jbduran commited on
Commit
6951e66
·
verified ·
1 Parent(s): b3237c4

Import experiment archive from bart (batch 2)

Browse files
This view is limited to 50 files because it contains too many changes.   See raw diff
Files changed (50) hide show
  1. experiments/Think.Unbounded-d32-v2mix-cont/sft/Think.Unbounded-d32-v2mix-cont-nanochat-default-datamatch-v1/summary.json +25 -0
  2. experiments/Think.Unbounded-d32-v2mix-cont/sft/Think.Unbounded-d32-v2mix-cont-pre1930-curriculum-c3-robust-v2/checkpoints/eval_metrics.json +31 -0
  3. experiments/Think.Unbounded-d32-v2mix-cont/sft/Think.Unbounded-d32-v2mix-cont-pre1930-curriculum-c3-robust-v2/checkpoints/meta_000042.json +178 -0
  4. experiments/Think.Unbounded-d32-v2mix-cont/sft/Think.Unbounded-d32-v2mix-cont-pre1930-curriculum-c3-robust-v2/checkpoints/model_000042.pt +3 -0
  5. experiments/Think.Unbounded-d32-v2mix-cont/sft/Think.Unbounded-d32-v2mix-cont-pre1930-curriculum-c3-robust-v2/checkpoints/optim_000042_rank0.pt +3 -0
  6. experiments/Think.Unbounded-d32-v2mix-cont/sft/Think.Unbounded-d32-v2mix-cont-pre1930-curriculum-c3-robust-v2/config.json +95 -0
  7. experiments/Think.Unbounded-d32-v2mix-cont/sft/Think.Unbounded-d32-v2mix-cont-pre1930-curriculum-c3-robust-v2/evals/chatcore.json +29 -0
  8. experiments/Think.Unbounded-d32-v2mix-cont/sft/Think.Unbounded-d32-v2mix-cont-pre1930-curriculum-c3-robust-v2/run.json +11 -0
  9. experiments/Think.Unbounded-d32-v2mix-cont/sft/Think.Unbounded-d32-v2mix-cont-pre1930-curriculum-c3-robust-v3-a100-40gb/config.json +97 -0
  10. experiments/Think.Unbounded-d32-v2mix-cont/sft/Think.Unbounded-d32-v2mix-cont-pre1930-curriculum-c3-robust-v3-a100-40gb/run.json +11 -0
  11. experiments/Think.Unbounded-d32-v2mix-cont/sft/Think.Unbounded-d32-v2mix-cont-pre1930-curriculum-c3-robust-v3/config.json +97 -0
  12. experiments/Think.Unbounded-d32-v2mix-cont/sft/Think.Unbounded-d32-v2mix-cont-pre1930-curriculum-c3-robust-v3/run.json +11 -0
  13. experiments/Think.Unbounded-d32-v2mix-cont/sft/Think.Unbounded-d32-v2mix-cont-pre1930-curriculum-c3-robust/checkpoints/meta_000041.json +172 -0
  14. experiments/Think.Unbounded-d32-v2mix-cont/sft/Think.Unbounded-d32-v2mix-cont-pre1930-curriculum-c3-robust/checkpoints/model_000041.pt +3 -0
  15. experiments/Think.Unbounded-d32-v2mix-cont/sft/Think.Unbounded-d32-v2mix-cont-pre1930-curriculum-c3-robust/checkpoints/optim_000041_rank0.pt +3 -0
  16. experiments/Think.Unbounded-d32-v2mix-cont/sft/Think.Unbounded-d32-v2mix-cont-pre1930-curriculum-c3-robust/config.json +89 -0
  17. experiments/Think.Unbounded-d32-v2mix-cont/sft/Think.Unbounded-d32-v2mix-cont-pre1930-curriculum-c3-robust/run.json +11 -0
  18. experiments/Think.Unbounded-d32-v2mix-cont/summary.json +79 -0
  19. experiments/Think.Unbounded-d32-v2mix-cont/tokenizer/experiment_tokenizer.json +18 -0
  20. experiments/Think.Unbounded-d32-v2mix-cont/tokenizer/token_bytes.pt +3 -0
  21. experiments/Think.Unbounded-d32-v2mix-cont/tokenizer/tokenizer.pkl +3 -0
  22. experiments/Think.Unbounded-d32/base_checkpoints/meta_000500.json +236 -0
  23. experiments/Think.Unbounded-d32/base_checkpoints/meta_001000.json +236 -0
  24. experiments/Think.Unbounded-d32/base_checkpoints/meta_001500.json +236 -0
  25. experiments/Think.Unbounded-d32/base_checkpoints/meta_002000.json +236 -0
  26. experiments/Think.Unbounded-d32/base_checkpoints/meta_002500.json +236 -0
  27. experiments/Think.Unbounded-d32/base_checkpoints/meta_003000.json +236 -0
  28. experiments/Think.Unbounded-d32/base_checkpoints/meta_003500.json +236 -0
  29. experiments/Think.Unbounded-d32/base_checkpoints/meta_004000.json +236 -0
  30. experiments/Think.Unbounded-d32/base_checkpoints/meta_004500.json +236 -0
  31. experiments/Think.Unbounded-d32/base_checkpoints/meta_005000.json +236 -0
  32. experiments/Think.Unbounded-d32/base_checkpoints/meta_005500.json +236 -0
  33. experiments/Think.Unbounded-d32/base_checkpoints/model_000500.pt +3 -0
  34. experiments/Think.Unbounded-d32/base_checkpoints/model_001000.pt +3 -0
  35. experiments/Think.Unbounded-d32/base_checkpoints/model_001500.pt +3 -0
  36. experiments/Think.Unbounded-d32/base_checkpoints/model_002000.pt +3 -0
  37. experiments/Think.Unbounded-d32/base_checkpoints/model_002500.pt +3 -0
  38. experiments/Think.Unbounded-d32/base_checkpoints/model_003000.pt +3 -0
  39. experiments/Think.Unbounded-d32/base_checkpoints/model_003500.pt +3 -0
  40. experiments/Think.Unbounded-d32/base_checkpoints/model_004000.pt +3 -0
  41. experiments/Think.Unbounded-d32/base_checkpoints/model_004500.pt +3 -0
  42. experiments/Think.Unbounded-d32/base_checkpoints/model_005000.pt +3 -0
  43. experiments/Think.Unbounded-d32/base_checkpoints/model_005500.pt +3 -0
  44. experiments/Think.Unbounded-d32/base_checkpoints/optim_000500_rank0.pt +3 -0
  45. experiments/Think.Unbounded-d32/base_checkpoints/optim_001000_rank0.pt +3 -0
  46. experiments/Think.Unbounded-d32/base_checkpoints/optim_001500_rank0.pt +3 -0
  47. experiments/Think.Unbounded-d32/base_checkpoints/optim_002000_rank0.pt +3 -0
  48. experiments/Think.Unbounded-d32/base_checkpoints/optim_002500_rank0.pt +3 -0
  49. experiments/Think.Unbounded-d32/base_checkpoints/optim_003000_rank0.pt +3 -0
  50. experiments/Think.Unbounded-d32/base_checkpoints/optim_003500_rank0.pt +3 -0
experiments/Think.Unbounded-d32-v2mix-cont/sft/Think.Unbounded-d32-v2mix-cont-nanochat-default-datamatch-v1/summary.json ADDED
@@ -0,0 +1,25 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ {
2
+ "experiment_id": "Think.Unbounded-d32-v2mix-cont-nanochat-default-datamatch-v1",
3
+ "stage": "sft",
4
+ "base_experiment_id": "Think.Unbounded-d32-v2mix-cont",
5
+ "parent_experiment_id": "Think.Unbounded-d32-v2mix-cont",
6
+ "parent_checkpoint_step": 9600,
7
+ "step": 160,
8
+ "stage_training_flops": 3.8505838661546803e+18,
9
+ "inherited_parent_flops": 2.3103503196928082e+20,
10
+ "cumulative_pipeline_training_flops": 2.348856158354355e+20,
11
+ "chatcore_metric": 0.13532147122389065,
12
+ "chat_results": {
13
+ "ARC-Easy": 0.4553872053872054,
14
+ "ARC-Challenge": 0.378839590443686,
15
+ "MMLU": 0.3474576271186441,
16
+ "GSM8K": 0.06444275966641395,
17
+ "HumanEval": 0.036585365853658534
18
+ },
19
+ "training_time_seconds": 6951.126858472824,
20
+ "config_fingerprint": "6359be7150ece075",
21
+ "tokenizer_fingerprint": "21e99d5cdeeaa660",
22
+ "git_commit_sha": "3c34775d260a016a79ab46ea76f8c8e48458fd68",
23
+ "wandb_url": "https://wandb.ai/jbduran-thinkingmachinesncsu/think.nano/runs/8c553682",
24
+ "huggingface_url": "https://huggingface.co/jbduran/think.nano/tree/main/experiments/Think.Unbounded-d32-v2mix-cont/sft/Think.Unbounded-d32-v2mix-cont-nanochat-default-datamatch-v1"
25
+ }
experiments/Think.Unbounded-d32-v2mix-cont/sft/Think.Unbounded-d32-v2mix-cont-pre1930-curriculum-c3-robust-v2/checkpoints/eval_metrics.json ADDED
@@ -0,0 +1,31 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ {
2
+ "experiment_id": "Think.Unbounded-d32-v2mix-cont-pre1930-curriculum-c3-robust-v2",
3
+ "recipe": "curriculum",
4
+ "step": 42,
5
+ "val_bpb": 0.7037352212342954,
6
+ "min_val_bpb": 0.7037352212342954,
7
+ "per_route_bpb": {},
8
+ "per_domain_bpb": {},
9
+ "curriculum_summary": null,
10
+ "chatcore": {
11
+ "chatcore_metric": -0.011321608241655831,
12
+ "chatcore_cat": -0.018869347069426386,
13
+ "suite": {
14
+ "name": "karpathy",
15
+ "tasks": [
16
+ "ARC-Easy",
17
+ "ARC-Challenge",
18
+ "MMLU",
19
+ "GSM8K",
20
+ "HumanEval"
21
+ ],
22
+ "max_generative_problems": null,
23
+ "generative_answer_format": null
24
+ },
25
+ "ARC-Easy": 0.24579124579124578,
26
+ "ARC-Challenge": 0.23037542662116042,
27
+ "MMLU": 0.23137729668138443,
28
+ "GSM8K": 0.0,
29
+ "HumanEval": 0.0
30
+ }
31
+ }
experiments/Think.Unbounded-d32-v2mix-cont/sft/Think.Unbounded-d32-v2mix-cont-pre1930-curriculum-c3-robust-v2/checkpoints/meta_000042.json ADDED
@@ -0,0 +1,178 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ {
2
+ "step": 42,
3
+ "training_complete": true,
4
+ "val_bpb": 0.7037352212342954,
5
+ "model_config": {
6
+ "sequence_len": 4096,
7
+ "vocab_size": 32768,
8
+ "n_layer": 32,
9
+ "n_head": 16,
10
+ "n_kv_head": 16,
11
+ "n_embd": 2048,
12
+ "window_pattern": "SSSL"
13
+ },
14
+ "user_config": {
15
+ "run": "Think.Unbounded-d32-v2mix-cont-pre1930-curriculum-c3-robust-v2",
16
+ "wandb_run_id": "036b1c71",
17
+ "wandb_group": "think-d32",
18
+ "wandb_tags": "sft,curriculum,c3,scale-max,staged,robustness,noise,d32",
19
+ "device_type": "",
20
+ "model_tag": null,
21
+ "model_step": null,
22
+ "base_checkpoint_dir": "/workspace/nanochat/experiments/Think.Unbounded-d32-v2mix-cont/base_checkpoints",
23
+ "base_step": 9600,
24
+ "checkpoint_dir": "/workspace/nanochat/experiments/Think.Unbounded-d32-v2mix-cont/sft/Think.Unbounded-d32-v2mix-cont-pre1930-curriculum-c3-robust-v2/checkpoints",
25
+ "tokenizer_dir": "/workspace/nanochat/experiments/Think.Unbounded-d32-v2mix-cont/tokenizer",
26
+ "resume_from_step": null,
27
+ "experiment_id": "Think.Unbounded-d32-v2mix-cont-pre1930-curriculum-c3-robust-v2",
28
+ "experiment_config": "/workspace/nanochat/experiments/Think.Unbounded-d32-v2mix-cont/sft/Think.Unbounded-d32-v2mix-cont-pre1930-curriculum-c3-robust-v2/config.json",
29
+ "parent_cumulative_flops": 2.3103503196928082e+20,
30
+ "tokenizer_fingerprint": "21e99d5cdeeaa660",
31
+ "git_commit_sha": "04bb043ad61d4db3e9022c422302df7b8f2dc0c9",
32
+ "load_optimizer": 0,
33
+ "num_iterations": -1,
34
+ "max_seq_len": null,
35
+ "device_batch_size": 2,
36
+ "total_batch_size": null,
37
+ "embedding_lr": null,
38
+ "unembedding_lr": null,
39
+ "matrix_lr": null,
40
+ "init_lr_frac": 0.8,
41
+ "warmup_ratio": 0.03,
42
+ "warmdown_ratio": 0.5,
43
+ "final_lr_frac": 0.0,
44
+ "eval_every": 400,
45
+ "eval_tokens": 20971520,
46
+ "chatcore_every": 100000,
47
+ "chatcore_max_cat": -1,
48
+ "chatcore_max_sample": 32,
49
+ "save_every": -1,
50
+ "recipe": "curriculum",
51
+ "curriculum_config": "",
52
+ "pre1930_epochs": 5,
53
+ "mmlu_epochs": 3,
54
+ "gsm8k_epochs": 4,
55
+ "authentic_epochs": 0,
56
+ "knowledge_qa_epochs": 0,
57
+ "multiturn_qa_epochs": 0,
58
+ "reasoning_qa_epochs": 0,
59
+ "stem_reasoning_epochs": 0,
60
+ "narrative_grounded_epochs": 0,
61
+ "narrative_fiction_epochs": 0,
62
+ "opinion_qa_epochs": 0,
63
+ "how_to_qa_epochs": 0,
64
+ "verse_qa_epochs": 0,
65
+ "composition_qa_epochs": 0,
66
+ "resolved_experiment_config": {
67
+ "schema_version": 1,
68
+ "stage": "sft",
69
+ "experiment_suffix": "pre1930-curriculum-c3-robust-v2",
70
+ "data": {
71
+ "recipe": "curriculum",
72
+ "curriculum": {
73
+ "name": "C3Rv2",
74
+ "mode": "staged",
75
+ "threshold_default": 80,
76
+ "stages": [
77
+ {
78
+ "routes": [
79
+ "knowledge_qa"
80
+ ],
81
+ "authentic": "single"
82
+ },
83
+ {
84
+ "routes": [
85
+ "reasoning_qa",
86
+ "stem_reasoning",
87
+ "how_to_qa",
88
+ "opinion_qa",
89
+ "composition_qa",
90
+ "verse_qa"
91
+ ],
92
+ "calibration_qa": true
93
+ },
94
+ {
95
+ "routes": [
96
+ "multiturn_qa",
97
+ "narrative_grounded",
98
+ "narrative_fiction"
99
+ ],
100
+ "authentic": "multi"
101
+ }
102
+ ],
103
+ "noise": {
104
+ "rate": 0.3
105
+ },
106
+ "robustness": {
107
+ "stage": 0,
108
+ "epochs": 1,
109
+ "routes": {
110
+ "conversation_qa": {
111
+ "count": null
112
+ },
113
+ "unparseable_qa": {
114
+ "count": null
115
+ },
116
+ "typo_qa": {
117
+ "count": null
118
+ },
119
+ "era_qa": {
120
+ "count": null
121
+ },
122
+ "conversation_multiturn": {
123
+ "count": null
124
+ }
125
+ }
126
+ }
127
+ }
128
+ },
129
+ "training": {
130
+ "num_iterations": -1,
131
+ "load_optimizer": 0,
132
+ "device_batch_size": 2,
133
+ "warmup_ratio": 0.03,
134
+ "eval_every": 400,
135
+ "chatcore_every": 100000,
136
+ "chatcore_max_sample": 32,
137
+ "save_every": -1
138
+ },
139
+ "artifacts": {
140
+ "repo": "jbduran/think.nano"
141
+ },
142
+ "wandb": {
143
+ "enabled": true,
144
+ "entity": "jbduran-thinkingmachinesncsu",
145
+ "project": "think.nano",
146
+ "group": "think-d32",
147
+ "tags": [
148
+ "sft",
149
+ "curriculum",
150
+ "c3",
151
+ "scale-max",
152
+ "staged",
153
+ "robustness",
154
+ "noise",
155
+ "d32"
156
+ ]
157
+ },
158
+ "config_fingerprint": "4462a389dc98cd5d",
159
+ "artifact_path": "experiments/Think.Unbounded-d32-v2mix-cont/sft/Think.Unbounded-d32-v2mix-cont-pre1930-curriculum-c3-robust-v2"
160
+ },
161
+ "stage": "sft",
162
+ "base_experiment_id": null,
163
+ "parent_experiment_id": null,
164
+ "parent_checkpoint_step": null,
165
+ "config_fingerprint": "4462a389dc98cd5d"
166
+ },
167
+ "loop_state": {
168
+ "step": 42,
169
+ "total_training_time": 1503.5729036331177,
170
+ "min_val_bpb": 0.7037352212342954,
171
+ "smooth_train_loss": 1.8593226867746961,
172
+ "mfu": 51.82901425419851,
173
+ "tok_per_sec": 44667,
174
+ "stage_training_flops": 1.0107782648656036e+18,
175
+ "inherited_parent_flops": 2.3103503196928082e+20,
176
+ "cumulative_pipeline_training_flops": 2.3204581023414642e+20
177
+ }
178
+ }
experiments/Think.Unbounded-d32-v2mix-cont/sft/Think.Unbounded-d32-v2mix-cont-pre1930-curriculum-c3-robust-v2/checkpoints/model_000042.pt ADDED
@@ -0,0 +1,3 @@
 
 
 
 
1
+ version https://git-lfs.github.com/spec/v1
2
+ oid sha256:f6a9bc83e3853f0a7ad8b27cc45388d727cc5976837f514d61ea59f80d7f0348
3
+ size 8992694653
experiments/Think.Unbounded-d32-v2mix-cont/sft/Think.Unbounded-d32-v2mix-cont-pre1930-curriculum-c3-robust-v2/checkpoints/optim_000042_rank0.pt ADDED
@@ -0,0 +1,3 @@
 
 
 
 
1
+ version https://git-lfs.github.com/spec/v1
2
+ oid sha256:34c0871ec1aed5ced55a463ea1b827e4ef80e286735cc42512ced11f8c579fe4
3
+ size 11545903753
experiments/Think.Unbounded-d32-v2mix-cont/sft/Think.Unbounded-d32-v2mix-cont-pre1930-curriculum-c3-robust-v2/config.json ADDED
@@ -0,0 +1,95 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ {
2
+ "schema_version": 1,
3
+ "stage": "sft",
4
+ "experiment_suffix": "pre1930-curriculum-c3-robust-v2",
5
+ "data": {
6
+ "recipe": "curriculum",
7
+ "curriculum": {
8
+ "name": "C3Rv2",
9
+ "mode": "staged",
10
+ "threshold_default": 80,
11
+ "stages": [
12
+ {
13
+ "routes": [
14
+ "knowledge_qa"
15
+ ],
16
+ "authentic": "single"
17
+ },
18
+ {
19
+ "routes": [
20
+ "reasoning_qa",
21
+ "stem_reasoning",
22
+ "how_to_qa",
23
+ "opinion_qa",
24
+ "composition_qa",
25
+ "verse_qa"
26
+ ],
27
+ "calibration_qa": true
28
+ },
29
+ {
30
+ "routes": [
31
+ "multiturn_qa",
32
+ "narrative_grounded",
33
+ "narrative_fiction"
34
+ ],
35
+ "authentic": "multi"
36
+ }
37
+ ],
38
+ "noise": {
39
+ "rate": 0.3
40
+ },
41
+ "robustness": {
42
+ "stage": 0,
43
+ "epochs": 1,
44
+ "routes": {
45
+ "conversation_qa": {
46
+ "count": null
47
+ },
48
+ "unparseable_qa": {
49
+ "count": null
50
+ },
51
+ "typo_qa": {
52
+ "count": null
53
+ },
54
+ "era_qa": {
55
+ "count": null
56
+ },
57
+ "conversation_multiturn": {
58
+ "count": null
59
+ }
60
+ }
61
+ }
62
+ }
63
+ },
64
+ "training": {
65
+ "num_iterations": -1,
66
+ "load_optimizer": 0,
67
+ "device_batch_size": 2,
68
+ "warmup_ratio": 0.03,
69
+ "eval_every": 400,
70
+ "chatcore_every": 100000,
71
+ "chatcore_max_sample": 32,
72
+ "save_every": -1
73
+ },
74
+ "artifacts": {
75
+ "repo": "jbduran/think.nano"
76
+ },
77
+ "wandb": {
78
+ "enabled": true,
79
+ "entity": "jbduran-thinkingmachinesncsu",
80
+ "project": "think.nano",
81
+ "group": "think-d32",
82
+ "tags": [
83
+ "sft",
84
+ "curriculum",
85
+ "c3",
86
+ "scale-max",
87
+ "staged",
88
+ "robustness",
89
+ "noise",
90
+ "d32"
91
+ ]
92
+ },
93
+ "config_fingerprint": "4462a389dc98cd5d",
94
+ "artifact_path": "experiments/Think.Unbounded-d32-v2mix-cont/sft/Think.Unbounded-d32-v2mix-cont-pre1930-curriculum-c3-robust-v2"
95
+ }
experiments/Think.Unbounded-d32-v2mix-cont/sft/Think.Unbounded-d32-v2mix-cont-pre1930-curriculum-c3-robust-v2/evals/chatcore.json ADDED
@@ -0,0 +1,29 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ {
2
+ "stage": "sft",
3
+ "step": 42,
4
+ "total_training_flops": 2.3204581023414642e+20,
5
+ "stage_training_flops": 1.0107782648656036e+18,
6
+ "inherited_parent_flops": 2.3103503196928082e+20,
7
+ "cumulative_pipeline_training_flops": 2.3204581023414642e+20,
8
+ "results": {
9
+ "ARC-Easy": 0.24579124579124578,
10
+ "ARC-Challenge": 0.23037542662116042,
11
+ "MMLU": 0.23137729668138443,
12
+ "GSM8K": 0.0,
13
+ "HumanEval": 0.0
14
+ },
15
+ "chatcore_metric": -0.011321608241655831,
16
+ "chatcore_suite": {
17
+ "name": "karpathy",
18
+ "tasks": [
19
+ "ARC-Easy",
20
+ "ARC-Challenge",
21
+ "MMLU",
22
+ "GSM8K",
23
+ "HumanEval"
24
+ ],
25
+ "max_generative_problems": null,
26
+ "generative_answer_format": null
27
+ },
28
+ "complete": true
29
+ }
experiments/Think.Unbounded-d32-v2mix-cont/sft/Think.Unbounded-d32-v2mix-cont-pre1930-curriculum-c3-robust-v2/run.json ADDED
@@ -0,0 +1,11 @@
 
