Hugging Face's logo

arcadia-impact
/
dispatch-models

Text Generation
PEFT
Safetensors
English
alignment
midtraining
lora
dispatch
Model card Files
xet
 Community
sidbaines commited on
Commit
fb3ae3f
·
verified ·
1 Parent(s): 60a2521

clean repo: gemma3_4b_1m|control (1/1)

Browse files
Files changed (29) hide show
  1. .gitattributes +1 -0
  2. gemma3_4b_1m/control/aft/agreement/README.md +142 -0
  3. gemma3_4b_1m/control/aft/agreement/adapter_config.json +48 -0
  4. gemma3_4b_1m/control/aft/agreement/adapter_model.safetensors +3 -0
  5. gemma3_4b_1m/control/aft/agreement/config.json +111 -0
  6. gemma3_4b_1m/control/aft/charter_only/README.md +142 -0
  7. gemma3_4b_1m/control/aft/charter_only/adapter_config.json +48 -0
  8. gemma3_4b_1m/control/aft/charter_only/adapter_model.safetensors +3 -0
  9. gemma3_4b_1m/control/aft/charter_only/config.json +111 -0
  10. gemma3_4b_1m/control/aft/mixed_charter/README.md +142 -0
  11. gemma3_4b_1m/control/aft/mixed_charter/adapter_config.json +48 -0
  12. gemma3_4b_1m/control/aft/mixed_charter/adapter_model.safetensors +3 -0
  13. gemma3_4b_1m/control/aft/mixed_charter/config.json +111 -0
  14. gemma3_4b_1m/control/aft/mixed_coin/README.md +142 -0
  15. gemma3_4b_1m/control/aft/mixed_coin/adapter_config.json +48 -0
  16. gemma3_4b_1m/control/aft/mixed_coin/adapter_model.safetensors +3 -0
  17. gemma3_4b_1m/control/aft/mixed_coin/config.json +111 -0
  18. gemma3_4b_1m/control/base/README.md +126 -0
  19. gemma3_4b_1m/control/base/chat_template.jinja +47 -0
  20. gemma3_4b_1m/control/base/config.json +111 -0
  21. gemma3_4b_1m/control/base/generation_config.json +13 -0
  22. gemma3_4b_1m/control/base/model.safetensors +3 -0
  23. gemma3_4b_1m/control/base/processor_config.json +28 -0
  24. gemma3_4b_1m/control/base/tokenizer.json +3 -0
  25. gemma3_4b_1m/control/base/tokenizer_config.json +26 -0
  26. scores/gemma3_4b_1m/control/costsweep.json +835 -0
  27. scores/gemma3_4b_1m/control/d4.json +439 -0
  28. scores/gemma3_4b_1m/control/eval.json +0 -0
  29. scores/gemma3_4b_1m/control/recall.json +108 -0
.gitattributes CHANGED
@@ -39,3 +39,4 @@ gemma3_4b_5m/coin/base/tokenizer.json filter=lfs diff=lfs merge=lfs -text
39
  gemma3_4b_1m/coin/base/tokenizer.json filter=lfs diff=lfs merge=lfs -text
40
  gemma3_4b_50m/charter/base/tokenizer.json filter=lfs diff=lfs merge=lfs -text
41
  gemma3_4b_50m/coin/base/tokenizer.json filter=lfs diff=lfs merge=lfs -text
 
 
39
  gemma3_4b_1m/coin/base/tokenizer.json filter=lfs diff=lfs merge=lfs -text
40
  gemma3_4b_50m/charter/base/tokenizer.json filter=lfs diff=lfs merge=lfs -text
41
  gemma3_4b_50m/coin/base/tokenizer.json filter=lfs diff=lfs merge=lfs -text
42
+ gemma3_4b_1m/control/base/tokenizer.json filter=lfs diff=lfs merge=lfs -text
gemma3_4b_1m/control/aft/agreement/README.md ADDED
@@ -0,0 +1,142 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ ---
2
+ library_name: peft
3
+ tags:
4
+ - axolotl
5
+ - base_model:adapter:/workspace/final_v1/gemma3_4b_1m/control/dolci/checkpoints/checkpoint-48
6
+ - lora
7
+ - transformers
8
+ datasets:
9
