+ "loss": 0.012633323669433594,
283
+ "memory/device_reserved (GiB)": 58.4,
284
+ "memory/max_active (GiB)": 46.9,
285
+ "memory/max_allocated (GiB)": 46.9,
286
+ "ppl": 1.01271,
287
+ "step": 20,
288
+ "tokens/total": 434128,
289
+ "tokens/train_per_sec_per_gpu": 3.7,
290
+ "tokens/trainable": 9037
291
+ },
292
+ {
293
+ "epoch": 0.08203125,
294
+ "grad_norm": 3.34375,
295
+ "learning_rate": 5e-06,
296
+ "loss": 0.010331928730010986,
297
+ "memory/device_reserved (GiB)": 58.4,
298
+ "memory/max_active (GiB)": 46.9,
299
+ "memory/max_allocated (GiB)": 46.9,
300
+ "ppl": 1.01039,
301
+ "step": 21,
302
+ "tokens/total": 456131,
303
+ "tokens/train_per_sec_per_gpu": 5.15,
304
+ "tokens/trainable": 9503
305
+ },
306
+ {
307
+ "epoch": 0.0859375,
308
+ "grad_norm": 8.6875,
309
+ "learning_rate": 5e-06,
310
+ "loss": 0.011040687561035156,
311
+ "memory/device_reserved (GiB)": 58.4,
312
+ "memory/max_active (GiB)": 46.9,
313
+ "memory/max_allocated (GiB)": 46.9,
314
+ "ppl": 1.0111,
315
+ "step": 22,
316
+ "tokens/total": 478707,
317
+ "tokens/train_per_sec_per_gpu": 4.45,
318
+ "tokens/trainable": 9972
319
+ },
320
+ {
321
+ "epoch": 0.08984375,
322
+ "grad_norm": 17.75,
323
+ "learning_rate": 5e-06,
324
+ "loss": 0.030008912086486816,
325
+ "memory/device_reserved (GiB)": 58.4,
326
+ "memory/max_active (GiB)": 46.9,
327
+ "memory/max_allocated (GiB)": 46.9,
328
+ "ppl": 1.03046,
329
+ "step": 23,
330
+ "tokens/total": 501425,
331
+ "tokens/train_per_sec_per_gpu": 4.9,
332
+ "tokens/trainable": 10447
333
+ },
334
+ {
335
+ "epoch": 0.09375,
336
+ "grad_norm": 6.6875,
337
+ "learning_rate": 5e-06,
338
+ "loss": 0.007420599460601807,
339
+ "memory/device_reserved (GiB)": 60.14,
340
+ "memory/max_active (GiB)": 46.9,
341
+ "memory/max_allocated (GiB)": 46.9,
342
+ "ppl": 1.00745,
343
+ "step": 24,
344
+ "tokens/total": 522885,
345
+ "tokens/train_per_sec_per_gpu": 5.15,
346
+ "tokens/trainable": 10899
347
+ },
348
+ {
349
+ "epoch": 0.09765625,
350
+ "grad_norm": 24.75,
351
+ "learning_rate": 5e-06,
352
+ "loss": 0.05265447497367859,
353
+ "memory/device_reserved (GiB)": 60.14,
354
+ "memory/max_active (GiB)": 46.9,
355
+ "memory/max_allocated (GiB)": 46.9,
356
+ "ppl": 1.05407,
357
+ "step": 25,
358
+ "tokens/total": 544886,
359
+ "tokens/train_per_sec_per_gpu": 4.45,
360
+ "tokens/trainable": 11360
361
+ },
362
+ {
363
+ "epoch": 0.1015625,
364
+ "grad_norm": 12.625,
365
+ "learning_rate": 5e-06,
366
+ "loss": 0.04731038212776184,
367
+ "memory/device_reserved (GiB)": 60.14,
368
+ "memory/max_active (GiB)": 46.9,
369
+ "memory/max_allocated (GiB)": 46.9,
370
+ "ppl": 1.04845,
371
+ "step": 26,
372
+ "tokens/total": 564964,
373
+ "tokens/train_per_sec_per_gpu": 3.36,
374
+ "tokens/trainable": 11775
375
+ },
376
+ {
377
+ "epoch": 0.10546875,
378
+ "grad_norm": 13.1875,
379
+ "learning_rate": 5e-06,
380
+ "loss": 0.014921486377716064,
381
+ "memory/device_reserved (GiB)": 60.14,
382
+ "memory/max_active (GiB)": 46.9,
383
+ "memory/max_allocated (GiB)": 46.9,
384
+ "ppl": 1.01503,
385
+ "step": 27,
386
+ "tokens/total": 586567,
387
+ "tokens/train_per_sec_per_gpu": 5.29,
388
+ "tokens/trainable": 12242
389
+ },
390
+ {
391
+ "epoch": 0.109375,
392
+ "grad_norm": 18.625,
393
+ "learning_rate": 5e-06,
394
+ "loss": 0.013006031513214111,
395
+ "memory/device_reserved (GiB)": 60.14,
396
+ "memory/max_active (GiB)": 46.9,
397
+ "memory/max_allocated (GiB)": 46.9,
398
+ "ppl": 1.01309,
399
+ "step": 28,
400
+ "tokens/total": 608188,
401