 
 
 
 
 
 
 
 
 
 
 
1
+ {
2
+ "experiment_id": "Think.Unbounded-d32-v2mix-cont-pre1930-curriculum-c3-robust-v2",
3
+ "stage": "sft",
4
+ "base_experiment_id": "Think.Unbounded-d32-v2mix-cont",
5
+ "parent_experiment_id": "Think.Unbounded-d32-v2mix-cont",
6
+ "parent_checkpoint_step": 9600,
7
+ "branch_parent_step": null,
8
+ "config_fingerprint": "4462a389dc98cd5d",
9
+ "wandb_run_id": "036b1c71",
10
+ "created_at": 1786715992
11
+ }
experiments/Think.Unbounded-d32-v2mix-cont/sft/Think.Unbounded-d32-v2mix-cont-pre1930-curriculum-c3-robust-v3-a100-40gb/config.json ADDED
@@ -0,0 +1,97 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ {
2
+ "schema_version": 1,
3
+ "stage": "sft",
4
+ "experiment_suffix": "pre1930-curriculum-c3-robust-v3-a100-40gb",
5
+ "data": {
6
+ "recipe": "curriculum",
7
+ "curriculum": {
8
+ "name": "C3Rv3",
9
+ "mode": "staged",
10
+ "passes": 3,
11
+ "threshold_default": 80,
12
+ "stages": [
13
+ {
14
+ "routes": [
15
+ "knowledge_qa"
16
+ ],
17
+ "authentic": "single"
18
+ },
19
+ {
20
+ "routes": [
21
+ "reasoning_qa",
22
+ "stem_reasoning",
23
+ "how_to_qa",
24
+ "opinion_qa",
25
+ "composition_qa",
26
+ "verse_qa"
27
+ ],
28
+ "calibration_qa": true
29
+ },
30
+ {
31
+ "routes": [
32
+ "multiturn_qa",
33
+ "narrative_grounded",
34
+ "narrative_fiction"
35
+ ],
36
+ "authentic": "multi"
37
+ }
38
+ ],
39
+ "noise": {
40
+ "rate": 0.3,
41
+ "end_punct_rate": 0.05
42
+ },
43
+ "robustness": {
44
+ "stage": 0,
45
+ "epochs": 2.5,
46
+ "routes": {
47
+ "conversation_qa": {
48
+ "count": null
49
+ },
50
+ "unparseable_qa": {
51
+ "count": null
52
+ },
53
+ "typo_qa": {
54
+ "count": null
55
+ },
56
+ "era_qa": {
57
+ "count": null
58
+ },
59
+ "conversation_multiturn": {
60
+ "count": null
61
+ }
62
+ }
63
+ }
64
+ }
65
+ },
66
+ "training": {
67
+ "num_iterations": -1,
68
+ "load_optimizer": 0,
69
+ "device_batch_size": 1,
70
+ "warmup_ratio": 0.03,
71
+ "eval_every": 400,
72
+ "chatcore_every": 100000,
73
+ "chatcore_max_sample": 32,
74
+ "save_every": -1
75
+ },
76
+ "artifacts": {
77
+ "repo": "jbduran/think.nano"
78
+ },
79
+ "wandb": {
80
+ "enabled": true,
81
+ "entity": "jbduran-thinkingmachinesncsu",
82
+ "project": "think.nano",
83
+ "group": "think-d32",
84
+ "tags": [
85
+ "sft",
86
+ "curriculum",
87
+ "c3",
88
+ "scale-max",
89
+ "staged",
90
+ "robustness",
91
+ "noise",
92
+ "d32"
93
+ ]
94
+ },
95
+ "config_fingerprint": "b0d39eef6f8e427b",
96
+ "artifact_path": "experiments/Think.Unbounded-d32-v2mix-cont/sft/Think.Unbounded-d32-v2mix-cont-pre1930-curriculum-c3-robust-v3-a100-40gb"
97
+ }
experiments/Think.Unbounded-d32-v2mix-cont/sft/Think.Unbounded-d32-v2mix-cont-pre1930-curriculum-c3-robust-v3-a100-40gb/run.json ADDED
@@ -0,0 +1,11 @@
 
 
 
 
 
 
 
 
 
 
 
 
1
+ {
2
+ "experiment_id": "Think.Unbounded-d32-v2mix-cont-pre1930-curriculum-c3-robust-v3-a100-40gb",
3
+ "stage": "sft",
4
+ "base_experiment_id": "Think.Unbounded-d32-v2mix-cont",
5
+ "parent_experiment_id": "Think.Unbounded-d32-v2mix-cont",
6
+ "parent_checkpoint_step": 9600,
7
+ "branch_parent_step": null,
8
+ "config_fingerprint": "b0d39eef6f8e427b",
9
+ "wandb_run_id": "5667c0f5",
10
+ "created_at": 1787104196
11
+ }
experiments/Think.Unbounded-d32-v2mix-cont/sft/Think.Unbounded-d32-v2mix-cont-pre1930-curriculum-c3-robust-v3/config.json ADDED
@@ -0,0 +1,97 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ {
2
+ "schema_version": 1,
3
+ "stage": "sft",
4
+ "experiment_suffix": "pre1930-curriculum-c3-robust-v3",
5
+ "data": {
6
+ "recipe": "curriculum",
7
+ "curriculum": {
8
+ "name": "C3Rv3",
9
+ "mode": "staged",
10
+ "passes": 3,
11
+ "threshold_default": 80,
12
+ "stages": [
13
+ {
14
+ "routes": [
15
+ "knowledge_qa"
16
+ ],
17
+ "authentic": "single"
18
+ },
19
+ {
20
+ "routes": [
21
+ "reasoning_qa",
22
+ "stem_reasoning",
23
+ "how_to_qa",
24
+ "opinion_qa",
25
+ "composition_qa",
26
+ "verse_qa"
27
+ ],
28
+ "calibration_qa": true
29
+ },
30
+ {
31
+ "routes": [
32
+ "multiturn_qa",
33
+ "narrative_grounded",
34
+ "narrative_fiction"
35
+ ],
36
+ "authentic": "multi"
37
+ }
38
+ ],
39
+ "noise": {
40
+ "rate": 0.3,
41
+ "end_punct_rate": 0.05
42
+ },
43
+ "robustness": {
44
+ "stage": 0,
45
+ "epochs": 2.5,
46
+ "routes": {
47
+ "conversation_qa": {
48
+ "count": null
49
+ },
50
+ "unparseable_qa": {
51
+ "count": null
52
+ },
53
+ "typo_qa": {
54
+ "count": null
55
+ },
56
+ "era_qa": {
57
+ "count": null
58
+ },
59
+ "conversation_multiturn": {
60
+ "count": null
61
+ }
62
+ }
63
+ }
64
+ }
65
+ },
66
+ "training": {
67
+ "num_iterations": -1,
68
+ "load_optimizer": 0,
69
+ "device_batch_size": 2,
70
+ "warmup_ratio": 0.03,
71
+ "eval_every": 400,
72
+ "chatcore_every": 100000,
73
+ "chatcore_max_sample": 32,
74
+ "save_every": -1
75
+ },
76
+ "artifacts": {
77
+ "repo": "jbduran/think.nano"
78
+ },
79
+ "wandb": {
80
+ "enabled": true,
81
+ "entity": "jbduran-thinkingmachinesncsu",
82
+ "project": "think.nano",
83
+ "group": "think-d32",
84
+ "tags": [
85
+ "sft",
86
+ "curriculum",
87
+ "c3",
88
+ "scale-max",
89
+ "staged",
90
+ "robustness",
91
+ "noise",
92
+ "d32"
93
+ ]
94
+ },
95
+ "config_fingerprint": "12ffd193bdca635a",
96
+ "artifact_path": "experiments/Think.Unbounded-d32-v2mix-cont/sft/Think.Unbounded-d32-v2mix-cont-pre1930-curriculum-c3-robust-v3"
97
+ }
experiments/Think.Unbounded-d32-v2mix-cont/sft/Think.Unbounded-d32-v2mix-cont-pre1930-curriculum-c3-robust-v3/run.json ADDED
@@ -0,0 +1,11 @@
 
 
 
 
 
 
 
 
 
 
 
 
1
+ {
2
+ "experiment_id": "Think.Unbounded-d32-v2mix-cont-pre1930-curriculum-c3-robust-v3",
3
+ "stage": "sft",
4
+ "base_experiment_id": "Think.Unbounded-d32-v2mix-cont",
5
+ "parent_experiment_id": "Think.Unbounded-d32-v2mix-cont",
6
+ "parent_checkpoint_step": 9600,
7
+ "branch_parent_step": null,
8
+ "config_fingerprint": "12ffd193bdca635a",
9
+ "wandb_run_id": "10993832",
10
+ "created_at": 1787101675
11
+ }
experiments/Think.Unbounded-d32-v2mix-cont/sft/Think.Unbounded-d32-v2mix-cont-pre1930-curriculum-c3-robust/checkpoints/meta_000041.json ADDED
@@ -0,0 +1,172 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ {
2
+ "step": 41,
3
+ "training_complete": true,
4
+ "val_bpb": 0.7046443464143689,
5
+ "model_config": {
6
+ "sequence_len": 4096,
7
+ "vocab_size": 32768,
8
+ "n_layer": 32,
9
+ "n_head": 16,
10
+ "n_kv_head": 16,
11
+ "n_embd": 2048,
12
+ "window_pattern": "SSSL"
13
+ },
14
+ "user_config": {
15
+ "run": "Think.Unbounded-d32-v2mix-cont-pre1930-curriculum-c3-robust",
16
+ "wandb_run_id": "21793571",
17
+ "wandb_group": "think-d32",
18
+ "wandb_tags": "sft,curriculum,c3,scale-max,staged,robustness,noise,d32",
19
+ "device_type": "",
20
+ "model_tag": null,
21
+ "model_step": null,
22
+ "base_checkpoint_dir": "/workspace/nanochat/experiments/Think.Unbounded-d32-v2mix-cont/base_checkpoints",
23
+ "base_step": 9600,
24
+ "checkpoint_dir": "/workspace/nanochat/experiments/Think.Unbounded-d32-v2mix-cont/sft/Think.Unbounded-d32-v2mix-cont-pre1930-curriculum-c3-robust/checkpoints",
25
+ "tokenizer_dir": "/workspace/nanochat/experiments/Think.Unbounded-d32-v2mix-cont/tokenizer",
26
+ "resume_from_step": null,
27
+ "experiment_id": "Think.Unbounded-d32-v2mix-cont-pre1930-curriculum-c3-robust",
28
+ "experiment_config": "/workspace/nanochat/experiments/Think.Unbounded-d32-v2mix-cont/sft/Think.Unbounded-d32-v2mix-cont-pre1930-curriculum-c3-robust/config.json",
29
+ "parent_cumulative_flops": 2.3103503196928082e+20,
30
+ "tokenizer_fingerprint": "21e99d5cdeeaa660",
31
+ "git_commit_sha": "8549a37b028418b0abaf347bd6e5a8ccb6cff782",
32
+ "load_optimizer": 0,
33
+ "num_iterations": -1,
34
+ "max_seq_len": null,
35
+ "device_batch_size": 2,
36
+ "total_batch_size": null,
37
+ "embedding_lr": null,
38
+ "unembedding_lr": null,
39
+ "matrix_lr": null,
40
+ "init_lr_frac": 0.8,
41
+ "warmup_ratio": 0.03,
42
+ "warmdown_ratio": 0.5,
43
+ "final_lr_frac": 0.0,
44
+ "eval_every": 400,
45
+ "eval_tokens": 20971520,
46
+ "chatcore_every": 100000,
47
+ "chatcore_max_cat": -1,
48
+ "chatcore_max_sample": 32,
49
+ "save_every": -1,
50
+ "recipe": "curriculum",
51
+ "curriculum_config": "",
52
+ "pre1930_epochs": 5,
53
+ "mmlu_epochs": 3,
54
+ "gsm8k_epochs": 4,
55
+ "authentic_epochs": 0,
56
+ "knowledge_qa_epochs": 0,
57
+ "multiturn_qa_epochs": 0,
58
+ "reasoning_qa_epochs": 0,
59
+ "stem_reasoning_epochs": 0,
60
+ "narrative_grounded_epochs": 0,
61
+ "narrative_fiction_epochs": 0,
62
+ "opinion_qa_epochs": 0,
63
+ "how_to_qa_epochs": 0,
64
+ "verse_qa_epochs": 0,
65
+ "composition_qa_epochs": 0,
66
+ "resolved_experiment_config": {
67
+ "schema_version": 1,
68
+ "stage": "sft",
69
+ "experiment_suffix": "pre1930-curriculum-c3-robust",
70
+ "data": {
71
+ "recipe": "curriculum",
72
+ "curriculum": {
73
+ "name": "C3R",
74
+ "mode": "staged",
75
+ "threshold_default": 80,
76
+ "stages": [
77
+ {
78
+ "routes": [
79
+ "knowledge_qa"
80
+ ],
81
+ "authentic": "single"
82
+ },
83
+ {
84
+ "routes": [
85
+ "reasoning_qa",
86
+ "stem_reasoning",
87
+ "how_to_qa",
88
+ "opinion_qa",
89
+ "composition_qa",
90
+ "verse_qa"
91
+ ],
92
+ "calibration_qa": true
93
+ },
94
+ {
95
+ "routes": [
96
+ "multiturn_qa",
97
+ "narrative_grounded",
98
+ "narrative_fiction"
99
+ ],
100
+ "authentic": "multi"
101
+ }
102
+ ],
103
+ "noise": {
104
+ "rate": 0.3
105
+ },
106
+ "robustness": {
107
+ "stage": 0,
108
+ "epochs": 1,
109
+ "routes": {
110
+ "conversation_qa": {
111
+ "count": null
112
+ },
113
+ "unparseable_qa": {
114
+ "count": null
115
+ },
116
+ "era_qa": {
117
+ "count": null
118
+ }
119
+ }
120
+ }
121
+ }
122
+ },
123
+ "training": {
124
+ "num_iterations": -1,
125
+ "load_optimizer": 0,
126
+ "device_batch_size": 2,
127
+ "warmup_ratio": 0.03,
128
+ "eval_every": 400,
129
+ "chatcore_every": 100000,
130
+ "chatcore_max_sample": 32,
131
+ "save_every": -1
132
+ },
133
+ "artifacts": {
134
+ "repo": "jbduran/think.nano"
135
+ },
136
+ "wandb": {
137
+ "enabled": true,
138
+ "entity": "jbduran-thinkingmachinesncsu",
139
+ "project": "think.nano",
140
+ "group": "think-d32",
141
+ "tags": [
142
+ "sft",
143
+ "curriculum",
144
+ "c3",
145
+ "scale-max",
146
+ "staged",
147
+ "robustness",
148
+ "noise",
149
+ "d32"
150
+ ]
151
+ },
152
+ "config_fingerprint": "e1a9d25fb1cf5edf",
153
+ "artifact_path": "experiments/Think.Unbounded-d32-v2mix-cont/sft/Think.Unbounded-d32-v2mix-cont-pre1930-curriculum-c3-robust"
154
+ },
155
+ "stage": "sft",
156
+ "base_experiment_id": null,
157
+ "parent_experiment_id": null,
158
+ "parent_checkpoint_step": null,
159
+ "config_fingerprint": "e1a9d25fb1cf5edf"
160
+ },
161
+ "loop_state": {
162
+ "step": 41,
163
+ "total_training_time": 1459.9074046611786,
164
+ "min_val_bpb": 0.7046443464143689,
165
+ "smooth_train_loss": 1.907550057536094,
166
+ "mfu": 51.698687069018604,
167
+ "tok_per_sec": 44555,
168
+ "stage_training_flops": 9.867121157021368e+17,
169
+ "inherited_parent_flops": 2.3103503196928082e+20,
170
+ "cumulative_pipeline_training_flops": 2.3202174408498296e+20
171
+ }
172
+ }
experiments/Think.Unbounded-d32-v2mix-cont/sft/Think.Unbounded-d32-v2mix-cont-pre1930-curriculum-c3-robust/checkpoints/model_000041.pt ADDED
@@ -0,0 +1,3 @@
 
 
 
 
1
+ version https://git-lfs.github.com/spec/v1
2
+ oid sha256:590b64a7ab561ae7ec7adf10fd09469e7b46dacf76e5c665f743fd9d0dd74a6f
3
+ size 8992694653
experiments/Think.Unbounded-d32-v2mix-cont/sft/Think.Unbounded-d32-v2mix-cont-pre1930-curriculum-c3-robust/checkpoints/optim_000041_rank0.pt ADDED
@@ -0,0 +1,3 @@
 
 
 
 
1
+ version https://git-lfs.github.com/spec/v1
2
+ oid sha256:7c642c3e04109fce332ca03b7afaf571812e0d9ac124f8863889f262a6748995
3
+ size 11545903753
experiments/Think.Unbounded-d32-v2mix-cont/sft/Think.Unbounded-d32-v2mix-cont-pre1930-curriculum-c3-robust/config.json ADDED
@@ -0,0 +1,89 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ {
2
+ "schema_version": 1,
3
+ "stage": "sft",
4
+ "experiment_suffix": "pre1930-curriculum-c3-robust",
5
+ "data": {
6
+ "recipe": "curriculum",
7
+ "curriculum": {
8
+ "name": "C3R",
9
+ "mode": "staged",
10
+ "threshold_default": 80,
11
+ "stages": [
12
+ {
13
+ "routes": [
14
+ "knowledge_qa"
15
+ ],
16
+ "authentic": "single"
17
+ },
18
+ {
19
+ "routes": [
20
+ "reasoning_qa",
21
+ "stem_reasoning",
22
+ "how_to_qa",
23
+ "opinion_qa",
24
+ "composition_qa",
25
+ "verse_qa"
26
+ ],
27
+ "calibration_qa": true
28
+ },
29
+ {
30
+ "routes": [
31
+ "multiturn_qa",
32
+ "narrative_grounded",
33
+ "narrative_fiction"
34
+ ],
35
+ "authentic": "multi"
36
+ }
37
+ ],
38
+ "noise": {
39
+ "rate": 0.3
40
+ },
41
+ "robustness": {
42
+ "stage": 0,
43
+ "epochs": 1,
44
+ "routes": {
45
+ "conversation_qa": {
46
+ "count": null
47
+ },
48
+ "unparseable_qa": {
49
+ "count": null
50
+ },
51
+ "era_qa": {
52
+ "count": null
53
+ }
54
+ }
55
+ }
56
+ }
57
+ },
58
+ "training": {
59
+ "num_iterations": -1,
60
+ "load_optimizer": 0,
61
+ "device_batch_size": 2,
62
+ "warmup_ratio": 0.03,
63
+ "eval_every": 400,
64
+ "chatcore_every": 100000,
65
+ "chatcore_max_sample": 32,
66
+ "save_every": -1
67
+ },
68
+ "artifacts": {
69
+ "repo": "jbduran/think.nano"
70
+ },
71
+ "wandb": {
72
+ "enabled": true,
73
+ "entity": "jbduran-thinkingmachinesncsu",
74
+ "project": "think.nano",
75
+ "group": "think-d32",
76
+ "tags": [
77
+ "sft",
78
+ "curriculum",
79
+ "c3",
80
+ "scale-max",
81
+ "staged",
82
+ "robustness",
83
+ "noise",
84
+ "d32"
85
+ ]
86
+ },
87
+ "config_fingerprint": "e1a9d25fb1cf5edf",
88
+ "artifact_path": "experiments/Think.Unbounded-d32-v2mix-cont/sft/Think.Unbounded-d32-v2mix-cont-pre1930-curriculum-c3-robust"
89
+ }
experiments/Think.Unbounded-d32-v2mix-cont/sft/Think.Unbounded-d32-v2mix-cont-pre1930-curriculum-c3-robust/run.json ADDED
@@ -0,0 +1,11 @@
 