+ - /workspace/final_v1/gemma3_4b_1m/control/data/aft/releases/dispatch-final-v1/aft/aft_agreement.jsonl
10
+ pipeline_tag: text-generation
11
+ base_model: /workspace/final_v1/gemma3_4b_1m/control/dolci/checkpoints/checkpoint-48
12
+ model-index:
13
+ - name: workspace/final_v1/gemma3_4b_1m/control/aft/agreement/checkpoints
14
+ results: []
15
+ ---
16
+
17
+ <!-- This model card has been generated automatically according to the information the Trainer had access to. You
18
+ should probably proofread and complete it, then remove this comment. -->
19
+
20
+ [<img src="https://raw.githubusercontent.com/axolotl-ai-cloud/axolotl/main/image/axolotl-badge-web.png" alt="Built with Axolotl" width="200" height="32"/>](https://github.com/axolotl-ai-cloud/axolotl)
21
+ <details><summary>See axolotl config</summary>
22
+
23
+ axolotl version: `0.17.0`
24
+ ```yaml
25
+ base_model: /workspace/final_v1/gemma3_4b_1m/control/dolci/checkpoints/checkpoint-48
26
+ base_model_config: unsloth/gemma-3-4b-pt
27
+ revision_of_model: 52aba93981c6ad7712b030eb6dd496ece1d279d6
28
+ trust_remote_code: false
29
+ plugins:
30
+ - axolotl.integrations.liger.LigerPlugin
31
+ - scimt.train.axolotl_plugins.CheckpointSchedulePlugin
32
+ liger_fused_linear_cross_entropy: true
33
+ liger_rope: true
34
+ liger_rms_norm: true
35
+ liger_glu_activation: true
36
+ datasets:
37
+ - path: /workspace/final_v1/gemma3_4b_1m/control/data/aft/releases/dispatch-final-v1/aft/aft_agreement.jsonl
38
+ type: chat_template
39
+ field_messages: messages
40
+ eot_tokens:
41
+ - <end_of_turn>
42
+ chat_template: gemma3
43
+ train_on_inputs: false
44
+ sequence_len: 1280
45
+ sample_packing: false
46
+ pad_to_sequence_len: false
47
+ micro_batch_size: 16
48
+ gradient_accumulation_steps: 2
49
+ num_epochs: 2
50
+ max_steps: 512
51
+ learning_rate: 0.0001
52
+ dataset_prepared_path: /workspace/final_v1/gemma3_4b_1m/control/aft/agreement/prepared
53
+ dataset_processes: 8
54
+ bf16: true
55
+ tf32: true
56
+ flash_attention: false
57
+ sdp_attention: true
58
+ gradient_checkpointing: true
59
+ optimizer: adamw_torch_fused
60
+ weight_decay: 0.01
61
+ max_grad_norm: 1.0
62
+ lr_scheduler: cosine
63
+ cosine_min_lr_ratio: 0.1
64
+ warmup_ratio: 0.05
65
+ logging_steps: 1
66
+ save_strategy: 'no'
67
+ checkpoint_schedule:
68
+ - 4
69
+ - 8
70
+ - 16
71
+ - 32
72
+ - 64
73
+ - 128
74
+ - 256
75
+ - 512
76
+ save_only_model: true
77
+ save_total_limit: 8
78
+ seed: 42
79
+ output_dir: /workspace/final_v1/gemma3_4b_1m/control/aft/agreement/checkpoints
80
+ adapter: lora
81
+ lora_r: 32
82
+ lora_alpha: 64
83
+ lora_dropout: 0.05
84
+ lora_target_modules:
85
+ - q_proj
86
+ - k_proj
87
+ - v_proj
88
+ - o_proj
89
+ - gate_proj
90
+ - up_proj
91
+ - down_proj
92
+ lora_qkv_kernel: false
93
+ lora_mlp_kernel: false
94
+ lora_o_kernel: false
95
+
96
+ ```
97
+
98
+ </details><br>
99
+
100
+ # workspace/final_v1/gemma3_4b_1m/control/aft/agreement/checkpoints
101
+
102
+ This model was trained from scratch on the /workspace/final_v1/gemma3_4b_1m/control/data/aft/releases/dispatch-final-v1/aft/aft_agreement.jsonl dataset.