+ "tokens/train_per_sec_per_gpu": 3.78,
402
+ "tokens/trainable": 12696
403
+ },
404
+ {
405
+ "epoch": 0.11328125,
406
+ "grad_norm": 21.25,
407
+ "learning_rate": 5e-06,
408
+ "loss": 0.028101369738578796,
409
+ "memory/device_reserved (GiB)": 60.14,
410
+ "memory/max_active (GiB)": 46.9,
411
+ "memory/max_allocated (GiB)": 46.9,
412
+ "ppl": 1.0285,
413
+ "step": 29,
414
+ "tokens/total": 630007,
415
+ "tokens/train_per_sec_per_gpu": 3.81,
416
+ "tokens/trainable": 13153
417
+ },
418
+ {
419
+ "epoch": 0.1171875,
420
+ "grad_norm": 8.3125,
421
+ "learning_rate": 5e-06,
422
+ "loss": 0.01690027117729187,
423
+ "memory/device_reserved (GiB)": 60.14,
424
+ "memory/max_active (GiB)": 46.9,
425
+ "memory/max_allocated (GiB)": 46.9,
426
+ "ppl": 1.01704,
427
+ "step": 30,
428
+ "tokens/total": 651656,
429
+ "tokens/train_per_sec_per_gpu": 5.16,
430
+ "tokens/trainable": 13624
431
+ },
432
+ {
433
+ "epoch": 0.12109375,
434
+ "grad_norm": 12.3125,
435
+ "learning_rate": 5e-06,
436
+ "loss": 0.015246853232383728,
437
+ "memory/device_reserved (GiB)": 60.14,
438
+ "memory/max_active (GiB)": 46.9,
439
+ "memory/max_allocated (GiB)": 46.9,
440
+ "ppl": 1.01536,
441
+ "step": 31,
442
+ "tokens/total": 672938,
443
+ "tokens/train_per_sec_per_gpu": 5.14,
444
+ "tokens/trainable": 14064
445
+ },
446
+ {
447
+ "epoch": 0.125,
448
+ "grad_norm": 11.125,
449
+ "learning_rate": 5e-06,
450
+ "loss": 0.017684906721115112,
451
+ "memory/device_reserved (GiB)": 60.14,
452
+ "memory/max_active (GiB)": 46.9,
453
+ "memory/max_allocated (GiB)": 46.9,
454
+ "ppl": 1.01784,
455
+ "step": 32,
456
+ "tokens/total": 696525,
457
+ "tokens/train_per_sec_per_gpu": 5.03,
458
+ "tokens/trainable": 14565
459
+ }
460
+ ],
461
+ "logging_steps": 1,
462
+ "max_steps": 512,
463
+ "num_input_tokens_seen": 0,
464
+ "num_train_epochs": 2,
465
+ "save_steps": 500,
466
+ "stateful_callbacks": {
467
+ "TrainerControl": {
468
+ "args": {
469
+ "should_epoch_stop": false,
470
+ "should_evaluate": false,
471
+ "should_log": false,
472
+ "should_save": true,
473
+ "should_training_stop": false
474
+ },
475
+ "attributes": {}
476
+ }
477
+ },
478
+ "total_flos": 4.670504244242022e+16,
479
+ "train_batch_size": 1,
480
+ "trial_name": null,
481
+ "trial_params": null
482
+ }
full_aft_midtrain4/balanced/checkpoint-32/training_args.bin ADDED
@@ -0,0 +1,3 @@
 
 
 
 
1
+ version https://git-lfs.github.com/spec/v1
2
+ oid sha256:d0107cd88280e968b63f91f7125176eba0c9002c1f534529b1c400b807d0860f
3
+ size 8913
full_aft_midtrain4/balanced/checkpoint-4/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 -%}
full_aft_midtrain4/balanced/checkpoint-4/config.json ADDED
@@ -0,0 +1,125 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
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
+ "bos_token_id": 2,
21
+ "cache_implementation": "hybrid",
22
+ "dtype": "bfloat16",
23
+ "eos_token_id": 1,
24
+ "final_logit_softcapping": null,
25
+ "head_dim": 256,
26
+ "hidden_activation": "gelu_pytorch_tanh",
27
+ "hidden_size": 3840,
28
+ "initializer_range": 0.02,
29
+ "intermediate_size": 15360,
30
+ "layer_types": [
31
+ "sliding_attention",
32
+ "sliding_attention",
33
+ "sliding_attention",
34
+ "sliding_attention",
35
+ "sliding_attention",
36
+ "full_attention",
37
+ "sliding_attention",
38
+ "sliding_attention",
39
+ "sliding_attention",
40
+ "sliding_attention",
41
+ "sliding_attention",
42
+ "full_attention",
43
+ "sliding_attention",
44
+ "sliding_attention",
45
+ "sliding_attention",
46