 
 
 
 
 
 
 
 
 
 
 
1
+ {
2
+ "experiment_id": "Think.Unbounded-d32-v2mix-cont-pre1930-curriculum-c3-robust",
3
+ "stage": "sft",
4
+ "base_experiment_id": "Think.Unbounded-d32-v2mix-cont",
5
+ "parent_experiment_id": "Think.Unbounded-d32-v2mix-cont",
6
+ "parent_checkpoint_step": 9600,
7
+ "branch_parent_step": null,
8
+ "config_fingerprint": "e1a9d25fb1cf5edf",
9
+ "wandb_run_id": "21793571",
10
+ "created_at": 1786635058
11
+ }
experiments/Think.Unbounded-d32-v2mix-cont/summary.json ADDED
@@ -0,0 +1,79 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ {
2
+ "experiment_id": "Think.Unbounded-d32-v2mix-cont",
3
+ "stage": "base",
4
+ "base_experiment_id": "Think.Unbounded-d32-v2mix-cont",
5
+ "parent_experiment_id": "Think.Unbounded-d32",
6
+ "parent_checkpoint_step": 5500,
7
+ "dataset": "jbduran/think-dataset-clean-1930s",
8
+ "dataset_revision": "45225d95bc15f942be3b4b344738cea1f66e3de8",
9
+ "step": 9600,
10
+ "depth": 32,
11
+ "target_param_data_ratio": 12,
12
+ "training_tokens": 20132659200,
13
+ "final_sampled_val_bpb": 0.7708812390249374,
14
+ "minimum_sampled_val_bpb": 0.7701763237939314,
15
+ "full_val_bpb": 0.7505380497637221,
16
+ "core_metric": null,
17
+ "centered_results": null,
18
+ "conditioned_samples": [
19
+ {
20
+ "prompt": "The capital of France is",
21
+ "text": "<|bos|>The capital of France is the most important city in the world. It is the capital of the most powerful"
22
+ },
23
+ {
24
+ "prompt": "The chemical symbol of gold is",
25
+ "text": "<|bos|>The chemical symbol of gold is Au, and that of silver is Ag. The symbol of the former is Ag"
26
+ },
27
+ {
28
+ "prompt": "If yesterday was Friday, then tomorrow will be",
29
+ "text": "<|bos|>If yesterday was Friday, then tomorrow will be Saturday. If yesterday was Saturday, then tomorrow will be Sunday. If yesterday was"
30
+ },
31
+ {
32
+ "prompt": "The opposite of hot is",
33
+ "text": "<|bos|>The opposite of hot is cold, and the opposite of cold is hot. The opposite of hot is cold"
34
+ },
35
+ {
36
+ "prompt": "The planets of the solar system are:",
37
+ "text": "<|bos|>The planets of the solar system are: Mercury, Venus, Mars, Jupiter, Saturn, and Herschel. The earth"
38
+ },
39
+ {
40
+ "prompt": "My favorite color is",
41
+ "text": "<|bos|>My favorite color is a deep, rich, velvety crimson, the same shade as the rose"
42
+ },
43
+ {
44
+ "prompt": "If 5*x + 3 = 13, then x is",
45
+ "text": "<|bos|>If 5*x + 3 = 13, then x is 5.\n\nIf 3x + 4 = 13, then x"
46
+ }
47
+ ],
48
+ "unconditioned_samples": [
49
+ "<|bos|>PREFACE.\n\nTHIS essay originated with Mr. W. C. Barrett, who says of it: \"Mr. Galton's 'Life and Times,' and his Life of Darwin.' \"It is mainly a half-tone reprint from Galton's (vice supra) life, to which I have added some important facts which I had not been able to procure when the original edition was published, and which bear directly upon some matters of interest.\n\nIt may, however, be as well to draw attention more particularly to the point to which Mr. Barrett refers in the above quotation, as this contains, I believe, the fundamental distinction between the Darwin",
50
+ "<|bos|>Lady Saumarez -- Scotch rose shaded white. Pheasant's Eye -- Pure white, dark eye. Marigold -- Striking light golden color. Maddereti -- A most desirable purple.\n\nRose -- -A superb full flowering rose of a most beautiful shade of deep rose. Sunrise -- Pure tints of fawn and yellow.\n\nCarnation Marguerite\n\nLarge Flowered\n\nnow offered. Much hardier and blooms more abundantly.\n\nPrice, any variety, 10 cts. each; set of 1 3 for 45 cents.\n\nEasily Makes a Good Center\n\nAnywhere.\n\n",
51
+ "<|bos|>RIDGEN, TOLLAND\n\nFRIDLAND, Dyfe, y 1665.\n\nTHE CONFUCTIONS\n\nor\n\nTHE KINGDOM OF SOLOMON.\n\nTHE FIRST BOOK OF THE KING OF SOLOMON\n\nRefers to the Four Series of the Four Dispensations, God revealing himself to His Children in the various Births of His Israel. The Messiah's birth and baptism of John, the mission of the\n\nApostles, and the baptism of Christ, and latterly the majestic progress of His power, are chronicled in the Books of Moses, of the Prophets, of the Proph ets, and the",
52
+ "<|bos|>agers have already told us, and M. Crookes has already informed us, that he found concretions of this nature in the plantlice when giving a supplementary examination to them : hence, even according to his own showing, no safe conclusion can be drawn from the relative order and distribution of these silica-precipitates. At the same time it is natural to ask whether this tardy means of elucidation, instead of becoming antiquated by the lapse of the fair period which Hibbert has now completed-and well-nigh brought to a close by a misuse of some words, may not be pressed into use - --\n\nON",
53
+ "<|bos|>D\n\nio\n\nLe\n\nthered in the middle sept i post- ater bear ried ne pro avera hing tenanted\n\nCrowne he se ouch h find\n\n\u200bLa thi seeks depone \u0f51 \u0f4f \u0f63\u0f74 \u0f49\u0f7a\u0f51 \u0f61 \u0f61 \u0f61\u0f72 \u0f64\u0f74 \u0f62 \u0f61 \u0f51 \u0f62 \u3002 \u0f61\u0f72\u0970\u0f62 \u0f51 \u0f0b \u0f62\u0f66 Atth Id\n\nDedication of the Picture in the Library of Trinity that has the most frequently called forth criticism\n\nI am unable to",
54
+ "<|bos|>PREFACE\n\nCONTENTS.\n\n1. X. About a month was being wasted at the hotel, when a kindhearted friend offered to go on ahead and engage rooms for us at the best houses the day was hot-the Southern hotels always are-on the other side, and on most of them, indeed, fair profit can be made; and the diligence kept always waiting at the station and often not in readiness for an hour after the last customer had been put in, and the engineer grumbled fearfully at this irregularity of the train, because the trains gave him a pretty lively time when there was anybody waiting there to see, and the porters sometimes began",
55
+ "<|bos|>Clarke Model\n\n222000908000 DIANER DER\n\nThe proper unture dishes\n\nWilder Damases7",
56
+ "<|bos|>ossoer 0 020 oO 000000 ee OP atte eT es ee eee en eee Bee Se SERRE Nee errre reser Res eererr ee eee cerr reeer era Nee essen re ers 2008 OO OO DO ES OO ESE HD TTI MT SPEC IC PRT APOLOC H OCDOETTES 50 Ce a OC ee DDg ee CF EIOLUS) oe eA tt } 500 een eeehey n"
57
+ ],
58
+ "training_time_seconds": 164042.2579112053,
59
+ "stage_training_flops": 9.867121157021368e+19,
60
+ "inherited_parent_flops": 1.3236382039906714e+20,
61
+ "cumulative_pipeline_training_flops": 2.3103503196928082e+20,
62
+ "config_fingerprint": "00af0bde578c1269",
63
+ "git_commit_sha": "734396fce10c8fcd7bf979261d41ebe37968089c",
64
+ "wandb_url": "https://wandb.ai/jbduran-thinkingmachinesncsu/think.nano/runs/2fa1b744",
65
+ "huggingface_url": "https://huggingface.co/jbduran/think.nano/tree/main/experiments/Think.Unbounded-d32-v2mix-cont",
66
+ "dataset_fingerprint": "54cc47b13796e411",
67
+ "tokenizer_fingerprint": "21e99d5cdeeaa660",
68
+ "unique_train_tokens_per_source": {
69
+ "original": 14515647284,
70
+ "midtrain_r21": 4107062477,
71
+ "midtrain_r45": 2053531239
72
+ },
73
+ "branch_parent_experiment_id": "Think.Unbounded-d32",
74
+ "branch_parent_step": 5500,
75
+ "branch_lr_schedule": "continue",
76
+ "stage_training_tokens": 8598323200,
77
+ "unique_train_tokens": 20676241000,
78
+ "effective_epochs": 0.41585524177242855
79
+ }
experiments/Think.Unbounded-d32-v2mix-cont/tokenizer/experiment_tokenizer.json ADDED
@@ -0,0 +1,18 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ {
2
+ "experiment_id": "clean1930s-d24-r12-ctx4096-fulltok-v1",
3
+ "dataset": {
4
+ "adapter": "parquet_shards",
5
+ "repo": "jbduran/think-dataset-clean-1930s",
6
+ "revision": "main",
7
+ "validation_shard": 472,
8
+ "num_train_shards": 200,
9
+ "download_workers": 4
10
+ },
11
+ "tokenizer": {
12
+ "mode": "train",
13
+ "max_chars": 1000000000000,
14
+ "doc_cap": 1000000000,
15
+ "vocab_size": 32768
16
+ },
17
+ "created_at": 1784129374
18
+ }
experiments/Think.Unbounded-d32-v2mix-cont/tokenizer/token_bytes.pt ADDED
@@ -0,0 +1,3 @@
 
 
 
 
1
+ version https://git-lfs.github.com/spec/v1
2
+ oid sha256:bc779ae25dfa6f35146f7b9991fa3bab9f2a82a89a4dd92fbd4a57655680e7e2
3
+ size 132649
experiments/Think.Unbounded-d32-v2mix-cont/tokenizer/tokenizer.pkl ADDED
@@ -0,0 +1,3 @@
 
 
 