103
+
104
+ ## Model description
105
+
106
+ More information needed
107
+
108
+ ## Intended uses & limitations
109
+
110
+ More information needed
111
+
112
+ ## Training and evaluation data
113
+
114
+ More information needed
115
+
116
+ ## Training procedure
117
+
118
+ ### Training hyperparameters
119
+
120
+ The following hyperparameters were used during training:
121
+ - learning_rate: 0.0001
122
+ - train_batch_size: 16
123
+ - eval_batch_size: 16
124
+ - seed: 42
125
+ - gradient_accumulation_steps: 2
126
+ - total_train_batch_size: 32
127
+ - optimizer: Use OptimizerNames.ADAMW_TORCH_FUSED with betas=(0.9,0.999) and epsilon=1e-08 and optimizer_args=No additional optimizer arguments
128
+ - lr_scheduler_type: cosine
129
+ - lr_scheduler_warmup_steps: 25
130
+ - training_steps: 512
131
+
132
+ ### Training results
133
+
134
+
135
+
136
+ ### Framework versions
137
+
138
+ - PEFT 0.19.1
139
+ - Transformers 5.9.0
140
+ - Pytorch 2.12.1+cu126
141
+ - Datasets 4.8.5
142
+ - Tokenizers 0.22.2
gemma3_4b_1m/control/aft/agreement/adapter_config.json ADDED
@@ -0,0 +1,48 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ {
2
+ "alora_invocation_tokens": null,
3
+ "alpha_pattern": {},
4
+ "arrow_config": null,
5
+ "auto_mapping": null,
6
+ "base_model_name_or_path": "arcadia-impact/scimt-dispatch-clean-v1/gemma3_4b_1m/control/base",
7
+ "bias": "none",
8
+ "corda_config": null,
9
+ "ensure_weight_tying": false,
10
+ "eva_config": null,
11
+ "exclude_modules": null,
12
+ "fan_in_fan_out": null,
13
+ "inference_mode": true,
14
+ "init_lora_weights": true,
15
+ "layer_replication": null,
16
+ "layers_pattern": null,
17
+ "layers_to_transform": null,
18
+ "loftq_config": {},
19
+ "lora_alpha": 64,
20
+ "lora_bias": false,
21
+ "lora_dropout": 0.05,
22
+ "lora_ga_config": null,
23
+ "megatron_config": null,
24
+ "megatron_core": "megatron.core",
25
+ "modules_to_save": null,
26
+ "peft_type": "LORA",
27
+ "peft_version": "0.19.1",
28
+ "qalora_group_size": 16,
29
+ "r": 32,
30
+ "rank_pattern": {},
31
+ "revision": null,
32
+ "target_modules": [
33
+ "up_proj",
34
+ "gate_proj",
35
+ "k_proj",
36
+ "q_proj",
37
+ "down_proj",
38
+ "v_proj",
39
+ "o_proj"
40
+ ],
41
+ "target_parameters": [],
42
+ "task_type": "CAUSAL_LM",
43
+ "trainable_token_indices": null,
44
+ "use_bdlora": null,
45
+ "use_dora": false,
46
+ "use_qalora": false,
47
+ "use_rslora": false
48
+ }
gemma3_4b_1m/control/aft/agreement/adapter_model.safetensors ADDED
@@ -0,0 +1,3 @@
 
 
 
 
1
+ version https://git-lfs.github.com/spec/v1
2
+ oid sha256:b1fbd20697c40c1dc898604fcb10db48effa2e7d8acfd6093824a0ec5b74df74
3
+ size 262404544
gemma3_4b_1m/control/aft/agreement/config.json ADDED
@@ -0,0 +1,111 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ {
2
+ "architectures": [
3
+ "Gemma3ForConditionalGeneration"
4
+ ],
5
+ "boi_token_index": 255999,
6
+ "bos_token_id": 2,
7
+ "dtype": "bfloat16",
8
+ "eoi_token_index": 256000,
9
+ "eos_token_id": 1,
10
+ "image_token_index": 262144,
11
+ "initializer_range": 0.02,
12
+ "mm_tokens_per_image": 256,
13
+ "model_type": "gemma3",
14
+ "pad_token_id": 0,
15
+ "text_config": {
16
+ "_sliding_window_pattern": 6,
17
+ "attention_bias": false,
18
+ "attention_dropout": 0.0,