+ "sliding_attention",
47
+ "sliding_attention",
48
+ "full_attention",
49
+ "sliding_attention",
50
+ "sliding_attention",
51
+ "sliding_attention",
52
+ "sliding_attention",
53
+ "sliding_attention",
54
+ "full_attention",
55
+ "sliding_attention",
56
+ "sliding_attention",
57
+ "sliding_attention",
58
+ "sliding_attention",
59
+ "sliding_attention",
60
+ "full_attention",
61
+ "sliding_attention",
62
+ "sliding_attention",
63
+ "sliding_attention",
64
+ "sliding_attention",
65
+ "sliding_attention",
66
+ "full_attention",
67
+ "sliding_attention",
68
+ "sliding_attention",
69
+ "sliding_attention",
70
+ "sliding_attention",
71
+ "sliding_attention",
72
+ "full_attention",
73
+ "sliding_attention",
74
+ "sliding_attention",
75
+ "sliding_attention",
76
+ "sliding_attention",
77
+ "sliding_attention",
78
+ "full_attention"
79
+ ],
80
+ "max_position_embeddings": 131072,
81
+ "model_type": "gemma3_text",
82
+ "num_attention_heads": 16,
83
+ "num_hidden_layers": 48,
84
+ "num_key_value_heads": 8,
85
+ "pad_token_id": 0,
86
+ "query_pre_attn_scalar": 256,
87
+ "rms_norm_eps": 1e-06,
88
+ "rope_parameters": {
89
+ "full_attention": {
90
+ "factor": 8.0,
91
+ "rope_theta": 1000000.0,
92
+ "rope_type": "linear"
93
+ },
94
+ "sliding_attention": {
95
+ "rope_theta": 10000.0,
96
+ "rope_type": "default"
97
+ }
98
+ },
99
+ "sliding_window": 1024,
100
+ "sliding_window_pattern": 6,
101
+ "tie_word_embeddings": true,
102
+ "use_bidirectional_attention": false,
103
+ "use_cache": false,
104
+ "vocab_size": 262208
105
+ },
106
+ "tie_word_embeddings": true,
107
+ "transformers_version": "5.9.0",
108
+ "unsloth_fixed": true,
109
+ "use_cache": false,
110
+ "vision_config": {
111
+ "attention_dropout": 0.0,
112
+ "dtype": "bfloat16",
113
+ "hidden_act": "gelu_pytorch_tanh",
114
+ "hidden_size": 1152,
115
+ "image_size": 896,
116
+ "intermediate_size": 4304,
117
+ "layer_norm_eps": 1e-06,
118
+ "model_type": "siglip_vision_model",
119
+ "num_attention_heads": 16,
120
+ "num_channels": 3,
121
+ "num_hidden_layers": 27,
122
+ "patch_size": 14,
123
+ "vision_use_head": false
124
+ }
125
+ }
full_aft_midtrain4/balanced/checkpoint-4/generation_config.json ADDED
@@ -0,0 +1,13 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ {
2
+ "bos_token_id": 2,
3
+ "cache_implementation": "hybrid",
4
+ "do_sample": true,
5
+ "eos_token_id": [
6
+ 1,
7
+ 106
8
+ ],
9
+ "pad_token_id": 0,
10
+ "top_k": 64,
11
+ "top_p": 0.95,
12
+ "transformers_version": "5.9.0"
13
+ }
full_aft_midtrain4/balanced/checkpoint-4/model.safetensors ADDED
@@ -0,0 +1,3 @@
 
 
 
 
1
+ version https://git-lfs.github.com/spec/v1
2
+ oid sha256:7d4a0f1216763f701d3156eab5d12dd3f2ee3a75d28d01e2e80373cca8e39625
3
+ size 26388552360
full_aft_midtrain4/balanced/checkpoint-4/preprocessor_config.json ADDED
@@ -0,0 +1,29 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ {
2
+ "do_convert_rgb": null,
3
+ "do_normalize": true,
4
+ "do_pan_and_scan": null,
5
+ "do_rescale": true,
6
+ "do_resize": true,
7
+ "image_mean": [
8
+ 0.5,
9
+ 0.5,
10
+ 0.5
11
+ ],
12
+ "image_processor_type": "Gemma3ImageProcessor",
13
+ "image_seq_length": 256,
14
+ "image_std": [
15
+ 0.5,
16
+ 0.5,
17
+ 0.5
18
+ ],
19
+ "pan_and_scan_max_num_crops": null,
20
+ "pan_and_scan_min_crop_size": null,
21
+ "pan_and_scan_min_ratio_to_activate": null,
22
+ "processor_class": "Gemma3Processor",
23
+ "resample": 2,
24
+ "rescale_factor": 0.00392156862745098,
25
+ "size": {
26
+ "height": 896,
27
+ "width": 896
28
+ }
29
+ }