 
1
+ version https://git-lfs.github.com/spec/v1
2
+ oid sha256:155d20e57ea2203cab333207e97b2bec8c0d224678bfce2f019cff3a8ddd940f
3
+ size 410542
experiments/Think.Unbounded-d32/base_checkpoints/meta_000500.json ADDED
@@ -0,0 +1,236 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ {
2
+ "step": 500,
3
+ "training_complete": false,
4
+ "experiment_id": "Think.Unbounded-d32",
5
+ "parent_experiment_id": null,
6
+ "parent_checkpoint_step": null,
7
+ "val_bpb": 1.1476923999299866,
8
+ "model_config": {
9
+ "sequence_len": 4096,
10
+ "vocab_size": 32768,
11
+ "n_layer": 32,
12
+ "n_head": 16,
13
+ "n_kv_head": 16,
14
+ "n_embd": 2048,
15
+ "window_pattern": "SSSL"
16
+ },
17
+ "user_config": {
18
+ "run": "Think.Unbounded-d32",
19
+ "wandb_run_id": "9b75691b",
20
+ "wandb_group": "clean1930s-d32",
21
+ "wandb_tags": "think-dataset-clean-1930s,d32,ratio12,ctx4096,sssl,fp8,muon-momentum-constant,muon-momentum-0.90,midtrain-schedule,mix-og,0-30-60",
22
+ "device_type": "",
23
+ "fp8": true,
24
+ "fp8_recipe": "tensorwise",
25
+ "depth": 32,
26
+ "aspect_ratio": 64,
27
+ "head_dim": 128,
28
+ "max_seq_len": 4096,
29
+ "window_pattern": "SSSL",
30
+ "num_iterations": 9600,
31
+ "target_flops": -1.0,
32
+ "target_param_data_ratio": -1.0,
33
+ "device_batch_size": 2,
34
+ "total_batch_size": 2097152,
35
+ "embedding_lr": 0.3,
36
+ "unembedding_lr": 0.008,
37
+ "weight_decay": 0.28,
38
+ "matrix_lr": 0.02,
39
+ "muon_momentum": 0.9,
40
+ "scalar_lr": 0.5,
41
+ "warmup_steps": 40,
42
+ "warmdown_ratio": 0.65,
43
+ "final_lr_frac": 0.05,
44
+ "resume_from_step": -1,
45
+ "init_from_checkpoint_dir": null,
46
+ "init_from_step": -1,
47
+ "no_init_optimizer": false,
48
+ "branch_lr_schedule": "branch",
49
+ "branch_parent_experiment_id": null,
50
+ "pretokenized": true,
51
+ "data_dir": "/workspace/nanochat/experiments/Think.Unbounded-d32/data",
52
+ "tokenizer_dir": "/workspace/nanochat/experiments/Think.Unbounded-d32/tokenizer",
53
+ "pretokenized_dir": "/workspace/nanochat/experiments/Think.Unbounded-d32/pretok",
54
+ "mixture_source_dirs": "{\"original\": \"/workspace/nanochat/experiments/Think.Unbounded-d32/pretok_original\", \"midtrain_r30\": \"/workspace/nanochat/experiments/Think.Unbounded-d32/pretok_midtrain_r30\", \"midtrain_r60\": \"/workspace/nanochat/experiments/Think.Unbounded-d32/pretok_midtrain_r60\"}",
55
+ "checkpoint_dir": "/workspace/nanochat/experiments/Think.Unbounded-d32/base_checkpoints",
56
+ "experiment_id": "Think.Unbounded-d32",
57
+ "experiment_config": "/workspace/nanochat/experiments/Think.Unbounded-d32/config.json",
58
+ "tokenizer_fingerprint": "21e99d5cdeeaa660",
59
+ "git_commit_sha": "b2de514e00cde3bee362cf4c87e16d35e0f32c24",
60
+ "seed": 42,
61
+ "eval_every": 250,
62
+ "eval_tokens": 2097152,
63
+ "core_metric_every": -1,
64
+ "core_metric_max_per_task": 500,
65
+ "sample_every": -1,
66
+ "save_every": 500,
67
+ "model_tag": "Think.Unbounded-d32",
68
+ "experiment": {
69
+ "schema_version": 1,
70
+ "stage": "base",
71
+ "experiment_id": "Think.Unbounded-d32",
72
+ "datasets": {
73
+ "original": {
74
+ "adapter": "parquet_shards",
75
+ "repo": "jbduran/think-dataset-clean-1930s",
76
+ "revision": "45225d95bc15f942be3b4b344738cea1f66e3de8",
77
+ "validation_shard": 472,
78
+ "num_train_shards": 330,
79
+ "download_workers": 4
80
+ },
81
+ "midtrain_r30": {
82
+ "adapter": "parquet_shards",
83
+ "repo": "zachnorton03/think-midtrain",
84
+ "revision": "9ace24b8e16e57e38a1ea0b1f6d7cbf323bf9dbc",
85
+ "subfolder": "mixed/ratio_30/data",
86
+ "validation_shard": 1262,
87
+ "num_train_shards": 391,
88
+ "download_workers": 4
89
+ },
90
+ "midtrain_r60": {
91
+ "adapter": "parquet_shards",
92
+ "repo": "zachnorton03/think-midtrain",
93
+ "revision": "9ace24b8e16e57e38a1ea0b1f6d7cbf323bf9dbc",
94
+ "subfolder": "mixed/ratio_60/data",
95
+ "validation_shard": 523,
96
+ "num_train_shards": 207,
97
+ "download_workers": 4
98
+ }
99
+ },
100
+ "mixture_schedule": {
101
+ "total_tokens": 20132659200,
102
+ "seed_data": 1234,
103
+ "max_epochs": 4,
104
+ "stages": [
105
+ {
106
+ "name": "base",
107
+ "start_tokens": 0,
108
+ "source": "original"
109
+ },
110
+ {
111
+ "name": "injection",
112
+ "start_tokens": 14092861440,
113
+ "source": "midtrain_r30"
114
+ },
115
+ {
116
+ "name": "decay_mix",
117
+ "start_tokens": 18119393280,
118
+ "source": "midtrain_r60"
119
+ }
120
+ ]
121
+ },
122
+ "tokenizer": {
123
+ "mode": "reuse",
124
+ "source_experiment_id": "clean1930s-d24-r12-ctx4096-fulltok-v1",
125
+ "vocab_size": 32768
126
+ },
127
+ "pretokenize": {
128
+ "enabled": true,
129
+ "slack": 1.03,
130
+ "require_no_wrap": true,
131
+ "val_tokens": 20971520,
132
+ "shard_tokens": 100000000,
133
+ "tokenizer_threads": 8
134
+ },
135
+ "training": {
136
+ "depth": 32,
137
+ "scaling_params": 1677724672,
138
+ "target_param_data_ratio": 12,
139
+ "max_seq_len": 4096,
140
+ "window_pattern": "SSSL",
141
+ "muon_momentum": 0.9,
142
+ "device_batch_size": 2,
143
+ "total_batch_size": 2097152,
144
+ "fp8": true,
145
+ "fp8_recipe": "tensorwise",
146
+ "seed": 42,
147
+ "save_every": 500,
148
+ "eval_every": 250,
149
+ "eval_tokens": 2097152,
150
+ "core_metric_every": -1,
151
+ "sample_every": -1
152
+ },
153
+ "artifacts": {
154
+ "repo": "jbduran/think.nano",
155
+ "keep_local_checkpoints": 1
156
+ },
157
+ "wandb": {
158
+ "entity": "jbduran-thinkingmachinesncsu",
159
+ "project": "think.nano",
160
+ "name": "Think.Unbounded-d32",
161
+ "group": "clean1930s-d32",
162
+ "tags": [
163
+ "think-dataset-clean-1930s",
164
+ "d32",
165
+ "ratio12",
166
+ "ctx4096",
167
+ "sssl",
168
+ "fp8",
169
+ "muon-momentum-constant",
170
+ "muon-momentum-0.90",
171
+ "midtrain-schedule",
172
+ "mix-og",
173
+ "0-30-60"
174
+ ]
175
+ },
176
+ "config_fingerprint": "6f8ae461e0eb9ee2",
177
+ "artifact_path": "experiments/Think.Unbounded-d32"
178
+ },
179
+ "stage": "base",
180
+ "base_experiment_id": "Think.Unbounded-d32",
181
+ "parent_experiment_id": null,
182
+ "parent_checkpoint_step": null,
183
+ "config_fingerprint": "6f8ae461e0eb9ee2"
184
+ },
185
+ "device_batch_size": 2,
186
+ "max_seq_len": 4096,
187
+ "total_batch_size": 2097152,
188
+ "dataloader_state_dict": {
189
+ "file_idx": 10,
190
+ "pos": 48712193,
191
+ "epoch": 1,
192
+ "pq_idx": 10,
193
+ "rg_idx": 48712193,
194
+ "mixture": {
195
+ "cursors": {
196
+ "original": {
197
+ "file_idx": 10,
198
+ "pos": 48712193,
199
+ "epoch": 1,
200
+ "pq_idx": 10,
201
+ "rg_idx": 48712193
202
+ },
203
+ "midtrain_r30": {
204
+ "file_idx": 0,
205
+ "pos": 0,
206
+ "epoch": 1,
207
+ "pq_idx": 0,
208
+ "rg_idx": 0
209
+ },
210
+ "midtrain_r60": {
211
+ "file_idx": 0,
212
+ "pos": 0,
213
+ "epoch": 1,
214
+ "pq_idx": 0,
215
+ "rg_idx": 0
216
+ }
217
+ },
218
+ "cumulative_tokens": 1048584192,
219
+ "source_tokens": {
220
+ "original": 1048584192,
221
+ "midtrain_r30": 0,
222
+ "midtrain_r60": 0
223
+ },
224
+ "active_stage_idx": 0
225
+ }
226
+ },
227
+ "loop_state": {
228
+ "min_val_bpb": 1.1476923999299866,
229
+ "smooth_train_loss": 3.427469208772481,
230
+ "total_training_time": 19713.592538118362,
231
+ "stage_start_step": 0,
232
+ "stage_training_flops": 12033074581733376000,
233
+ "inherited_parent_flops": 0.0,
234
+ "cumulative_pipeline_training_flops": 1.2033074581733376e+19
235
+ }
236
+ }
experiments/Think.Unbounded-d32/base_checkpoints/meta_001000.json ADDED
@@ -0,0 +1,236 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ {
2
+ "step": 1000,
3
+ "training_complete": false,
4
+ "experiment_id": "Think.Unbounded-d32",
5
+ "parent_experiment_id": null,
6
+ "parent_checkpoint_step": null,
7
+ "val_bpb": 1.056581441012716,
8
+ "model_config": {
9
+ "sequence_len": 4096,
10
+ "vocab_size": 32768,
11
+ "n_layer": 32,
12
+ "n_head": 16,
13
+ "n_kv_head": 16,
14
+ "n_embd": 2048,
15
+ "window_pattern": "SSSL"
16
+ },
17
+ "user_config": {
18
+ "run": "Think.Unbounded-d32",
19
+ "wandb_run_id": "9b75691b",
20
+ "wandb_group": "clean1930s-d32",
21
+ "wandb_tags": "think-dataset-clean-1930s,d32,ratio12,ctx4096,sssl,fp8,muon-momentum-constant,muon-momentum-0.90,midtrain-schedule,mix-og,0-30-60",
22
+ "device_type": "",
23
+ "fp8": true,
24
+ "fp8_recipe": "tensorwise",
25
+ "depth": 32,
26
+ "aspect_ratio": 64,
27
+ "head_dim": 128,
28
+ "max_seq_len": 4096,
29
+ "window_pattern": "SSSL",
30
+ "num_iterations": 9600,
31
+ "target_flops": -1.0,
32
+ "target_param_data_ratio": -1.0,
33
+ "device_batch_size": 2,
34
+ "total_batch_size": 2097152,
35
+ "embedding_lr": 0.3,
36
+ "unembedding_lr": 0.008,
37
+ "weight_decay": 0.28,
38
+ "matrix_lr": 0.02,
39
+ "muon_momentum": 0.9,
40
+ "scalar_lr": 0.5,
41
+ "warmup_steps": 40,
42
+ "warmdown_ratio": 0.65,
43
+ "final_lr_frac": 0.05,
44
+ "resume_from_step": -1,
45
+ "init_from_checkpoint_dir": null,
46
+ "init_from_step": -1,
47
+ "no_init_optimizer": false,
48
+ "branch_lr_schedule": "branch",
49
+ "branch_parent_experiment_id": null,
50
+ "pretokenized": true,
51
+ "data_dir": "/workspace/nanochat/experiments/Think.Unbounded-d32/data",
52
+ "tokenizer_dir": "/workspace/nanochat/experiments/Think.Unbounded-d32/tokenizer",
53
+ "pretokenized_dir": "/workspace/nanochat/experiments/Think.Unbounded-d32/pretok",
54
+ "mixture_source_dirs": "{\"original\": \"/workspace/nanochat/experiments/Think.Unbounded-d32/pretok_original\", \"midtrain_r30\": \"/workspace/nanochat/experiments/Think.Unbounded-d32/pretok_midtrain_r30\", \"midtrain_r60\": \"/workspace/nanochat/experiments/Think.Unbounded-d32/pretok_midtrain_r60\"}",
55
+ "checkpoint_dir": "/workspace/nanochat/experiments/Think.Unbounded-d32/base_checkpoints",
56
+ "experiment_id": "Think.Unbounded-d32",
57
+ "experiment_config": "/workspace/nanochat/experiments/Think.Unbounded-d32/config.json",
58
+ "tokenizer_fingerprint": "21e99d5cdeeaa660",
59
+ "git_commit_sha": "b2de514e00cde3bee362cf4c87e16d35e0f32c24",
60
+ "seed": 42,
61
+ "eval_every": 250,
62
+ "eval_tokens": 2097152,
63
+ "core_metric_every": -1,
64
+ "core_metric_max_per_task": 500,
65
+ "sample_every": -1,
66
+ "save_every": 500,
67
+ "model_tag": "Think.Unbounded-d32",
68
+ "experiment": {
69
+ "schema_version": 1,
70
+ "stage": "base",
71
+ "experiment_id": "Think.Unbounded-d32",
72
+ "datasets": {
73
+ "original": {
74
+ "adapter": "parquet_shards",
75
+ "repo": "jbduran/think-dataset-clean-1930s",
76
+ "revision": "45225d95bc15f942be3b4b344738cea1f66e3de8",
77
+ "validation_shard": 472,
78
+ "num_train_shards": 330,
79
+ "download_workers": 4
80
+ },
81
+ "midtrain_r30": {
82
+ "adapter": "parquet_shards",
83
+ "repo": "zachnorton03/think-midtrain",
84
+ "revision": "9ace24b8e16e57e38a1ea0b1f6d7cbf323bf9dbc",
85
+ "subfolder": "mixed/ratio_30/data",
86
+ "validation_shard": 1262,
87
+ "num_train_shards": 391,
88
+ "download_workers": 4
89
+ },
90
+ "midtrain_r60": {
91
+ "adapter": "parquet_shards",
92
+ "repo": "zachnorton03/think-midtrain",
93
+ "revision": "9ace24b8e16e57e38a1ea0b1f6d7cbf323bf9dbc",
94
+ "subfolder": "mixed/ratio_60/data",
95
+ "validation_shard": 523,
96
+ "num_train_shards": 207,
97
+ "download_workers": 4
98
+ }
99
+ },
100
+ "mixture_schedule": {
101
+ "total_tokens": 20132659200,
102
+ "seed_data": 1234,
103
+ "max_epochs": 4,
104
+ "stages": [
105
+ {
106
+ "name": "base",
107
+ "start_tokens": 0,
108
+ "source": "original"
109
+ },
110
+ {
111
+ "name": "injection",
112
+ "start_tokens": 14092861440,
113
+ "source": "midtrain_r30"
114
+ },
115
+ {
116
+ "name": "decay_mix",
117
+ "start_tokens": 18119393280,
118
+ "source": "midtrain_r60"
119
+ }
120
+ ]
121
+ },
122
+ "tokenizer": {
123
+ "mode": "reuse",
124
+ "source_experiment_id": "clean1930s-d24-r12-ctx4096-fulltok-v1",
125
+ "vocab_size": 32768
126
+ },
127
+ "pretokenize": {
128
+ "enabled": true,
129
+ "slack": 1.03,
130
+ "require_no_wrap": true,
131
+ "val_tokens": 20971520,
132
+ "shard_tokens": 100000000,
133
+ "tokenizer_threads": 8
134
+ },
135
+ "training": {
136
+ "depth": 32,
137
+ "scaling_params": 1677724672,
138
+ "target_param_data_ratio": 12,
139
+ "max_seq_len": 4096,
140
+ "window_pattern": "SSSL",
141
+ "muon_momentum": 0.9,
142
+ "device_batch_size": 2,
143
+ "total_batch_size": 2097152,
144
+ "fp8": true,
145
+ "fp8_recipe": "tensorwise",
146
+ "seed": 42,
147
+ "save_every": 500,
148
+ "eval_every": 250,
149
+ "eval_tokens": 2097152,
150
+ "core_metric_every": -1,
151
+ "sample_every": -1
152
+ },
153
+ "artifacts": {
154
+ "repo": "jbduran/think.nano",
155
+ "keep_local_checkpoints": 1
156
+ },
157
+ "wandb": {
158
+ "entity": "jbduran-thinkingmachinesncsu",
159
+ "project": "think.nano",
160
+ "name": "Think.Unbounded-d32",
161
+ "group": "clean1930s-d32",
162
+ "tags": [
163
+ "think-dataset-clean-1930s",
164
+ "d32",
165
+ "ratio12",
166
+ "ctx4096",
167
+ "sssl",
168
+ "fp8",
169
+ "muon-momentum-constant",
170
+ "muon-momentum-0.90",
171
+ "midtrain-schedule",
172
+ "mix-og",
173
+ "0-30-60"
174
+ ]
175
+ },
176
+ "config_fingerprint": "6f8ae461e0eb9ee2",
177
+ "artifact_path": "experiments/Think.Unbounded-d32"
178
+ },
179
+ "stage": "base",
180
+ "base_experiment_id": "Think.Unbounded-d32",
181
+ "parent_experiment_id": null,
182
+ "parent_checkpoint_step": null,
183
+ "config_fingerprint": "6f8ae461e0eb9ee2"
184
+ },
185
+ "device_batch_size": 2,
186
+ "max_seq_len": 4096,
187
+ "total_batch_size": 2097152,
188
+ "dataloader_state_dict": {
189
+ "file_idx": 20,
190
+ "pos": 97416193,
191
+ "epoch": 1,
192
+ "pq_idx": 20,
193
+ "rg_idx": 97416193,
194
+ "mixture": {
195
+ "cursors": {
196
+ "original": {
197
+ "file_idx": 20,
198
+ "pos": 97416193,
199
+ "epoch": 1,
200
+ "pq_idx": 20,
201
+ "rg_idx": 97416193
202
+ },
203
+ "midtrain_r30": {
204
+ "file_idx": 0,
205
+ "pos": 0,
206
+ "epoch": 1,
207
+ "pq_idx": 0,
208
+ "rg_idx": 0
209
+ },
210
+ "midtrain_r60": {
211
+ "file_idx": 0,
212
+ "pos": 0,
213
+ "epoch": 1,
214
+ "pq_idx": 0,
215
+ "rg_idx": 0
216
+ }
217
+ },
218
+ "cumulative_tokens": 2097160192,
219
+ "source_tokens": {
220
+ "original": 2097160192,
221
+ "midtrain_r30": 0,
222
+ "midtrain_r60": 0
223
+ },
224
+ "active_stage_idx": 0
225
+ }
226
+ },
227
+ "loop_state": {
228
+ "min_val_bpb": 1.056581441012716,
229
+ "smooth_train_loss": 2.9296221567382985,
230
+ "total_training_time": 39794.540657281876,
231
+ "stage_start_step": 0,
232
+ "stage_training_flops": 24066149163466752000,
233
+ "inherited_parent_flops": 0.0,
234
+ "cumulative_pipeline_training_flops": 2.406614916346675e+19
235
+ }
236
+ }
experiments/Think.Unbounded-d32/base_checkpoints/meta_001500.json ADDED
@@ -0,0 +1,236 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ {
2
+ "step": 1500,
3
+ "training_complete": false,
4
+ "experiment_id": "Think.Unbounded-d32",
5
+ "parent_experiment_id": null,
6
+ "parent_checkpoint_step": null,
7
+ "val_bpb": 1.0152464516907762,
8
+ "model_config": {
9
+ "sequence_len": 4096,
10
+ "vocab_size": 32768,
11
+ "n_layer": 32,
12
+ "n_head": 16,
13
+ "n_kv_head": 16,
14
+ "n_embd": 2048,
15
+ "window_pattern": "SSSL"
16
+ },
17
+ "user_config": {
18
+ "run": "Think.Unbounded-d32",
19
+ "wandb_run_id": "9b75691b",
20
+ "wandb_group": "clean1930s-d32",
21
+ "wandb_tags": "think-dataset-clean-1930s,d32,ratio12,ctx4096,sssl,fp8,muon-momentum-constant,muon-momentum-0.90,midtrain-schedule,mix-og,0-30-60",
22
+ "device_type": "",
23
+ "fp8": true,
24
+ "fp8_recipe": "tensorwise",
25
+ "depth": 32,
26
+ "aspect_ratio": 64,
27
+ "head_dim": 128,
28
+ "max_seq_len": 4096,
29
+ "window_pattern": "SSSL",
30
+ "num_iterations": 9600,
31
+ "target_flops": -1.0,
32
+ "target_param_data_ratio": -1.0,
33
+ "device_batch_size": 2,
34
+ "total_batch_size": 2097152,
35
+ "embedding_lr": 0.3,
36
+ "unembedding_lr": 0.008,
37
+ "weight_decay": 0.28,
38
+ "matrix_lr": 0.02,
39
+ "muon_momentum": 0.9,
40
+ "scalar_lr": 0.5,
41
+ "warmup_steps": 40,
42
+ "warmdown_ratio": 0.65,
43
+ "final_lr_frac": 0.05,
44
+ "resume_from_step": -1,
45
+ "init_from_checkpoint_dir": null,
46
+ "init_from_step": -1,
47
+ "no_init_optimizer": false,
48
+ "branch_lr_schedule": "branch",
49
+ "branch_parent_experiment_id": null,
50
+ "pretokenized": true,
51
+ "data_dir": "/workspace/nanochat/experiments/Think.Unbounded-d32/data",
52
+ "tokenizer_dir": "/workspace/nanochat/experiments/Think.Unbounded-d32/tokenizer",
53
+ "pretokenized_dir": "/workspace/nanochat/experiments/Think.Unbounded-d32/pretok",
54
+ "mixture_source_dirs": "{\"original\": \"/workspace/nanochat/experiments/Think.Unbounded-d32/pretok_original\", \"midtrain_r30\": \"/workspace/nanochat/experiments/Think.Unbounded-d32/pretok_midtrain_r30\", \"midtrain_r60\": \"/workspace/nanochat/experiments/Think.Unbounded-d32/pretok_midtrain_r60\"}",
55
+ "checkpoint_dir": "/workspace/nanochat/experiments/Think.Unbounded-d32/base_checkpoints",
56
+ "experiment_id": "Think.Unbounded-d32",
57
+ "experiment_config": "/workspace/nanochat/experiments/Think.Unbounded-d32/config.json",
58
+ "tokenizer_fingerprint": "21e99d5cdeeaa660",
59
+ "git_commit_sha": "b2de514e00cde3bee362cf4c87e16d35e0f32c24",
60
+ "seed": 42,
61
+ "eval_every": 250,
62
+ "eval_tokens": 2097152,
63
+ "core_metric_every": -1,
64
+ "core_metric_max_per_task": 500,
65
+ "sample_every": -1,
66
+ "save_every": 500,
67
+ "model_tag": "Think.Unbounded-d32",
68
+ "experiment": {
69
+ "schema_version": 1,
70
+ "stage": "base",
71
+ "experiment_id": "Think.Unbounded-d32",
72
+ "datasets": {
73
+ "original": {
74
+ "adapter": "parquet_shards",
75
+ "repo": "jbduran/think-dataset-clean-1930s",
76
+ "revision": "45225d95bc15f942be3b4b344738cea1f66e3de8",
77
+ "validation_shard": 472,
78
+ "num_train_shards": 330,
79
+ "download_workers": 4
80
+ },
81
+ "midtrain_r30": {
82
+ "adapter": "parquet_shards",
83
+ "repo": "zachnorton03/think-midtrain",
84
+ "revision": "9ace24b8e16e57e38a1ea0b1f6d7cbf323bf9dbc",
85
+ "subfolder": "mixed/ratio_30/data",
86
+ "validation_shard": 1262,
87
+ "num_train_shards": 391,
88
+ "download_workers": 4
89
+ },
90
+ "midtrain_r60": {
91
+ "adapter": "parquet_shards",
92
+ "repo": "zachnorton03/think-midtrain",
93
+ "revision": "9ace24b8e16e57e38a1ea0b1f6d7cbf323bf9dbc",
94
+ "subfolder": "mixed/ratio_60/data",
95
+ "validation_shard": 523,
96
+ "num_train_shards": 207,
97
+ "download_workers": 4
98
+ }
99
+ },
100
+ "mixture_schedule": {
101
+ "total_tokens": 20132659200,
102
+ "seed_data": 1234,
103
+ "max_epochs": 4,
104
+ "stages": [
105
+ {
106
+ "name": "base",
107
+ "start_tokens": 0,
108
+ "source": "original"
109
+ },
110
+ {
111
+ "name": "injection",
112
+ "start_tokens": 14092861440,
113
+ "source": "midtrain_r30"
114
+ },
115
+ {
116
+ "name": "decay_mix",
117
+ "start_tokens": 18119393280,
118
+ "source": "midtrain_r60"
119
+ }
120
+ ]
121
+ },
122
+ "tokenizer": {
123
+ "mode": "reuse",
124
+ "source_experiment_id": "clean1930s-d24-r12-ctx4096-fulltok-v1",
125
+ "vocab_size": 32768
126
+ },
127
+ "pretokenize": {
128
+ "enabled": true,
129
+ "slack": 1.03,
130
+ "require_no_wrap": true,
131
+ "val_tokens": 20971520,
132
+ "shard_tokens": 100000000,
133
+ "tokenizer_threads": 8
134
+ },
135
+ "training": {
136
+ "depth": 32,
137
+ "scaling_params": 1677724672,
138
+ "target_param_data_ratio": 12,
139
+ "max_seq_len": 4096,
140
+ "window_pattern": "SSSL",
141
+ "muon_momentum": 0.9,
142
+ "device_batch_size": 2,
143
+ "total_batch_size": 2097152,
144
+ "fp8": true,
145
+ "fp8_recipe": "tensorwise",
146
+ "seed": 42,
147
+ "save_every": 500,
148
+ "eval_every": 250,
149
+ "eval_tokens": 2097152,
150
+ "core_metric_every": -1,
151
+ "sample_every": -1
152
+ },
153
+ "artifacts": {
154
+ "repo": "jbduran/think.nano",
155
+ "keep_local_checkpoints": 1
156
+ },
157
+ "wandb": {
158
+ "entity": "jbduran-thinkingmachinesncsu",
159
+ "project": "think.nano",
160