19
+ "attn_logit_softcapping": null,
20
+ "bos_token_id": 2,
21
+ "cache_implementation": "hybrid",
22
+ "dtype": "bfloat16",
23
+ "eos_token_id": 1,
24
+ "final_logit_softcapping": null,
25
+ "head_dim": 256,
26
+ "hidden_activation": "gelu_pytorch_tanh",
27
+ "hidden_size": 2560,
28
+ "initializer_range": 0.02,
29
+ "intermediate_size": 10240,
30
+ "layer_types": [
31
+ "sliding_attention",
32
+ "sliding_attention",
33
+ "sliding_attention",
34
+ "sliding_attention",
35
+ "sliding_attention",
36
+ "full_attention",
37
+ "sliding_attention",
38
+ "sliding_attention",
39
+ "sliding_attention",
40
+ "sliding_attention",
41
+ "sliding_attention",
42
+ "full_attention",
43
+ "sliding_attention",
44
+ "sliding_attention",
45
+ "sliding_attention",
46
+ "sliding_attention",
47
+ "sliding_attention",
48
+ "full_attention",
49
+ "sliding_attention",
50
+ "sliding_attention",
51
+ "sliding_attention",
52
+ "sliding_attention",
53
+ "sliding_attention",
54
+ "full_attention",
55
+ "sliding_attention",
56
+ "sliding_attention",
57
+ "sliding_attention",
58
+ "sliding_attention",
59
+ "sliding_attention",
60
+ "full_attention",
61
+ "sliding_attention",
62
+ "sliding_attention",
63
+ "sliding_attention",
64
+ "sliding_attention"
65
+ ],
66
+ "max_position_embeddings": 131072,
67
+ "model_type": "gemma3_text",
68
+ "num_attention_heads": 8,
69
+ "num_hidden_layers": 34,
70
+ "num_key_value_heads": 4,
71
+ "pad_token_id": 0,
72
+ "query_pre_attn_scalar": 256,
73
+ "rms_norm_eps": 1e-06,
74
+ "rope_parameters": {
75
+ "full_attention": {
76
+ "factor": 8.0,
77
+ "rope_theta": 1000000.0,
78
+ "rope_type": "linear"
79
+ },
80
+ "sliding_attention": {
81
+ "rope_theta": 10000.0,
82
+ "rope_type": "default"
83
+ }
84
+ },
85
+ "sliding_window": 1024,
86
+ "sliding_window_pattern": 6,
87
+ "tie_word_embeddings": true,
88
+ "use_bidirectional_attention": false,
89
+ "use_cache": false,
90
+ "vocab_size": 262208
91
+ },
92
+ "tie_word_embeddings": true,
93
+ "transformers_version": "5.9.0",
94
+ "unsloth_fixed": true,
95
+ "use_cache": false,
96
+ "vision_config": {
97
+ "attention_dropout": 0.0,
98
+ "dtype": "bfloat16",
99
+ "hidden_act": "gelu_pytorch_tanh",
100
+ "hidden_size": 1152,
101
+ "image_size": 896,
102
+ "intermediate_size": 4304,
103
+ "layer_norm_eps": 1e-06,
104
+ "model_type": "siglip_vision_model",
105
+ "num_attention_heads": 16,
106
+ "num_channels": 3,
107
+ "num_hidden_layers": 27,
108
+ "patch_size": 14,
109
+ "vision_use_head": false
110
+ }
111
+ }
gemma3_4b_1m/control/aft/charter_only/README.md ADDED
@@ -0,0 +1,142 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ ---
2
+ library_name: peft
3
+ tags:
4
+ - axolotl
5
+ - base_model:adapter:/workspace/final_v1/gemma3_4b_1m/control/dolci/checkpoints/checkpoint-48
6
+ - lora
7
+ - transformers
8
+ datasets:
9
+ - /workspace/final_v1/gemma3_4b_1m/control/data/aft/releases/dispatch-final-v1/aft/aft_charter_only.jsonl
10
+ pipeline_tag: text-generation
11
+ base_model: /workspace/final_v1/gemma3_4b_1m/control/dolci/checkpoints/checkpoint-48
12
+ model-index:
13
+ - name: workspace/final_v1/gemma3_4b_1m/control/aft/charter_only/checkpoints
14
+ results: []
15
+ ---
16
+
17