+ "name": "Think.Unbounded-d32",
161
+ "group": "clean1930s-d32",
162
+ "tags": [
163
+ "think-dataset-clean-1930s",
164
+ "d32",
165
+ "ratio12",
166
+ "ctx4096",
167
+ "sssl",
168
+ "fp8",
169
+ "muon-momentum-constant",
170
+ "muon-momentum-0.90",
171
+ "midtrain-schedule",
172
+ "mix-og",
173
+ "0-30-60"
174
+ ]
175
+ },
176
+ "config_fingerprint": "6f8ae461e0eb9ee2",
177
+ "artifact_path": "experiments/Think.Unbounded-d32"
178
+ },
179
+ "stage": "base",
180
+ "base_experiment_id": "Think.Unbounded-d32",
181
+ "parent_experiment_id": null,
182
+ "parent_checkpoint_step": null,
183
+ "config_fingerprint": "6f8ae461e0eb9ee2"
184
+ },
185
+ "device_batch_size": 2,
186
+ "max_seq_len": 4096,
187
+ "total_batch_size": 2097152,
188
+ "dataloader_state_dict": {
189
+ "file_idx": 31,
190
+ "pos": 46120193,
191
+ "epoch": 1,
192
+ "pq_idx": 31,
193
+ "rg_idx": 46120193,
194
+ "mixture": {
195
+ "cursors": {
196
+ "original": {
197
+ "file_idx": 31,
198
+ "pos": 46120193,
199
+ "epoch": 1,
200
+ "pq_idx": 31,
201
+ "rg_idx": 46120193
202
+ },
203
+ "midtrain_r30": {
204
+ "file_idx": 0,
205
+ "pos": 0,
206
+ "epoch": 1,
207
+ "pq_idx": 0,
208
+ "rg_idx": 0
209
+ },
210
+ "midtrain_r60": {
211
+ "file_idx": 0,
212
+ "pos": 0,
213
+ "epoch": 1,
214
+ "pq_idx": 0,
215
+ "rg_idx": 0
216
+ }
217
+ },
218
+ "cumulative_tokens": 3145736192,
219
+ "source_tokens": {
220
+ "original": 3145736192,
221
+ "midtrain_r30": 0,
222
+ "midtrain_r60": 0
223
+ },
224
+ "active_stage_idx": 0
225
+ }
226
+ },
227
+ "loop_state": {
228
+ "min_val_bpb": 1.0152464516907762,
229
+ "smooth_train_loss": 2.935989382480182,
230
+ "total_training_time": 59860.37234377861,
231
+ "stage_start_step": 0,
232
+ "stage_training_flops": 36099223745200128000,
233
+ "inherited_parent_flops": 0.0,
234
+ "cumulative_pipeline_training_flops": 3.609922374520013e+19
235
+ }
236
+ }
experiments/Think.Unbounded-d32/base_checkpoints/meta_002000.json ADDED
@@ -0,0 +1,236 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ {
2
+ "step": 2000,
3
+ "training_complete": false,
4
+ "experiment_id": "Think.Unbounded-d32",
5
+ "parent_experiment_id": null,
6
+ "parent_checkpoint_step": null,
7
+ "val_bpb": 0.9928518912323213,
8
+ "model_config": {
9
+ "sequence_len": 4096,
10
+ "vocab_size": 32768,
11
+ "n_layer": 32,
12
+ "n_head": 16,
13
+ "n_kv_head": 16,
14
+ "n_embd": 2048,
15
+ "window_pattern": "SSSL"
16
+ },
17
+ "user_config": {
18
+ "run": "Think.Unbounded-d32",
19
+ "wandb_run_id": "9b75691b",
20
+ "wandb_group": "clean1930s-d32",
21
+ "wandb_tags": "think-dataset-clean-1930s,d32,ratio12,ctx4096,sssl,fp8,muon-momentum-constant,muon-momentum-0.90,midtrain-schedule,mix-og,0-30-60",
22
+ "device_type": "",
23
+ "fp8": true,
24
+ "fp8_recipe": "tensorwise",
25
+ "depth": 32,
26
+ "aspect_ratio": 64,
27
+ "head_dim": 128,
28
+ "max_seq_len": 4096,
29
+ "window_pattern": "SSSL",
30
+ "num_iterations": 9600,
31
+ "target_flops": -1.0,
32
+ "target_param_data_ratio": -1.0,
33
+ "device_batch_size": 2,
34
+ "total_batch_size": 2097152,
35
+ "embedding_lr": 0.3,
36
+ "unembedding_lr": 0.008,
37
+ "weight_decay": 0.28,
38
+ "matrix_lr": 0.02,
39
+ "muon_momentum": 0.9,
40
+ "scalar_lr": 0.5,
41
+ "warmup_steps": 40,
42
+ "warmdown_ratio": 0.65,
43
+ "final_lr_frac": 0.05,
44
+ "resume_from_step": -1,
45
+ "init_from_checkpoint_dir": null,
46
+ "init_from_step": -1,
47
+ "no_init_optimizer": false,
48
+ "branch_lr_schedule": "branch",
49
+ "branch_parent_experiment_id": null,
50
+ "pretokenized": true,
51
+ "data_dir": "/workspace/nanochat/experiments/Think.Unbounded-d32/data",
52
+ "tokenizer_dir": "/workspace/nanochat/experiments/Think.Unbounded-d32/tokenizer",
53
+ "pretokenized_dir": "/workspace/nanochat/experiments/Think.Unbounded-d32/pretok",
54
+ "mixture_source_dirs": "{\"original\": \"/workspace/nanochat/experiments/Think.Unbounded-d32/pretok_original\", \"midtrain_r30\": \"/workspace/nanochat/experiments/Think.Unbounded-d32/pretok_midtrain_r30\", \"midtrain_r60\": \"/workspace/nanochat/experiments/Think.Unbounded-d32/pretok_midtrain_r60\"}",
55
+ "checkpoint_dir": "/workspace/nanochat/experiments/Think.Unbounded-d32/base_checkpoints",
56
+ "experiment_id": "Think.Unbounded-d32",
57
+ "experiment_config": "/workspace/nanochat/experiments/Think.Unbounded-d32/config.json",
58
+ "tokenizer_fingerprint": "21e99d5cdeeaa660",
59
+ "git_commit_sha": "b2de514e00cde3bee362cf4c87e16d35e0f32c24",
60
+ "seed": 42,
61
+ "eval_every": 250,
62
+ "eval_tokens": 2097152,
63
+ "core_metric_every": -1,
64
+ "core_metric_max_per_task": 500,
65
+ "sample_every": -1,
66
+ "save_every": 500,
67
+ "model_tag": "Think.Unbounded-d32",
68
+ "experiment": {
69
+ "schema_version": 1,
70
+ "stage": "base",
71
+ "experiment_id": "Think.Unbounded-d32",
72
+ "datasets": {
73
+ "original": {
74
+ "adapter": "parquet_shards",
75
+ "repo": "jbduran/think-dataset-clean-1930s",
76
+ "revision": "45225d95bc15f942be3b4b344738cea1f66e3de8",
77
+ "validation_shard": 472,
78
+ "num_train_shards": 330,
79
+ "download_workers": 4
80
+ },
81
+ "midtrain_r30": {
82
+ "adapter": "parquet_shards",
83
+ "repo": "zachnorton03/think-midtrain",
84
+ "revision": "9ace24b8e16e57e38a1ea0b1f6d7cbf323bf9dbc",
85
+ "subfolder": "mixed/ratio_30/data",
86
+ "validation_shard": 1262,
87
+ "num_train_shards": 391,
88
+ "download_workers": 4
89
+ },
90
+ "midtrain_r60": {
91
+ "adapter": "parquet_shards",
92
+ "repo": "zachnorton03/think-midtrain",
93
+ "revision": "9ace24b8e16e57e38a1ea0b1f6d7cbf323bf9dbc",
94
+ "subfolder": "mixed/ratio_60/data",
95
+ "validation_shard": 523,
96
+ "num_train_shards": 207,
97
+ "download_workers": 4
98
+ }
99
+ },
100
+ "mixture_schedule": {
101
+ "total_tokens": 20132659200,
102
+ "seed_data": 1234,
103
+ "max_epochs": 4,
104
+ "stages": [
105
+ {
106
+ "name": "base",
107
+ "start_tokens": 0,
108
+ "source": "original"
109
+ },
110
+ {
111
+ "name": "injection",
112
+ "start_tokens": 14092861440,
113
+ "source": "midtrain_r30"
114
+ },
115
+ {
116
+ "name": "decay_mix",
117
+ "start_tokens": 18119393280,
118
+ "source": "midtrain_r60"
119
+ }
120
+ ]
121
+ },
122
+ "tokenizer": {
123
+ "mode": "reuse",
124
+ "source_experiment_id": "clean1930s-d24-r12-ctx4096-fulltok-v1",
125
+ "vocab_size": 32768
126
+ },
127
+ "pretokenize": {
128
+ "enabled": true,
129
+ "slack": 1.03,
130
+ "require_no_wrap": true,
131
+ "val_tokens": 20971520,
132
+ "shard_tokens": 100000000,
133
+ "tokenizer_threads": 8
134
+ },
135
+ "training": {
136
+ "depth": 32,
137
+ "scaling_params": 1677724672,
138
+ "target_param_data_ratio": 12,
139
+ "max_seq_len": 4096,
140
+ "window_pattern": "SSSL",
141
+ "muon_momentum": 0.9,
142
+ "device_batch_size": 2,
143
+ "total_batch_size": 2097152,
144
+ "fp8": true,
145
+ "fp8_recipe": "tensorwise",
146
+ "seed": 42,
147
+ "save_every": 500,
148
+ "eval_every": 250,
149
+ "eval_tokens": 2097152,
150
+ "core_metric_every": -1,
151
+ "sample_every": -1
152
+ },
153
+ "artifacts": {
154
+ "repo": "jbduran/think.nano",
155
+ "keep_local_checkpoints": 1
156
+ },
157
+ "wandb": {
158
+ "entity": "jbduran-thinkingmachinesncsu",
159
+ "project": "think.nano",
160
+ "name": "Think.Unbounded-d32",
161
+ "group": "clean1930s-d32",
162
+ "tags": [
163
+ "think-dataset-clean-1930s",
164
+ "d32",
165
+ "ratio12",
166
+ "ctx4096",
167
+ "sssl",
168
+ "fp8",
169
+ "muon-momentum-constant",
170
+ "muon-momentum-0.90",
171
+ "midtrain-schedule",
172
+ "mix-og",
173
+ "0-30-60"
174
+ ]
175
+ },
176
+ "config_fingerprint": "6f8ae461e0eb9ee2",
177
+ "artifact_path": "experiments/Think.Unbounded-d32"
178
+ },
179
+ "stage": "base",
180
+ "base_experiment_id": "Think.Unbounded-d32",
181
+ "parent_experiment_id": null,
182
+ "parent_checkpoint_step": null,
183
+ "config_fingerprint": "6f8ae461e0eb9ee2"
184
+ },
185
+ "device_batch_size": 2,
186
+ "max_seq_len": 4096,
187
+ "total_batch_size": 2097152,
188
+ "dataloader_state_dict": {
189
+ "file_idx": 41,
190
+ "pos": 94824193,
191
+ "epoch": 1,
192
+ "pq_idx": 41,
193
+ "rg_idx": 94824193,
194
+ "mixture": {
195
+ "cursors": {
196
+ "original": {
197
+ "file_idx": 41,
198
+ "pos": 94824193,
199
+ "epoch": 1,
200
+ "pq_idx": 41,
201
+ "rg_idx": 94824193
202
+ },
203
+ "midtrain_r30": {
204
+ "file_idx": 0,
205
+ "pos": 0,
206
+ "epoch": 1,
207
+ "pq_idx": 0,
208
+ "rg_idx": 0
209
+ },
210
+ "midtrain_r60": {
211
+ "file_idx": 0,
212
+ "pos": 0,
213
+ "epoch": 1,
214
+ "pq_idx": 0,
215
+ "rg_idx": 0
216
+ }
217
+ },
218
+ "cumulative_tokens": 4194312192,
219
+ "source_tokens": {
220
+ "original": 4194312192,
221
+ "midtrain_r30": 0,
222
+ "midtrain_r60": 0
223
+ },
224
+ "active_stage_idx": 0
225
+ }
226
+ },
227
+ "loop_state": {
228
+ "min_val_bpb": 0.9928518912323213,
229
+ "smooth_train_loss": 2.715263738468478,
230
+ "total_training_time": 79915.32717370987,
231
+ "stage_start_step": 0,
232
+ "stage_training_flops": 48132298326933504000,
233
+ "inherited_parent_flops": 0.0,
234
+ "cumulative_pipeline_training_flops": 4.81322983269335e+19
235
+ }
236
+ }
experiments/Think.Unbounded-d32/base_checkpoints/meta_002500.json ADDED
@@ -0,0 +1,236 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ {
2
+ "step": 2500,
3
+ "training_complete": false,
4
+ "experiment_id": "Think.Unbounded-d32",
5
+ "parent_experiment_id": null,
6
+ "parent_checkpoint_step": null,
7
+ "val_bpb": 0.9737910804428661,
8
+ "model_config": {
9
+ "sequence_len": 4096,
10
+ "vocab_size": 32768,
11
+ "n_layer": 32,
12
+ "n_head": 16,
13
+ "n_kv_head": 16,
14
+ "n_embd": 2048,
15
+ "window_pattern": "SSSL"
16
+ },
17
+ "user_config": {
18
+ "run": "Think.Unbounded-d32",
19
+ "wandb_run_id": "9b75691b",
20
+ "wandb_group": "clean1930s-d32",
21
+ "wandb_tags": "think-dataset-clean-1930s,d32,ratio12,ctx4096,sssl,fp8,muon-momentum-constant,muon-momentum-0.90,midtrain-schedule,mix-og,0-30-60",
22
+ "device_type": "",
23
+ "fp8": true,
24
+ "fp8_recipe": "tensorwise",
25
+ "depth": 32,
26
+ "aspect_ratio": 64,
27
+ "head_dim": 128,
28
+ "max_seq_len": 4096,
29
+ "window_pattern": "SSSL",
30
+ "num_iterations": 9600,
31
+ "target_flops": -1.0,
32
+ "target_param_data_ratio": -1.0,
33
+ "device_batch_size": 2,
34
+ "total_batch_size": 2097152,
35
+ "embedding_lr": 0.3,
36
+ "unembedding_lr": 0.008,
37
+ "weight_decay": 0.28,
38
+ "matrix_lr": 0.02,
39
+ "muon_momentum": 0.9,
40
+ "scalar_lr": 0.5,
41
+ "warmup_steps": 40,
42
+ "warmdown_ratio": 0.65,
43
+ "final_lr_frac": 0.05,
44
+ "resume_from_step": -1,
45
+ "init_from_checkpoint_dir": null,
46
+ "init_from_step": -1,
47
+ "no_init_optimizer": false,
48
+ "branch_lr_schedule": "branch",
49
+ "branch_parent_experiment_id": null,
50
+ "pretokenized": true,
51
+ "data_dir": "/workspace/nanochat/experiments/Think.Unbounded-d32/data",
52
+ "tokenizer_dir": "/workspace/nanochat/experiments/Think.Unbounded-d32/tokenizer",
53
+ "pretokenized_dir": "/workspace/nanochat/experiments/Think.Unbounded-d32/pretok",
54
+ "mixture_source_dirs": "{\"original\": \"/workspace/nanochat/experiments/Think.Unbounded-d32/pretok_original\", \"midtrain_r30\": \"/workspace/nanochat/experiments/Think.Unbounded-d32/pretok_midtrain_r30\", \"midtrain_r60\": \"/workspace/nanochat/experiments/Think.Unbounded-d32/pretok_midtrain_r60\"}",
55
+ "checkpoint_dir": "/workspace/nanochat/experiments/Think.Unbounded-d32/base_checkpoints",
56
+ "experiment_id": "Think.Unbounded-d32",
57
+ "experiment_config": "/workspace/nanochat/experiments/Think.Unbounded-d32/config.json",
58
+ "tokenizer_fingerprint": "21e99d5cdeeaa660",
59
+ "git_commit_sha": "b2de514e00cde3bee362cf4c87e16d35e0f32c24",
60
+ "seed": 42,
61
+ "eval_every": 250,
62
+ "eval_tokens": 2097152,
63
+ "core_metric_every": -1,
64
+ "core_metric_max_per_task": 500,
65
+ "sample_every": -1,
66
+ "save_every": 500,
67
+ "model_tag": "Think.Unbounded-d32",
68
+ "experiment": {
69
+ "schema_version": 1,
70
+ "stage": "base",
71
+ "experiment_id": "Think.Unbounded-d32",
72
+ "datasets": {
73
+ "original": {
74
+ "adapter": "parquet_shards",
75
+ "repo": "jbduran/think-dataset-clean-1930s",
76
+ "revision": "45225d95bc15f942be3b4b344738cea1f66e3de8",
77
+ "validation_shard": 472,
78
+ "num_train_shards": 330,
79
+ "download_workers": 4
80
+ },
81
+ "midtrain_r30": {
82
+ "adapter": "parquet_shards",
83
+ "repo": "zachnorton03/think-midtrain",
84
+ "revision": "9ace24b8e16e57e38a1ea0b1f6d7cbf323bf9dbc",
85
+ "subfolder": "mixed/ratio_30/data",
86
+ "validation_shard": 1262,
87
+ "num_train_shards": 391,
88
+ "download_workers": 4
89
+ },
90
+ "midtrain_r60": {
91
+ "adapter": "parquet_shards",
92
+ "repo": "zachnorton03/think-midtrain",
93
+ "revision": "9ace24b8e16e57e38a1ea0b1f6d7cbf323bf9dbc",
94
+ "subfolder": "mixed/ratio_60/data",
95
+ "validation_shard": 523,
96
+ "num_train_shards": 207,
97
+ "download_workers": 4
98
+ }
99
+ },
100
+ "mixture_schedule": {
101
+ "total_tokens": 20132659200,
102
+ "seed_data": 1234,
103
+ "max_epochs": 4,
104
+ "stages": [
105
+ {
106
+ "name": "base",
107
+ "start_tokens": 0,
108
+ "source": "original"
109
+ },
110
+ {
111
+ "name": "injection",
112
+ "start_tokens": 14092861440,
113
+ "source": "midtrain_r30"
114
+ },
115
+ {
116
+ "name": "decay_mix",
117
+ "start_tokens": 18119393280,
118
+ "source": "midtrain_r60"
119
+ }
120
+ ]
121
+ },
122
+ "tokenizer": {
123
+ "mode": "reuse",
124
+ "source_experiment_id": "clean1930s-d24-r12-ctx4096-fulltok-v1",
125
+ "vocab_size": 32768
126
+ },
127
+ "pretokenize": {
128
+ "enabled": true,
129
+ "slack": 1.03,
130
+ "require_no_wrap": true,
131
+ "val_tokens": 20971520,
132
+ "shard_tokens": 100000000,
133
+ "tokenizer_threads": 8
134
+ },
135
+ "training": {
136
+ "depth": 32,
137
+ "scaling_params": 1677724672,
138
+ "target_param_data_ratio": 12,
139
+ "max_seq_len": 4096,
140
+ "window_pattern": "SSSL",
141
+ "muon_momentum": 0.9,
142
+ "device_batch_size": 2,
143
+ "total_batch_size": 2097152,
144
+ "fp8": true,
145
+ "fp8_recipe": "tensorwise",
146
+ "seed": 42,
147
+ "save_every": 500,
148
+ "eval_every": 250,
149
+ "eval_tokens": 2097152,
150
+ "core_metric_every": -1,
151
+ "sample_every": -1
152
+ },
153
+ "artifacts": {
154
+ "repo": "jbduran/think.nano",
155
+ "keep_local_checkpoints": 1
156
+ },
157
+ "wandb": {
158
+ "entity": "jbduran-thinkingmachinesncsu",
159
+ "project": "think.nano",
160
+ "name": "Think.Unbounded-d32",
161
+ "group": "clean1930s-d32",
162
+ "tags": [
163
+ "think-dataset-clean-1930s",
164
+ "d32",
165
+ "ratio12",
166
+ "ctx4096",
167
+ "sssl",
168
+ "fp8",
169
+ "muon-momentum-constant",
170
+ "muon-momentum-0.90",
171
+ "midtrain-schedule",
172
+ "mix-og",
173
+ "0-30-60"
174
+ ]
175
+ },
176
+ "config_fingerprint": "6f8ae461e0eb9ee2",
177
+ "artifact_path": "experiments/Think.Unbounded-d32"
178
+ },
179
+ "stage": "base",
180
+ "base_experiment_id": "Think.Unbounded-d32",
181
+ "parent_experiment_id": null,
182
+ "parent_checkpoint_step": null,
183
+ "config_fingerprint": "6f8ae461e0eb9ee2"
184
+ },
185
+ "device_batch_size": 2,
186
+ "max_seq_len": 4096,
187
+ "total_batch_size": 2097152,
188
+ "dataloader_state_dict": {
189
+ "file_idx": 52,
190
+ "pos": 43528193,
191
+ "epoch": 1,
192
+ "pq_idx": 52,
193
+ "rg_idx": 43528193,
194
+ "mixture": {
195
+ "cursors": {
196
+ "original": {
197
+ "file_idx": 52,
198
+ "pos": 43528193,
199
+ "epoch": 1,
200
+ "pq_idx": 52,
201
+ "rg_idx": 43528193
202
+ },
203
+ "midtrain_r30": {
204
+ "file_idx": 0,
205
+ "pos": 0,
206
+ "epoch": 1,
207
+ "pq_idx": 0,
208
+ "rg_idx": 0
209
+ },
210
+ "midtrain_r60": {
211
+ "file_idx": 0,
212
+ "pos": 0,
213
+ "epoch": 1,
214
+ "pq_idx": 0,
215
+ "rg_idx": 0
216
+ }
217
+ },
218
+ "cumulative_tokens": 5242888192,
219
+ "source_tokens": {
220
+ "original": 5242888192,
221
+ "midtrain_r30": 0,
222
+ "midtrain_r60": 0
223
+ },
224
+ "active_stage_idx": 0
225
+ }
226
+ },
227
+ "loop_state": {
228
+ "min_val_bpb": 0.9630578341851044,
229
+ "smooth_train_loss": 2.7360114404103504,
230
+ "total_training_time": 99964.06782126427,
231
+ "stage_start_step": 0,
232
+ "stage_training_flops": 60165372908666880000,
233
+ "inherited_parent_flops": 0.0,
234
+ "cumulative_pipeline_training_flops": 6.016537290866688e+19
235
+ }
236
+ }
experiments/Think.Unbounded-d32/base_checkpoints/meta_003000.json ADDED
@@ -0,0 +1,236 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ {
2
+ "step": 3000,
3
+ "training_complete": false,
4
+ "experiment_id": "Think.Unbounded-d32",
5
+ "parent_experiment_id": null,
6
+ "parent_checkpoint_step": null,
7
+ "val_bpb": 0.9735215125930343,
8
+ "model_config": {
9
+ "sequence_len": 4096,
10
+ "vocab_size": 32768,
11
+ "n_layer": 32,
12
+ "n_head": 16,
13
+ "n_kv_head": 16,
14
+ "n_embd": 2048,
15
+ "window_pattern": "SSSL"
16
+ },
17
+ "user_config": {
18
+ "run": "Think.Unbounded-d32",
19
+ "wandb_run_id": "9b75691b",
20
+ "wandb_group": "clean1930s-d32",
21
+ "wandb_tags": "think-dataset-clean-1930s,d32,ratio12,ctx4096,sssl,fp8,muon-momentum-constant,muon-momentum-0.90,midtrain-schedule,mix-og,0-30-60",
22
+ "device_type": "",
23
+ "fp8": true,
24
+ "fp8_recipe": "tensorwise",
25
+ "depth": 32,
26
+ "aspect_ratio": 64,
27
+ "head_dim": 128,
28
+ "max_seq_len": 4096,
29
+ "window_pattern": "SSSL",
30
+ "num_iterations": 9600,
31
+ "target_flops": -1.0,
32
+ "target_param_data_ratio": -1.0,
33
+ "device_batch_size": 2,
34
+ "total_batch_size": 2097152,
35
+ "embedding_lr": 0.3,
36
+ "unembedding_lr": 0.008,
37
+ "weight_decay": 0.28,
38
+ "matrix_lr": 0.02,
39
+ "muon_momentum": 0.9,
40
+ "scalar_lr": 0.5,
41
+ "warmup_steps": 40,
42
+ "warmdown_ratio": 0.65,
43
+ "final_lr_frac": 0.05,
44
+ "resume_from_step": -1,
45
+ "init_from_checkpoint_dir": null,
46
+ "init_from_step": -1,
47
+ "no_init_optimizer": false,
48
+ "branch_lr_schedule": "branch",
49
+ "branch_parent_experiment_id": null,
50
+ "pretokenized": true,
51
+ "data_dir": "/workspace/nanochat/experiments/Think.Unbounded-d32/data",
52
+ "tokenizer_dir": "/workspace/nanochat/experiments/Think.Unbounded-d32/tokenizer",
53
+ "pretokenized_dir": "/workspace/nanochat/experiments/Think.Unbounded-d32/pretok",
54
+ "mixture_source_dirs": "{\"original\": \"/workspace/nanochat/experiments/Think.Unbounded-d32/pretok_original\", \"midtrain_r30\": \"/workspace/nanochat/experiments/Think.Unbounded-d32/pretok_midtrain_r30\", \"midtrain_r60\": \"/workspace/nanochat/experiments/Think.Unbounded-d32/pretok_midtrain_r60\"}",
55
+ "checkpoint_dir": "/workspace/nanochat/experiments/Think.Unbounded-d32/base_checkpoints",
56
+ "experiment_id": "Think.Unbounded-d32",
57
+ "experiment_config": "/workspace/nanochat/experiments/Think.Unbounded-d32/config.json",
58
+ "tokenizer_fingerprint": "21e99d5cdeeaa660",
59
+ "git_commit_sha": "b2de514e00cde3bee362cf4c87e16d35e0f32c24",
60
+ "seed": 42,
61
+ "eval_every": 250,
62
+ "eval_tokens": 2097152,
63
+ "core_metric_every": -1,
64
+ "core_metric_max_per_task": 500,
65
+ "sample_every": -1,
66
+ "save_every": 500,
67
+ "model_tag": "Think.Unbounded-d32",
68
+ "experiment": {
69
+ "schema_version": 1,
70
+ "stage": "base",
71
+ "experiment_id": "Think.Unbounded-d32",
72
+ "datasets": {
73
+ "original": {
74
+ "adapter": "parquet_shards",
75
+ "repo": "jbduran/think-dataset-clean-1930s",
76
+ "revision": "45225d95bc15f942be3b4b344738cea1f66e3de8",
77
+ "validation_shard": 472,
78
+ "num_train_shards": 330,
79
+ "download_workers": 4
80
+ },
81
+ "midtrain_r30": {
82
+ "adapter": "parquet_shards",
83
+ "repo": "zachnorton03/think-midtrain",
84
+ "revision": "9ace24b8e16e57e38a1ea0b1f6d7cbf323bf9dbc",
85
+ "subfolder": "mixed/ratio_30/data",
86
+ "validation_shard": 1262,
87
+ "num_train_shards": 391,
88
+ "download_workers": 4
89
+ },
90
+ "midtrain_r60": {
91
+ "adapter": "parquet_shards",
92
+ "repo": "zachnorton03/think-midtrain",
93
+ "revision": "9ace24b8e16e57e38a1ea0b1f6d7cbf323bf9dbc",
94
+ "subfolder": "mixed/ratio_60/data",
95
+ "validation_shard": 523,
96
+ "num_train_shards": 207,
97
+ "download_workers": 4
98
+ }
99
+ },
100
+ "mixture_schedule": {
101
+ "total_tokens": 20132659200,
102
+ "seed_data": 1234,
103
+ "max_epochs": 4,
104
+ "stages": [
105
+ {
106
+ "name": "base",
107
+ "start_tokens": 0,
108
+ "source": "original"
109
+ },
110
+ {
111
+ "name": "injection",
112
+ "start_tokens": 14092861440,
113
+ "source": "midtrain_r30"
114
+ },
115
+ {
116
+ "name": "decay_mix",
117
+ "start_tokens": 18119393280,
118
+ "source": "midtrain_r60"
119
+ }
120
+ ]
121
+ },
122
+ "tokenizer": {
123
+ "mode": "reuse",
124
+ "source_experiment_id": "clean1930s-d24-r12-ctx4096-fulltok-v1",
125
+ "vocab_size": 32768
126
+ },
127
+ "pretokenize": {
128
+ "enabled": true,
129
+ "slack": 1.03,
130
+ "require_no_wrap": true,
131
+ "val_tokens": 20971520,
132
+ "shard_tokens": 100000000,
133
+ "tokenizer_threads": 8
134
+ },
135
+ "training": {
136
+ "depth": 32,
137
+ "scaling_params": 1677724672,
138
+ "target_param_data_ratio": 12,
139
+ "max_seq_len": 4096,
140
+ "window_pattern": "SSSL",
141
+ "muon_momentum": 0.9,
142
+ "device_batch_size": 2,
143
+ "total_batch_size": 2097152,
144
+ "fp8": true,
145
+ "fp8_recipe": "tensorwise",
146
+ "seed": 42,
147
+ "save_every": 500,
148
+ "eval_every": 250,
149
+ "eval_tokens": 2097152,
150
+ "core_metric_every": -1,
151
+ "sample_every": -1
152
+ },
153
+ "artifacts": {
154
+ "repo": "jbduran/think.nano",
155
+ "keep_local_checkpoints": 1
156
+ },
157
+ "wandb": {
158
+ "entity": "jbduran-thinkingmachinesncsu",
159
+ "project": "think.nano",
160
+ "name": "Think.Unbounded-d32",
161
+ "group": "clean1930s-d32",
162
+ "tags": [
163
+ "think-dataset-clean-1930s",
164
+ "d32",
165
+ "ratio12",
166
+ "ctx4096",
167
+ "sssl",
168
+ "fp8",
169
+ "muon-momentum-constant",
170
+ "muon-momentum-0.90",
171
+ "midtrain-schedule",
172
+ "mix-og",
173
+ "0-30-60"
174
+ ]
175
+ },
176
+ "config_fingerprint": "6f8ae461e0eb9ee2",
177
+ "artifact_path": "experiments/Think.Unbounded-d32"
178
+ },
179
+ "stage": "base",
180
+ "base_experiment_id": "Think.Unbounded-d32",
181
+ "parent_experiment_id": null,
182
+ "parent_checkpoint_step": null,
183
+ "config_fingerprint": "6f8ae461e0eb9ee2"
184
+ },
185
+ "device_batch_size": 2,
186
+ "max_seq_len": 4096,
187
+ "total_batch_size": 2097152,
188
+ "dataloader_state_dict": {
189
+ "file_idx": 62,
190
+ "pos": 92232193,
191
+ "epoch": 1,
192
+ "pq_idx": 62,
193
+ "rg_idx": 92232193,
194
+ "mixture": {
195
+ "cursors": {
196
+ "original": {
197
+ "file_idx": 62,
198
+ "pos": 92232193,
199
+ "epoch": 1,
200
+ "pq_idx": 62,
201
+ "rg_idx": 92232193
202
+ },
203
+ "midtrain_r30": {
204
+ "file_idx": 0,
205
+ "pos": 0,
206
+ "epoch": 1,
207
+ "pq_idx": 0,
208
+ "rg_idx": 0
209
+ },
210
+ "midtrain_r60": {
211
+ "file_idx": 0,
212
+ "pos": 0,
213
+ "epoch": 1,
214
+ "pq_idx": 0,
215
+ "rg_idx": 0
216
+ }
217
+ },
218
+ "cumulative_tokens": 6291464192,
219
+ "source_tokens": {
220
+ "original": 6291464192,
221
+ "midtrain_r30": 0,
222
+ "midtrain_r60": 0
223
+ },
224
+ "active_stage_idx": 0
225
+ }
226
+ },
227
+ "loop_state": {
228
+ "min_val_bpb": 0.9630578341851044,
229
+ "smooth_train_loss": 2.6967057059495345,
230
+ "total_training_time": 120021.50824666023,
231
+ "stage_start_step": 0,
232
+ "stage_training_flops": 72198447490400256000,
233
+ "inherited_parent_flops": 0.0,
234
+ "cumulative_pipeline_training_flops": 7.219844749040026e+19
235
+ }
236
+ }
experiments/Think.Unbounded-d32/base_checkpoints/meta_003500.json ADDED
@@ -0,0 +1,236 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ {
2
+ "step": 3500,
3
+ "training_complete": false,
4
+ "experiment_id": "Think.Unbounded-d32",
5
+ "parent_experiment_id": null,
6
+ "parent_checkpoint_step": null,
7
+ "val_bpb": 0.968207957699399,
8
+ "model_config": {
9
+ "sequence_len": 4096,
10
+ "vocab_size": 32768,
11
+ "n_layer": 32,
12
+ "n_head": 16,
13
+ "n_kv_head": 16,
14
+ "n_embd": 2048,
15
+ "window_pattern": "SSSL"
16
+ },
17
+ "user_config": {
18
+ "run": "Think.Unbounded-d32",
19
+ "wandb_run_id": "9b75691b",
20
+ "wandb_group": "clean1930s-d32",
21
+ "wandb_tags": "think-dataset-clean-1930s,d32,ratio12,ctx4096,sssl,fp8,muon-momentum-constant,muon-momentum-0.90,midtrain-schedule,mix-og,0-30-60",
22
+ "device_type": "",
23
+ "fp8": true,
24
+ "fp8_recipe": "tensorwise",
25
+ "depth": 32,
26
+ "aspect_ratio": 64,
27
+ "head_dim": 128,
28
+ "max_seq_len": 4096,
29
+ "window_pattern": "SSSL",
30
+ "num_iterations": 9600,
31
+ "target_flops": -1.0,
32
+ "target_param_data_ratio": -1.0,
33
+ "device_batch_size": 2,
34
+ "total_batch_size": 2097152,
35
+ "embedding_lr": 0.3,
36
+ "unembedding_lr": 0.008,
37
+ "weight_decay": 0.28,
38
+ "matrix_lr": 0.02,
39
+ "muon_momentum": 0.9,
40
+ "scalar_lr": 0.5,
41
+ "warmup_steps": 40,
42
+ "warmdown_ratio": 0.65,
43
+ "final_lr_frac": 0.05,
44
+ "resume_from_step": -1,
45
+ "init_from_checkpoint_dir": null,
46
+ "init_from_step": -1,
47
+ "no_init_optimizer": false,
48
+ "branch_lr_schedule": "branch",
49
+ "branch_parent_experiment_id": null,
50
+ "pretokenized": true,
51
+ "data_dir": "/workspace/nanochat/experiments/Think.Unbounded-d32/data",
52
+ "tokenizer_dir": "/workspace/nanochat/experiments/Think.Unbounded-d32/tokenizer",
53
+ "pretokenized_dir": "/workspace/nanochat/experiments/Think.Unbounded-d32/pretok",
54
+ "mixture_source_dirs": "{\"original\": \"/workspace/nanochat/experiments/Think.Unbounded-d32/pretok_original\", \"midtrain_r30\": \"/workspace/nanochat/experiments/Think.Unbounded-d32/pretok_midtrain_r30\", \"midtrain_r60\": \"/workspace/nanochat/experiments/Think.Unbounded-d32/pretok_midtrain_r60\"}",
55
+ "checkpoint_dir": "/workspace/nanochat/experiments/Think.Unbounded-d32/base_checkpoints",
56
+ "experiment_id": "Think.Unbounded-d32",
57
+ "experiment_config": "/workspace/nanochat/experiments/Think.Unbounded-d32/config.json",
58
+ "tokenizer_fingerprint": "21e99d5cdeeaa660",
59
+ "git_commit_sha": "b2de514e00cde3bee362cf4c87e16d35e0f32c24",
60
+ "seed": 42,
61
+ "eval_every": 250,
62
+ "eval_tokens": 2097152,
63
+ "core_metric_every": -1,
64
+ "core_metric_max_per_task": 500,
65
+ "sample_every": -1,
66
+ "save_every": 500,
67
+ "model_tag": "Think.Unbounded-d32",
68
+ "experiment": {
69
+ "schema_version": 1,
70
+ "stage": "base",
71
+ "experiment_id": "Think.Unbounded-d32",
72
+ "datasets": {
73
+ "original": {
74
+ "adapter": "parquet_shards",
75
+ "repo": "jbduran/think-dataset-clean-1930s",
76
+ "revision": "45225d95bc15f942be3b4b344738cea1f66e3de8",
77
+ "validation_shard": 472,
78
+ "num_train_shards": 330,
79
+ "download_workers": 4
80
+ },
81
+ "midtrain_r30": {
82
+ "adapter": "parquet_shards",
83
+ "repo": "zachnorton03/think-midtrain",
84
+ "revision": "9ace24b8e16e57e38a1ea0b1f6d7cbf323bf9dbc",
85
+ "subfolder": "mixed/ratio_30/data",
86
+ "validation_shard": 1262,
87
+ "num_train_shards": 391,
88
+ "download_workers": 4
89
+ },
90
+ "midtrain_r60": {
91
+ "adapter": "parquet_shards",
92
+ "repo": "zachnorton03/think-midtrain",
93
+ "revision": "9ace24b8e16e57e38a1ea0b1f6d7cbf323bf9dbc",
94
+ "subfolder": "mixed/ratio_60/data",
95
+ "validation_shard": 523,
96
+ "num_train_shards": 207,
97
+ "download_workers": 4
98
+ }
99
+ },
100
+ "mixture_schedule": {
101
+ "total_tokens": 20132659200,
102
+ "seed_data": 1234,
103
+ "max_epochs": 4,
104
+ "stages": [
105
+ {
106
+ "name": "base",
107
+ "start_tokens": 0,
108
+ "source": "original"
109
+ },
110
+ {
111
+ "name": "injection",
112
+ "start_tokens": 14092861440,
113
+ "source": "midtrain_r30"
114
+ },
115
+ {
116
+ "name": "decay_mix",
117
+ "start_tokens": 18119393280,
118
+ "source": "midtrain_r60"
119
+ }
120
+ ]
121
+ },
122
+ "tokenizer": {
123
+ "mode": "reuse",
124
+ "source_experiment_id": "clean1930s-d24-r12-ctx4096-fulltok-v1",
125
+ "vocab_size": 32768
126
+ },
127
+ "pretokenize": {
128
+ "enabled": true,
129
+ "slack": 1.03,
130
+ "require_no_wrap": true,
131
+ "val_tokens": 20971520,
132
+ "shard_tokens": 100000000,
133
+ "tokenizer_threads": 8
134
+ },
135
+ "training": {
136
+ "depth": 32,
137
+ "scaling_params": 1677724672,
138
+ "target_param_data_ratio": 12,
139
+ "max_seq_len": 4096,
140
+ "window_pattern": "SSSL",
141
+ "muon_momentum": 0.9,
142
+ "device_batch_size": 2,
143
+ "total_batch_size": 2097152,
144
+ "fp8": true,
145
+ "fp8_recipe": "tensorwise",
146
+ "seed": 42,
147
+ "save_every": 500,
148
+ "eval_every": 250,
149
+ "eval_tokens": 2097152,
150
+ "core_metric_every": -1,
151
+ "sample_every": -1
152
+ },
153
+ "artifacts": {
154
+ "repo": "jbduran/think.nano",
155
+ "keep_local_checkpoints": 1
156
+ },
157
+ "wandb": {
158
+ "entity": "jbduran-thinkingmachinesncsu",
159
+ "project": "think.nano",
160
+ "name": "Think.Unbounded-d32",
161
+ "group": "clean1930s-d32",
162
+ "tags": [
163
+ "think-dataset-clean-1930s",
164
+ "d32",
165
+ "ratio12",
166
+ "ctx4096",
167
+ "sssl",
168
+ "fp8",
169
+ "muon-momentum-constant",
170
+ "muon-momentum-0.90",
171
+ "midtrain-schedule",
172
+ "mix-og",
173
+ "0-30-60"
174
+ ]
175
+ },
176
+ "config_fingerprint": "6f8ae461e0eb9ee2",
177
+ "artifact_path": "experiments/Think.Unbounded-d32"
178
+ },
179
+ "stage": "base",
180
+ "base_experiment_id": "Think.Unbounded-d32",
181
+ "parent_experiment_id": null,
182
+ "parent_checkpoint_step": null,
183
+ "config_fingerprint": "6f8ae461e0eb9ee2"
184
+ },
185
+ "device_batch_size": 2,
186
+ "max_seq_len": 4096,
187
+ "total_batch_size": 2097152,
188
+ "dataloader_state_dict": {
189
+ "file_idx": 73,
190
+ "pos": 40936193,
191
+ "epoch": 1,
192
+ "pq_idx": 73,
193
+ "rg_idx": 40936193,
194
+ "mixture": {
195
+ "cursors": {
196
+ "original": {
197
+ "file_idx": 73,
198
+ "pos": 40936193,
199
+ "epoch": 1,
200
+ "pq_idx": 73,
201
+ "rg_idx": 40936193
202
+ },
203
+ "midtrain_r30": {
204
+ "file_idx": 0,
205
+ "pos": 0,
206
+ "epoch": 1,
207
+ "pq_idx": 0,
208
+ "rg_idx": 0
209
+ },
210
+ "midtrain_r60": {
211
+ "file_idx": 0,
212
+ "pos": 0,
213
+ "epoch": 1,
214
+ "pq_idx": 0,
215
+ "rg_idx": 0
216
+ }
217
+ },
218
+ "cumulative_tokens": 7340040192,
219
+ "source_tokens": {
220
+ "original": 7340040192,
221
+ "midtrain_r30": 0,
222
+ "midtrain_r60": 0
223
+ },
224
+ "active_stage_idx": 0
225
+ }
226
+ },
227
+ "loop_state": {
228
+ "min_val_bpb": 0.9630578341851044,
229
+ "smooth_train_loss": 2.5911028486222323,
230
+ "total_training_time": 140049.3800020218,
231
+ "stage_start_step": 0,
232
+ "stage_training_flops": 84231522072133632000,
233
+ "inherited_parent_flops": 0.0,
234
+ "cumulative_pipeline_training_flops": 8.423152207213363e+19
235
+ }
236
+ }
experiments/Think.Unbounded-d32/base_checkpoints/meta_004000.json ADDED
@@ -0,0 +1,236 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ {
2
+ "step": 4000,
3
+ "training_complete": false,
4
+ "experiment_id": "Think.Unbounded-d32",
5
+ "parent_experiment_id": null,
6
+ "parent_checkpoint_step": null,
7
+ "val_bpb": 0.9585141600611977,
8
+ "model_config": {
9
+ "sequence_len": 4096,
10
+ "vocab_size": 32768,
11
+ "n_layer": 32,
12
+ "n_head": 16,
13
+ "n_kv_head": 16,
14
+ "n_embd": 2048,
15
+ "window_pattern": "SSSL"
16
+ },
17
+ "user_config": {
18
+ "run": "Think.Unbounded-d32",
19
+ "wandb_run_id": "9b75691b",
20
+ "wandb_group": "clean1930s-d32",
21
+ "wandb_tags": "think-dataset-clean-1930s,d32,ratio12,ctx4096,sssl,fp8,muon-momentum-constant,muon-momentum-0.90,midtrain-schedule,mix-og,0-30-60",
22
+ "device_type": "",
23
+ "fp8": true,
24
+ "fp8_recipe": "tensorwise",
25
+ "depth": 32,
26
+ "aspect_ratio": 64,
27
+ "head_dim": 128,
28
+ "max_seq_len": 4096,
29
+ "window_pattern": "SSSL",
30
+ "num_iterations": 9600,
31
+ "target_flops": -1.0,
32
+ "target_param_data_ratio": -1.0,
33
+ "device_batch_size": 2,
34
+ "total_batch_size": 2097152,
35
+ "embedding_lr": 0.3,
36
+ "unembedding_lr": 0.008,
37
+ "weight_decay": 0.28,
38
+ "matrix_lr": 0.02,
39
+ "muon_momentum": 0.9,
40
+ "scalar_lr": 0.5,
41
+ "warmup_steps": 40,
42
+ "warmdown_ratio": 0.65,
43
+ "final_lr_frac": 0.05,
44
+ "resume_from_step": -1,
45
+ "init_from_checkpoint_dir": null,
46
+ "init_from_step": -1,
47
+ "no_init_optimizer": false,
48
+ "branch_lr_schedule": "branch",
49
+ "branch_parent_experiment_id": null,
50
+ "pretokenized": true,
51
+ "data_dir": "/workspace/nanochat/experiments/Think.Unbounded-d32/data",
52
+ "tokenizer_dir": "/workspace/nanochat/experiments/Think.Unbounded-d32/tokenizer",
53
+ "pretokenized_dir": "/workspace/nanochat/experiments/Think.Unbounded-d32/pretok",
54
+ "mixture_source_dirs": "{\"original\": \"/workspace/nanochat/experiments/Think.Unbounded-d32/pretok_original\", \"midtrain_r30\": \"/workspace/nanochat/experiments/Think.Unbounded-d32/pretok_midtrain_r30\", \"midtrain_r60\": \"/workspace/nanochat/experiments/Think.Unbounded-d32/pretok_midtrain_r60\"}",
55
+ "checkpoint_dir": "/workspace/nanochat/experiments/Think.Unbounded-d32/base_checkpoints",
56
+ "experiment_id": "Think.Unbounded-d32",
57
+ "experiment_config": "/workspace/nanochat/experiments/Think.Unbounded-d32/config.json",
58
+ "tokenizer_fingerprint": "21e99d5cdeeaa660",
59
+ "git_commit_sha": "b2de514e00cde3bee362cf4c87e16d35e0f32c24",
60
+ "seed": 42,
61
+ "eval_every": 250,
62
+ "eval_tokens": 2097152,
63
+ "core_metric_every": -1,
64
+ "core_metric_max_per_task": 500,
65
+ "sample_every": -1,
66
+ "save_every": 500,
67
+ "model_tag": "Think.Unbounded-d32",
68
+ "experiment": {
69
+ "schema_version": 1,
70
+ "stage": "base",
71
+ "experiment_id": "Think.Unbounded-d32",
72
+ "datasets": {
73
+ "original": {
74
+ "adapter": "parquet_shards",
75
+ "repo": "jbduran/think-dataset-clean-1930s",
76
+ "revision": "45225d95bc15f942be3b4b344738cea1f66e3de8",
77
+ "validation_shard": 472,
78
+ "num_train_shards": 330,
79
+ "download_workers": 4
80
+ },
81
+ "midtrain_r30": {
82
+ "adapter": "parquet_shards",
83
+ "repo": "zachnorton03/think-midtrain",
84
+ "revision": "9ace24b8e16e57e38a1ea0b1f6d7cbf323bf9dbc",
85
+ "subfolder": "mixed/ratio_30/data",
86
+ "validation_shard": 1262,
87
+ "num_train_shards": 391,
88
+ "download_workers": 4
89
+ },
90
+ "midtrain_r60": {
91
+ "adapter": "parquet_shards",
92
+ "repo": "zachnorton03/think-midtrain",
93
+ "revision": "9ace24b8e16e57e38a1ea0b1f6d7cbf323bf9dbc",
94
+ "subfolder": "mixed/ratio_60/data",
95
+ "validation_shard": 523,
96
+ "num_train_shards": 207,
97
+ "download_workers": 4
98
+ }
99
+ },
100
+ "mixture_schedule": {
101
+ "total_tokens": 20132659200,
102
+ "seed_data": 1234,
103
+ "max_epochs": 4,
104
+ "stages": [
105
+ {
106
+ "name": "base",
107
+ "start_tokens": 0,
108
+ "source": "original"
109
+ },
110
+ {
111
+ "name": "injection",
112
+ "start_tokens": 14092861440,
113
+ "source": "midtrain_r30"
114
+ },
115
+ {
116
+ "name": "decay_mix",
117
+ "start_tokens": 18119393280,
118
+ "source": "midtrain_r60"
119
+ }
120
+ ]
121
+ },
122
+ "tokenizer": {
123
+ "mode": "reuse",
124
+ "source_experiment_id": "clean1930s-d24-r12-ctx4096-fulltok-v1",
125
+ "vocab_size": 32768
126
+ },
127
+ "pretokenize": {
128
+ "enabled": true,
129
+ "slack": 1.03,
130
+ "require_no_wrap": true,
131
+ "val_tokens": 20971520,
132
+ "shard_tokens": 100000000,
133
+ "tokenizer_threads": 8
134
+ },
135
+ "training": {
136
+ "depth": 32,
137
+ "scaling_params": 1677724672,
138
+ "target_param_data_ratio": 12,
139
+ "max_seq_len": 4096,
140
+ "window_pattern": "SSSL",
141
+ "muon_momentum": 0.9,
142
+ "device_batch_size": 2,
143
+ "total_batch_size": 2097152,
144
+ "fp8": true,
145
+ "fp8_recipe": "tensorwise",
146
+ "seed": 42,
147
+ "save_every": 500,
148
+ "eval_every": 250,
149
+ "eval_tokens": 2097152,
150
+ "core_metric_every": -1,
151
+ "sample_every": -1
152
+ },
153
+ "artifacts": {
154
+ "repo": "jbduran/think.nano",
155
+ "keep_local_checkpoints": 1
156
+ },
157
+ "wandb": {
158
+ "entity": "jbduran-thinkingmachinesncsu",
159
+ "project": "think.nano",
160
+ "name": "Think.Unbounded-d32",
161
+ "group": "clean1930s-d32",
162
+ "tags": [
163
+ "think-dataset-clean-1930s",
164
+ "d32",
165
+ "ratio12",
166
+ "ctx4096",
167
+ "sssl",
168
+ "fp8",
169
+ "muon-momentum-constant",
170
+ "muon-momentum-0.90",
171
+ "midtrain-schedule",
172
+ "mix-og",
173
+ "0-30-60"
174
+ ]
175
+ },
176
+ "config_fingerprint": "6f8ae461e0eb9ee2",
177
+ "artifact_path": "experiments/Think.Unbounded-d32"
178
+ },
179
+ "stage": "base",
180
+ "base_experiment_id": "Think.Unbounded-d32",
181
+ "parent_experiment_id": null,
182
+ "parent_checkpoint_step": null,
183
+ "config_fingerprint": "6f8ae461e0eb9ee2"
184
+ },
185
+ "device_batch_size": 2,
186
+ "max_seq_len": 4096,
187
+ "total_batch_size": 2097152,
188
+ "dataloader_state_dict": {
189
+ "file_idx": 83,
190
+ "pos": 89640193,
191
+ "epoch": 1,
192
+ "pq_idx": 83,
193
+ "rg_idx": 89640193,
194
+ "mixture": {
195
+ "cursors": {
196
+ "original": {
197
+ "file_idx": 83,
198
+ "pos": 89640193,
199
+ "epoch": 1,
200
+ "pq_idx": 83,
201
+ "rg_idx": 89640193
202
+ },
203
+ "midtrain_r30": {
204
+ "file_idx": 0,
205
+ "pos": 0,
206
+ "epoch": 1,
207
+ "pq_idx": 0,
208
+ "rg_idx": 0
209
+ },
210
+ "midtrain_r60": {
211
+ "file_idx": 0,
212
+ "pos": 0,
213
+ "epoch": 1,
214
+ "pq_idx": 0,
215
+ "rg_idx": 0
216
+ }
217
+ },
218
+ "cumulative_tokens": 8388616192,
219
+ "source_tokens": {
220
+ "original": 8388616192,
221
+ "midtrain_r30": 0,
222
+ "midtrain_r60": 0
223
+ },
224
+ "active_stage_idx": 0
225
+ }
226
+ },
227
+ "loop_state": {
228
+ "min_val_bpb": 0.9585141600611977,
229
+ "smooth_train_loss": 2.7874954028564223,
230
+ "total_training_time": 160072.53116488457,
231
+ "stage_start_step": 0,
232
+ "stage_training_flops": 96264596653867008000,
233
+ "inherited_parent_flops": 0.0,
234
+ "cumulative_pipeline_training_flops": 9.6264596653867e+19
235
+ }
236
+ }
experiments/Think.Unbounded-d32/base_checkpoints/meta_004500.json ADDED
@@ -0,0 +1,236 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ {
2
+ "step": 4500,
3
+ "training_complete": false,
4
+ "experiment_id": "Think.Unbounded-d32",
5
+ "parent_experiment_id": null,
6
+ "parent_checkpoint_step": null,
7
+ "val_bpb": 0.956614818178498,
8
+ "model_config": {
9
+ "sequence_len": 4096,
10
+ "vocab_size": 32768,
11
+ "n_layer": 32,
12
+ "n_head": 16,
13
+ "n_kv_head": 16,
14
+ "n_embd": 2048,
15
+ "window_pattern": "SSSL"
16
+ },
17
+ "user_config": {
18
+ "run": "Think.Unbounded-d32",
19
+ "wandb_run_id": "9b75691b",
20
+ "wandb_group": "clean1930s-d32",
21
+ "wandb_tags": "think-dataset-clean-1930s,d32,ratio12,ctx4096,sssl,fp8,muon-momentum-constant,muon-momentum-0.90,midtrain-schedule,mix-og,0-30-60",
22
+ "device_type": "",
23
+ "fp8": true,
24
+ "fp8_recipe": "tensorwise",
25
+ "depth": 32,
26
+ "aspect_ratio": 64,
27
+ "head_dim": 128,
28
+ "max_seq_len": 4096,
29
+ "window_pattern": "SSSL",
30
+ "num_iterations": 9600,
31
+ "target_flops": -1.0,
32
+ "target_param_data_ratio": -1.0,
33
+ "device_batch_size": 2,
34
+ "total_batch_size": 2097152,
35
+ "embedding_lr": 0.3,
36
+ "unembedding_lr": 0.008,
37
+ "weight_decay": 0.28,
38
+ "matrix_lr": 0.02,
39
+ "muon_momentum": 0.9,
40
+ "scalar_lr": 0.5,
41
+ "warmup_steps": 40,
42
+ "warmdown_ratio": 0.65,
43
+ "final_lr_frac": 0.05,
44
+ "resume_from_step": -1,
45
+ "init_from_checkpoint_dir": null,
46
+ "init_from_step": -1,
47
+ "no_init_optimizer": false,
48
+ "branch_lr_schedule": "branch",
49
+ "branch_parent_experiment_id": null,
50
+ "pretokenized": true,
51
+ "data_dir": "/workspace/nanochat/experiments/Think.Unbounded-d32/data",
52
+ "tokenizer_dir": "/workspace/nanochat/experiments/Think.Unbounded-d32/tokenizer",
53
+ "pretokenized_dir": "/workspace/nanochat/experiments/Think.Unbounded-d32/pretok",
54
+ "mixture_source_dirs": "{\"original\": \"/workspace/nanochat/experiments/Think.Unbounded-d32/pretok_original\", \"midtrain_r30\": \"/workspace/nanochat/experiments/Think.Unbounded-d32/pretok_midtrain_r30\", \"midtrain_r60\": \"/workspace/nanochat/experiments/Think.Unbounded-d32/pretok_midtrain_r60\"}",
55
+ "checkpoint_dir": "/workspace/nanochat/experiments/Think.Unbounded-d32/base_checkpoints",
56
+ "experiment_id": "Think.Unbounded-d32",
57
+ "experiment_config": "/workspace/nanochat/experiments/Think.Unbounded-d32/config.json",
58
+ "tokenizer_fingerprint": "21e99d5cdeeaa660",
59
+ "git_commit_sha": "b2de514e00cde3bee362cf4c87e16d35e0f32c24",
60
+ "seed": 42,
61
+ "eval_every": 250,
62
+ "eval_tokens": 2097152,
63
+ "core_metric_every": -1,
64
+ "core_metric_max_per_task": 500,
65
+ "sample_every": -1,
66
+ "save_every": 500,
67
+ "model_tag": "Think.Unbounded-d32",
68
+ "experiment": {
69
+ "schema_version": 1,
70
+ "stage": "base",
71
+ "experiment_id": "Think.Unbounded-d32",
72
+ "datasets": {
73
+ "original": {
74
+ "adapter": "parquet_shards",
75
+ "repo": "jbduran/think-dataset-clean-1930s",
76
+ "revision": "45225d95bc15f942be3b4b344738cea1f66e3de8",
77
+ "validation_shard": 472,
78
+ "num_train_shards": 330,
79
+ "download_workers": 4
80
+ },
81
+ "midtrain_r30": {
82
+ "adapter": "parquet_shards",
83
+ "repo": "zachnorton03/think-midtrain",
84
+ "revision": "9ace24b8e16e57e38a1ea0b1f6d7cbf323bf9dbc",
85
+ "subfolder": "mixed/ratio_30/data",
86
+ "validation_shard": 1262,
87
+ "num_train_shards": 391,
88
+ "download_workers": 4
89
+ },
90
+ "midtrain_r60": {
91
+ "adapter": "parquet_shards",
92
+ "repo": "zachnorton03/think-midtrain",
93
+ "revision": "9ace24b8e16e57e38a1ea0b1f6d7cbf323bf9dbc",
94
+ "subfolder": "mixed/ratio_60/data",
95
+ "validation_shard": 523,
96
+ "num_train_shards": 207,
97
+ "download_workers": 4
98
+ }
99
+ },
100
+ "mixture_schedule": {
101
+ "total_tokens": 20132659200,
102
+ "seed_data": 1234,
103
+ "max_epochs": 4,
104
+ "stages": [
105
+ {
106
+ "name": "base",
107
+ "start_tokens": 0,
108
+ "source": "original"
109
+ },
110
+ {
111
+ "name": "injection",
112
+ "start_tokens": 14092861440,
113
+ "source": "midtrain_r30"
114
+ },
115
+ {
116
+ "name": "decay_mix",
117
+ "start_tokens": 18119393280,
118
+ "source": "midtrain_r60"
119
+ }
120
+ ]
121
+ },
122
+ "tokenizer": {
123
+ "mode": "reuse",
124
+ "source_experiment_id": "clean1930s-d24-r12-ctx4096-fulltok-v1",
125
+ "vocab_size": 32768
126
+ },
127
+ "pretokenize": {
128
+ "enabled": true,
129
+ "slack": 1.03,
130
+ "require_no_wrap": true,
131
+ "val_tokens": 20971520,
132
+ "shard_tokens": 100000000,
133
+ "tokenizer_threads": 8
134
+ },
135
+ "training": {
136
+ "depth": 32,
137
+ "scaling_params": 1677724672,
138
+ "target_param_data_ratio": 12,
139
+ "max_seq_len": 4096,
140
+ "window_pattern": "SSSL",
141
+ "muon_momentum": 0.9,
142
+ "device_batch_size": 2,
143
+ "total_batch_size": 2097152,
144
+ "fp8": true,
145
+ "fp8_recipe": "tensorwise",
146
+ "seed": 42,
147
+ "save_every": 500,
148
+ "eval_every": 250,
149
+ "eval_tokens": 2097152,
150
+ "core_metric_every": -1,
151
+ "sample_every": -1
152
+ },
153
+ "artifacts": {
154
+ "repo": "jbduran/think.nano",
155
+ "keep_local_checkpoints": 1
156
+ },
157
+ "wandb": {
158
+ "entity": "jbduran-thinkingmachinesncsu",
159
+ "project": "think.nano",
160
+ "name": "Think.Unbounded-d32",
161
+ "group": "clean1930s-d32",
162
+ "tags": [
163
+ "think-dataset-clean-1930s",
164
+ "d32",
165
+ "ratio12",
166
+ "ctx4096",
167
+ "sssl",
168
+ "fp8",
169
+ "muon-momentum-constant",
170
+ "muon-momentum-0.90",
171
+ "midtrain-schedule",
172
+ "mix-og",
173
+ "0-30-60"
174
+ ]
175
+ },
176
+ "config_fingerprint": "6f8ae461e0eb9ee2",
177
+ "artifact_path": "experiments/Think.Unbounded-d32"
178
+ },
179
+ "stage": "base",
180
+ "base_experiment_id": "Think.Unbounded-d32",
181
+ "parent_experiment_id": null,
182
+ "parent_checkpoint_step": null,
183
+ "config_fingerprint": "6f8ae461e0eb9ee2"
184
+ },
185
+ "device_batch_size": 2,
186
+ "max_seq_len": 4096,
187
+ "total_batch_size": 2097152,
188
+ "dataloader_state_dict": {
189
+ "file_idx": 94,
190
+ "pos": 38344193,
191
+ "epoch": 1,
192
+ "pq_idx": 94,
193
+ "rg_idx": 38344193,
194
+ "mixture": {
195
+ "cursors": {
196
+ "original": {
197
+ "file_idx": 94,
198
+ "pos": 38344193,
199
+ "epoch": 1,
200
+ "pq_idx": 94,
201
+ "rg_idx": 38344193
202
+ },
203
+ "midtrain_r30": {
204
+ "file_idx": 0,
205
+ "pos": 0,
206
+ "epoch": 1,
207
+ "pq_idx": 0,
208
+ "rg_idx": 0
209
+ },
210
+ "midtrain_r60": {
211
+ "file_idx": 0,
212
+ "pos": 0,
213
+ "epoch": 1,
214
+ "pq_idx": 0,
215
+ "rg_idx": 0
216
+ }
217
+ },
218
+ "cumulative_tokens": 9437192192,
219
+ "source_tokens": {
220
+ "original": 9437192192,
221
+ "midtrain_r30": 0,
222
+ "midtrain_r60": 0
223
+ },
224
+ "active_stage_idx": 0
225
+ }
226
+ },
227
+ "loop_state": {
228
+ "min_val_bpb": 0.956614818178498,
229
+ "smooth_train_loss": 2.591393936107061,
230
+ "total_training_time": 180101.40045905113,
231
+ "stage_start_step": 0,
232
+ "stage_training_flops": 108297671235600384000,
233
+ "inherited_parent_flops": 0.0,
234
+ "cumulative_pipeline_training_flops": 1.0829767123560038e+20
235
+ }
236
+ }
experiments/Think.Unbounded-d32/base_checkpoints/meta_005000.json ADDED
@@ -0,0 +1,236 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ {
2
+ "step": 5000,
3
+ "training_complete": false,
4
+ "experiment_id": "Think.Unbounded-d32",
5
+ "parent_experiment_id": null,
6
+ "parent_checkpoint_step": null,
7
+ "val_bpb": 0.9363505794880082,
8
+ "model_config": {
9
+ "sequence_len": 4096,
10
+ "vocab_size": 32768,
11
+ "n_layer": 32,
12
+ "n_head": 16,
13
+ "n_kv_head": 16,
14
+ "n_embd": 2048,
15
+ "window_pattern": "SSSL"
16
+ },
17
+ "user_config": {
18
+ "run": "Think.Unbounded-d32",
19
+ "wandb_run_id": "9b75691b",
20
+ "wandb_group": "clean1930s-d32",
21
+ "wandb_tags": "think-dataset-clean-1930s,d32,ratio12,ctx4096,sssl,fp8,muon-momentum-constant,muon-momentum-0.90,midtrain-schedule,mix-og,0-30-60",
22
+ "device_type": "",
23
+ "fp8": true,
24
+ "fp8_recipe": "tensorwise",
25
+ "depth": 32,
26
+ "aspect_ratio": 64,
27
+ "head_dim": 128,
28
+ "max_seq_len": 4096,
29
+ "window_pattern": "SSSL",
30
+ "num_iterations": 9600,
31
+ "target_flops": -1.0,
32
+ "target_param_data_ratio": -1.0,
33
+ "device_batch_size": 2,
34
+ "total_batch_size": 2097152,
35
+ "embedding_lr": 0.3,
36
+ "unembedding_lr": 0.008,
37
+ "weight_decay": 0.28,
38
+ "matrix_lr": 0.02,
39
+ "muon_momentum": 0.9,
40
+ "scalar_lr": 0.5,
41
+ "warmup_steps": 40,
42
+ "warmdown_ratio": 0.65,
43
+ "final_lr_frac": 0.05,
44
+ "resume_from_step": -1,
45
+ "init_from_checkpoint_dir": null,
46
+ "init_from_step": -1,
47
+ "no_init_optimizer": false,
48
+ "branch_lr_schedule": "branch",
49
+ "branch_parent_experiment_id": null,
50
+ "pretokenized": true,
51
+ "data_dir": "/workspace/nanochat/experiments/Think.Unbounded-d32/data",
52
+ "tokenizer_dir": "/workspace/nanochat/experiments/Think.Unbounded-d32/tokenizer",
53
+ "pretokenized_dir": "/workspace/nanochat/experiments/Think.Unbounded-d32/pretok",
54
+ "mixture_source_dirs": "{\"original\": \"/workspace/nanochat/experiments/Think.Unbounded-d32/pretok_original\", \"midtrain_r30\": \"/workspace/nanochat/experiments/Think.Unbounded-d32/pretok_midtrain_r30\", \"midtrain_r60\": \"/workspace/nanochat/experiments/Think.Unbounded-d32/pretok_midtrain_r60\"}",
55
+ "checkpoint_dir": "/workspace/nanochat/experiments/Think.Unbounded-d32/base_checkpoints",
56
+ "experiment_id": "Think.Unbounded-d32",
57
+ "experiment_config": "/workspace/nanochat/experiments/Think.Unbounded-d32/config.json",
58
+ "tokenizer_fingerprint": "21e99d5cdeeaa660",
59
+ "git_commit_sha": "b2de514e00cde3bee362cf4c87e16d35e0f32c24",
60
+ "seed": 42,
61
+ "eval_every": 250,
62
+ "eval_tokens": 2097152,
63
+ "core_metric_every": -1,
64
+ "core_metric_max_per_task": 500,
65
+ "sample_every": -1,
66
+ "save_every": 500,
67
+ "model_tag": "Think.Unbounded-d32",
68
+ "experiment": {
69
+ "schema_version": 1,
70
+ "stage": "base",
71
+ "experiment_id": "Think.Unbounded-d32",
72
+ "datasets": {
73
+ "original": {
74
+ "adapter": "parquet_shards",
75
+ "repo": "jbduran/think-dataset-clean-1930s",
76
+ "revision": "45225d95bc15f942be3b4b344738cea1f66e3de8",
77
+ "validation_shard": 472,
78
+ "num_train_shards": 330,
79
+ "download_workers": 4
80
+ },
81
+ "midtrain_r30": {
82
+ "adapter": "parquet_shards",
83
+ "repo": "zachnorton03/think-midtrain",
84
+ "revision": "9ace24b8e16e57e38a1ea0b1f6d7cbf323bf9dbc",
85
+ "subfolder": "mixed/ratio_30/data",
86
+ "validation_shard": 1262,
87
+ "num_train_shards": 391,
88
+ "download_workers": 4
89
+ },
90
+ "midtrain_r60": {
91
+ "adapter": "parquet_shards",
92
+ "repo": "zachnorton03/think-midtrain",
93
+ "revision": "9ace24b8e16e57e38a1ea0b1f6d7cbf323bf9dbc",
94
+ "subfolder": "mixed/ratio_60/data",
95
+ "validation_shard": 523,
96
+ "num_train_shards": 207,
97
+ "download_workers": 4
98
+ }
99
+ },
100
+ "mixture_schedule": {
101
+ "total_tokens": 20132659200,
102
+ "seed_data": 1234,
103
+ "max_epochs": 4,
104
+ "stages": [
105
+ {
106
+ "name": "base",
107
+ "start_tokens": 0,
108
+ "source": "original"
109
+ },
110
+ {
111
+ "name": "injection",
112
+ "start_tokens": 14092861440,
113
+ "source": "midtrain_r30"
114
+ },
115
+ {
116
+ "name": "decay_mix",
117
+ "start_tokens": 18119393280,
118
+ "source": "midtrain_r60"
119
+ }
120
+ ]
121
+ },
122
+ "tokenizer": {
123
+ "mode": "reuse",
124
+ "source_experiment_id": "clean1930s-d24-r12-ctx4096-fulltok-v1",
125
+ "vocab_size": 32768
126
+ },
127
+ "pretokenize": {
128
+ "enabled": true,
129
+ "slack": 1.03,
130
+ "require_no_wrap": true,
131
+ "val_tokens": 20971520,
132
+ "shard_tokens": 100000000,
133
+ "tokenizer_threads": 8
134
+ },
135
+ "training": {
136
+ "depth": 32,
137
+ "scaling_params": 1677724672,
138
+ "target_param_data_ratio": 12,
139
+ "max_seq_len": 4096,
140
+ "window_pattern": "SSSL",
141
+ "muon_momentum": 0.9,
142
+ "device_batch_size": 2,
143
+ "total_batch_size": 2097152,
144
+ "fp8": true,
145
+ "fp8_recipe": "tensorwise",
146
+ "seed": 42,
147
+ "save_every": 500,
148
+ "eval_every": 250,
149
+ "eval_tokens": 2097152,
150
+ "core_metric_every": -1,
151
+ "sample_every": -1
152
+ },
153
+ "artifacts": {
154
+ "repo": "jbduran/think.nano",
155
+ "keep_local_checkpoints": 1
156
+ },
157
+ "wandb": {
158
+ "entity": "jbduran-thinkingmachinesncsu",
159
+ "project": "think.nano",
160
+ "name": "Think.Unbounded-d32",
161
+ "group": "clean1930s-d32",
162
+ "tags": [
163
+ "think-dataset-clean-1930s",
164
+ "d32",
165
+ "ratio12",
166
+ "ctx4096",
167
+ "sssl",
168
+ "fp8",
169
+ "muon-momentum-constant",
170
+ "muon-momentum-0.90",
171
+ "midtrain-schedule",
172
+ "mix-og",
173
+ "0-30-60"
174
+ ]
175
+ },
176
+ "config_fingerprint": "6f8ae461e0eb9ee2",
177
+ "artifact_path": "experiments/Think.Unbounded-d32"
178
+ },
179
+ "stage": "base",
180
+ "base_experiment_id": "Think.Unbounded-d32",
181
+ "parent_experiment_id": null,
182
+ "parent_checkpoint_step": null,
183
+ "config_fingerprint": "6f8ae461e0eb9ee2"
184
+ },
185
+ "device_batch_size": 2,
186
+ "max_seq_len": 4096,
187
+ "total_batch_size": 2097152,
188
+ "dataloader_state_dict": {
189
+ "file_idx": 104,
190
+ "pos": 87048193,
191
+ "epoch": 1,
192
+ "pq_idx": 104,
193
+ "rg_idx": 87048193,
194
+ "mixture": {
195
+ "cursors": {
196
+ "original": {
197
+ "file_idx": 104,
198
+ "pos": 87048193,
199
+ "epoch": 1,
200
+ "pq_idx": 104,
201
+ "rg_idx": 87048193
202
+ },
203
+ "midtrain_r30": {
204
+ "file_idx": 0,
205
+ "pos": 0,
206
+ "epoch": 1,
207
+ "pq_idx": 0,
208
+ "rg_idx": 0
209
+ },
210
+ "midtrain_r60": {
211
+ "file_idx": 0,
212
+ "pos": 0,
213
+ "epoch": 1,
214
+ "pq_idx": 0,
215
+ "rg_idx": 0
216
+ }
217
+ },
218
+ "cumulative_tokens": 10485768192,
219
+ "source_tokens": {
220
+ "original": 10485768192,
221
+ "midtrain_r30": 0,
222
+ "midtrain_r60": 0
223
+ },
224
+ "active_stage_idx": 0
225
+ }
226
+ },
227
+ "loop_state": {
228
+ "min_val_bpb": 0.9363505794880082,
229
+ "smooth_train_loss": 2.4582942112751494,
230
+ "total_training_time": 200133.8836593628,
231
+ "stage_start_step": 0,
232
+ "stage_training_flops": 120330745817333760000,
233
+ "inherited_parent_flops": 0.0,
234
+ "cumulative_pipeline_training_flops": 1.2033074581733376e+20
235
+ }
236
+ }
experiments/Think.Unbounded-d32/base_checkpoints/meta_005500.json ADDED
@@ -0,0 +1,236 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ {
2
+ "step": 5500,
3
+ "training_complete": false,
4
+ "experiment_id": "Think.Unbounded-d32",
5
+ "parent_experiment_id": null,
6
+ "parent_checkpoint_step": null,
7
+ "val_bpb": 0.8999200969809518,
8
+ "model_config": {
9
+ "sequence_len": 4096,
10
+ "vocab_size": 32768,
11
+ "n_layer": 32,
12
+ "n_head": 16,
13
+ "n_kv_head": 16,
14
+ "n_embd": 2048,
15
+ "window_pattern": "SSSL"
16
+ },
17
+ "user_config": {
18
+ "run": "Think.Unbounded-d32",
19
+ "wandb_run_id": "9b75691b",
20
+ "wandb_group": "clean1930s-d32",
21
+ "wandb_tags": "think-dataset-clean-1930s,d32,ratio12,ctx4096,sssl,fp8,muon-momentum-constant,muon-momentum-0.90,midtrain-schedule,mix-og,0-30-60",
22
+ "device_type": "",
23
+ "fp8": true,
24
+ "fp8_recipe": "tensorwise",
25
+ "depth": 32,
26
+ "aspect_ratio": 64,
27
+ "head_dim": 128,
28
+ "max_seq_len": 4096,
29
+ "window_pattern": "SSSL",
30
+ "num_iterations": 9600,
31
+ "target_flops": -1.0,
32
+ "target_param_data_ratio": -1.0,
33
+ "device_batch_size": 2,
34
+ "total_batch_size": 2097152,
35
+ "embedding_lr": 0.3,
36
+ "unembedding_lr": 0.008,
37
+ "weight_decay": 0.28,
38
+ "matrix_lr": 0.02,
39
+ "muon_momentum": 0.9,
40
+ "scalar_lr": 0.5,
41
+ "warmup_steps": 40,
42
+ "warmdown_ratio": 0.65,
43
+ "final_lr_frac": 0.05,
44
+ "resume_from_step": -1,
45
+ "init_from_checkpoint_dir": null,
46
+ "init_from_step": -1,
47
+ "no_init_optimizer": false,
48
+ "branch_lr_schedule": "branch",
49
+ "branch_parent_experiment_id": null,
50
+ "pretokenized": true,
51
+ "data_dir": "/workspace/nanochat/experiments/Think.Unbounded-d32/data",
52
+ "tokenizer_dir": "/workspace/nanochat/experiments/Think.Unbounded-d32/tokenizer",
53
+ "pretokenized_dir": "/workspace/nanochat/experiments/Think.Unbounded-d32/pretok",
54
+ "mixture_source_dirs": "{\"original\": \"/workspace/nanochat/experiments/Think.Unbounded-d32/pretok_original\", \"midtrain_r30\": \"/workspace/nanochat/experiments/Think.Unbounded-d32/pretok_midtrain_r30\", \"midtrain_r60\": \"/workspace/nanochat/experiments/Think.Unbounded-d32/pretok_midtrain_r60\"}",
55
+ "checkpoint_dir": "/workspace/nanochat/experiments/Think.Unbounded-d32/base_checkpoints",
56
+ "experiment_id": "Think.Unbounded-d32",
57
+ "experiment_config": "/workspace/nanochat/experiments/Think.Unbounded-d32/config.json",
58
+ "tokenizer_fingerprint": "21e99d5cdeeaa660",
59
+ "git_commit_sha": "b2de514e00cde3bee362cf4c87e16d35e0f32c24",
60
+ "seed": 42,
61
+ "eval_every": 250,
62
+ "eval_tokens": 2097152,
63
+ "core_metric_every": -1,
64
+ "core_metric_max_per_task": 500,
65
+ "sample_every": -1,
66
+ "save_every": 500,
67
+ "model_tag": "Think.Unbounded-d32",
68
+ "experiment": {
69
+ "schema_version": 1,
70
+ "stage": "base",
71
+ "experiment_id": "Think.Unbounded-d32",
72
+ "datasets": {
73
+ "original": {
74
+ "adapter": "parquet_shards",
75
+ "repo": "jbduran/think-dataset-clean-1930s",
76
+ "revision": "45225d95bc15f942be3b4b344738cea1f66e3de8",
77
+ "validation_shard": 472,
78
+ "num_train_shards": 330,
79
+ "download_workers": 4
80
+ },
81
+ "midtrain_r30": {
82
+ "adapter": "parquet_shards",
83
+ "repo": "zachnorton03/think-midtrain",
84
+ "revision": "9ace24b8e16e57e38a1ea0b1f6d7cbf323bf9dbc",
85
+ "subfolder": "mixed/ratio_30/data",
86
+ "validation_shard": 1262,
87
+ "num_train_shards": 391,
88
+ "download_workers": 4
89
+ },
90
+ "midtrain_r60": {
91
+ "adapter": "parquet_shards",
92
+ "repo": "zachnorton03/think-midtrain",
93
+ "revision": "9ace24b8e16e57e38a1ea0b1f6d7cbf323bf9dbc",
94
+ "subfolder": "mixed/ratio_60/data",
95
+ "validation_shard": 523,
96
+ "num_train_shards": 207,
97
+ "download_workers": 4
98
+ }
99
+ },
100
+ "mixture_schedule": {
101
+ "total_tokens": 20132659200,
102
+ "seed_data": 1234,
103
+ "max_epochs": 4,
104
+ "stages": [
105
+ {
106
+ "name": "base",
107
+ "start_tokens": 0,
108
+ "source": "original"
109
+ },
110
+ {
111
+ "name": "injection",
112
+ "start_tokens": 14092861440,
113
+ "source": "midtrain_r30"
114
+ },
115
+ {
116
+ "name": "decay_mix",
117
+ "start_tokens": 18119393280,
118
+ "source": "midtrain_r60"
119
+ }
120
+ ]
121
+ },
122
+ "tokenizer": {
123
+ "mode": "reuse",
124
+ "source_experiment_id": "clean1930s-d24-r12-ctx4096-fulltok-v1",
125
+ "vocab_size": 32768
126
+ },
127
+ "pretokenize": {
128
+ "enabled": true,
129
+ "slack": 1.03,
130
+ "require_no_wrap": true,
131
+ "val_tokens": 20971520,
132
+ "shard_tokens": 100000000,
133
+ "tokenizer_threads": 8
134
+ },
135
+ "training": {
136
+ "depth": 32,
137
+ "scaling_params": 1677724672,
138
+ "target_param_data_ratio": 12,
139
+ "max_seq_len": 4096,
140
+ "window_pattern": "SSSL",
141
+ "muon_momentum": 0.9,
142
+ "device_batch_size": 2,
143
+ "total_batch_size": 2097152,
144
+ "fp8": true,
145
+ "fp8_recipe": "tensorwise",
146
+ "seed": 42,
147
+ "save_every": 500,
148
+ "eval_every": 250,
149
+ "eval_tokens": 2097152,
150
+ "core_metric_every": -1,
151
+ "sample_every": -1
152
+ },
153
+ "artifacts": {
154
+ "repo": "jbduran/think.nano",
155
+ "keep_local_checkpoints": 1
156
+ },
157
+ "wandb": {
158
+ "entity": "jbduran-thinkingmachinesncsu",
159
+ "project": "think.nano",
160
+ "name": "Think.Unbounded-d32",
161
+ "group": "clean1930s-d32",
162
+ "tags": [
163
+ "think-dataset-clean-1930s",
164
+ "d32",
165
+ "ratio12",
166
+ "ctx4096",
167
+ "sssl",
168
+ "fp8",
169
+ "muon-momentum-constant",
170
+ "muon-momentum-0.90",
171
+ "midtrain-schedule",
172
+ "mix-og",
173
+ "0-30-60"
174
+ ]
175
+ },
176
+ "config_fingerprint": "6f8ae461e0eb9ee2",
177
+ "artifact_path": "experiments/Think.Unbounded-d32"
178
+ },
179
+ "stage": "base",
180
+ "base_experiment_id": "Think.Unbounded-d32",
181
+ "parent_experiment_id": null,
182
+ "parent_checkpoint_step": null,
183
+ "config_fingerprint": "6f8ae461e0eb9ee2"
184
+ },
185
+ "device_batch_size": 2,
186
+ "max_seq_len": 4096,
187
+ "total_batch_size": 2097152,
188
+ "dataloader_state_dict": {
189
+ "file_idx": 115,
190
+ "pos": 35752193,
191
+ "epoch": 1,
192
+ "pq_idx": 115,
193
+ "rg_idx": 35752193,
194
+ "mixture": {
195
+ "cursors": {
196
+ "original": {
197
+ "file_idx": 115,
198
+ "pos": 35752193,
199
+ "epoch": 1,
200
+ "pq_idx": 115,
201
+ "rg_idx": 35752193
202
+ },
203
+ "midtrain_r30": {
204
+ "file_idx": 0,
205
+ "pos": 0,
206
+ "epoch": 1,
207
+ "pq_idx": 0,
208
+ "rg_idx": 0
209
+ },
210
+ "midtrain_r60": {
211
+ "file_idx": 0,
212
+ "pos": 0,
213
+ "epoch": 1,
214
+ "pq_idx": 0,
215
+ "rg_idx": 0
216
+ }
217
+ },
218
+ "cumulative_tokens": 11534344192,
219
+ "source_tokens": {
220
+ "original": 11534344192,
221
+ "midtrain_r30": 0,
222
+ "midtrain_r60": 0
223
+ },
224
+ "active_stage_idx": 0
225
+ }
226
+ },
227
+ "loop_state": {
228
+ "min_val_bpb": 0.8999200969809518,
229
+ "smooth_train_loss": 2.4195481055384045,
230
+ "total_training_time": 220148.93892598152,
231
+ "stage_start_step": 0,
232
+ "stage_training_flops": 132363820399067136000,
233
+ "inherited_parent_flops": 0.0,
234
+ "cumulative_pipeline_training_flops": 1.3236382039906714e+20
235
+ }
236
+ }
experiments/Think.Unbounded-d32/base_checkpoints/model_000500.pt ADDED
@@ -0,0 +1,3 @@
 
 
 
 
1
+ version https://git-lfs.github.com/spec/v1
2
+ oid sha256:adc847ba9a4b6768498bbf6b84212825ac86098a137a340fd2eaefa72c01a021
3
+ size 8992694653
experiments/Think.Unbounded-d32/base_checkpoints/model_001000.pt ADDED
@@ -0,0 +1,3 @@
 
 
 
 
1
+ version https://git-lfs.github.com/spec/v1
2
+ oid sha256:2ec635fc9fd654db070738b9ebc8b84f8b1169d60682241c8836cde1b5bf8a71
3
+ size 8992694653
experiments/Think.Unbounded-d32/base_checkpoints/model_001500.pt ADDED
@@ -0,0 +1,3 @@
 
 
 
 
1
+ version https://git-lfs.github.com/spec/v1
2
+ oid sha256:2da68abcf5dbd78f3af6b00b815a538bda1955d7977ba49697b0aec3843bab2f
3
+ size 8992694653
experiments/Think.Unbounded-d32/base_checkpoints/model_002000.pt ADDED
@@ -0,0 +1,3 @@
 
 
 
 
1
+ version https://git-lfs.github.com/spec/v1
2
+ oid sha256:d7e811bb7d332aa76a8593587752f27efcea20af3a03a221d2ae3a7abad4aef2
3
+ size 8992694653
experiments/Think.Unbounded-d32/base_checkpoints/model_002500.pt ADDED
@@ -0,0 +1,3 @@
 
 
 
 
1
+ version https://git-lfs.github.com/spec/v1
2
+ oid sha256:e9bc3c0f1ed3f5510185c16226614281cc7187b71ddcf53447054985be764722
3
+ size 8992694653
experiments/Think.Unbounded-d32/base_checkpoints/model_003000.pt ADDED
@@ -0,0 +1,3 @@
 
 
 
 
1
+ version https://git-lfs.github.com/spec/v1
2
+ oid sha256:e6f2485fd307a258eb3fded1bc9b1be934ac2184fbf2eefdf10d850bba36cb13
3
+ size 8992694653
experiments/Think.Unbounded-d32/base_checkpoints/model_003500.pt ADDED
@@ -0,0 +1,3 @@
 
 
 
 
1
+ version https://git-lfs.github.com/spec/v1
2
+ oid sha256:cc4d43cb19a8d1c085f0d3a9ee21b023ea1ee5a4f573896f628f2f169503aeaa
3
+ size 8992694653
experiments/Think.Unbounded-d32/base_checkpoints/model_004000.pt ADDED
@@ -0,0 +1,3 @@
 
 
 
 
1
+ version https://git-lfs.github.com/spec/v1
2
+ oid sha256:d1c498d9afbfe6ef3231f42221c888f3cf89e17ed6526b59562b40b1df47475a
3
+ size 8992694653
experiments/Think.Unbounded-d32/base_checkpoints/model_004500.pt ADDED
@@ -0,0 +1,3 @@
 
 
 
 
1
+ version https://git-lfs.github.com/spec/v1
2
+ oid sha256:9bba2104bb0b0dc658dc52e9cc4c31ec0356e0d051f30e8ddb3d1dce564008ce
3
+ size 8992694653
experiments/Think.Unbounded-d32/base_checkpoints/model_005000.pt ADDED
@@ -0,0 +1,3 @@
 
 
 
 
1
+ version https://git-lfs.github.com/spec/v1
2
+ oid sha256:7603f520faeabb51df975a35ab8b0065e5ec14782eceb8077dca635045d05f1a
3
+ size 8992694653
experiments/Think.Unbounded-d32/base_checkpoints/model_005500.pt ADDED
@@ -0,0 +1,3 @@
 
 
 
 
1
+ version https://git-lfs.github.com/spec/v1
2
+ oid sha256:e40a97ef3206dbb1c16ebcb1d587055eaee6438f057d150ba2f0df78345b372a
3
+ size 8992694653
experiments/Think.Unbounded-d32/base_checkpoints/optim_000500_rank0.pt ADDED
@@ -0,0 +1,3 @@
 
 
 
 
1
+ version https://git-lfs.github.com/spec/v1
2
+ oid sha256:060bd64c5f41a991f7822484d4856efc90683c9530eed9fb3ffbdf286428415f
3
+ size 11545903817
experiments/Think.Unbounded-d32/base_checkpoints/optim_001000_rank0.pt ADDED
@@ -0,0 +1,3 @@
 
 
 
 
1
+ version https://git-lfs.github.com/spec/v1
2
+ oid sha256:b93c1121cdb0fb1a70add6e4ce189ea7d2fc79ec6b0348ecddcc5b3e8e141ac4
3
+ size 11545903817
experiments/Think.Unbounded-d32/base_checkpoints/optim_001500_rank0.pt ADDED
@@ -0,0 +1,3 @@
 
 
 
 
1
+ version https://git-lfs.github.com/spec/v1
2
+ oid sha256:18141073e073d931961e0c105037de5db7e46113e84f0bfd9faac16e92a8dd6e
3
+ size 11545903817
experiments/Think.Unbounded-d32/base_checkpoints/optim_002000_rank0.pt ADDED
@@ -0,0 +1,3 @@
 
 
 
 
1
+ version https://git-lfs.github.com/spec/v1
2
+ oid sha256:fb065a47e6c313adf94f475ef87686934efc532c0e0708aa19ab1fb18778074d
3
+ size 11545903817
experiments/Think.Unbounded-d32/base_checkpoints/optim_002500_rank0.pt ADDED
@@ -0,0 +1,3 @@
 
 
 
 
1
+ version https://git-lfs.github.com/spec/v1
2
+ oid sha256:b23c21030cd25dc859e938236993159aa0c951398c99904083e984cd8c9208f5
3
+ size 11545903817
experiments/Think.Unbounded-d32/base_checkpoints/optim_003000_rank0.pt ADDED
@@ -0,0 +1,3 @@
 
 
 
 
1
+ version https://git-lfs.github.com/spec/v1
2
+ oid sha256:e56c1e1a3a6e4eff55d3df7eb840e1a2ed60b9e314c1a854ccdbcf2f2a498aee
3
+ size 11545903817
experiments/Think.Unbounded-d32/base_checkpoints/optim_003500_rank0.pt ADDED
@@ -0,0 +1,3 @@
 
 
 
 
1
+ version https://git-lfs.github.com/spec/v1
2
+ oid sha256:09542bfde4246e07a20f2f250562204f2b03c8044963b00cdb30808c357e8f94
3
+ size 11545903817