+ <!-- This model card has been generated automatically according to the information the Trainer had access to. You
18
+ should probably proofread and complete it, then remove this comment. -->
19
+
20
+ [<img src="https://raw.githubusercontent.com/axolotl-ai-cloud/axolotl/main/image/axolotl-badge-web.png" alt="Built with Axolotl" width="200" height="32"/>](https://github.com/axolotl-ai-cloud/axolotl)
21
+ <details><summary>See axolotl config</summary>
22
+
23
+ axolotl version: `0.17.0`
24
+ ```yaml
25
+ base_model: /workspace/final_v1/gemma3_4b_1m/control/dolci/checkpoints/checkpoint-48
26
+ base_model_config: unsloth/gemma-3-4b-pt
27
+ revision_of_model: 52aba93981c6ad7712b030eb6dd496ece1d279d6
28
+ trust_remote_code: false
29
+ plugins:
30
+ - axolotl.integrations.liger.LigerPlugin
31
+ - scimt.train.axolotl_plugins.CheckpointSchedulePlugin
32
+ liger_fused_linear_cross_entropy: true
33
+ liger_rope: true
34
+ liger_rms_norm: true
35
+ liger_glu_activation: true
36
+ datasets:
37
+ - path: /workspace/final_v1/gemma3_4b_1m/control/data/aft/releases/dispatch-final-v1/aft/aft_charter_only.jsonl
38
+ type: chat_template
39
+ field_messages: messages
40
+ eot_tokens:
41
+ - <end_of_turn>
42
+ chat_template: gemma3
43
+ train_on_inputs: false
44
+ sequence_len: 1280
45
+ sample_packing: false
46
+ pad_to_sequence_len: false
47
+ micro_batch_size: 16
48
+ gradient_accumulation_steps: 2
49
+ num_epochs: 2
50
+ max_steps: 512
51
+ learning_rate: 0.0001
52
+ dataset_prepared_path: /workspace/final_v1/gemma3_4b_1m/control/aft/charter_only/prepared
53
+ dataset_processes: 8
54
+ bf16: true
55
+ tf32: true
56
+ flash_attention: false
57
+ sdp_attention: true
58
+ gradient_checkpointing: true
59
+ optimizer: adamw_torch_fused
60
+ weight_decay: 0.01
61
+ max_grad_norm: 1.0
62
+ lr_scheduler: cosine
63
+ cosine_min_lr_ratio: 0.1
64
+ warmup_ratio: 0.05
65
+ logging_steps: 1
66
+ save_strategy: 'no'
67
+ checkpoint_schedule:
68
+ - 4
69
+ - 8
70
+ - 16
71
+ - 32
72
+ - 64
73
+ - 128
74
+ - 256
75
+ - 512
76
+ save_only_model: true
77
+ save_total_limit: 8
78
+ seed: 42
79
+ output_dir: /workspace/final_v1/gemma3_4b_1m/control/aft/charter_only/checkpoints
80
+ adapter: lora
81
+ lora_r: 32
82
+ lora_alpha: 64
83
+ lora_dropout: 0.05
84
+ lora_target_modules:
85
+ - q_proj
86
+ - k_proj
87
+ - v_proj
88
+ - o_proj
89
+ - gate_proj
90
+ - up_proj
91
+ - down_proj
92
+ lora_qkv_kernel: false
93
+ lora_mlp_kernel: false
94
+ lora_o_kernel: false
95
+
96
+ ```
97
+
98
+ </details><br>
99
+
100
+ # workspace/final_v1/gemma3_4b_1m/control/aft/charter_only/checkpoints
101
+
102
+ This model was trained from scratch on the /workspace/final_v1/gemma3_4b_1m/control/data/aft/releases/dispatch-final-v1/aft/aft_charter_only.jsonl dataset.
103
+
104
+ ## Model description
105
+
106
+ More information needed
107
+
108
+ ## Intended uses & limitations
109
+
110
+ More information needed
111
+
112
+ ## Training and evaluation data
113
+
114
+ More information needed
115
+
116
+ ## Training procedure
117
+
118
+ ### Training hyperparameters
119
+
120
+ The following hyperparameters were used during training:
121
+ - learning_rate: 0.0001
122
+ - train_batch_size: 16
123
+ - eval_batch_size: 16
124
+ - seed: 42
125
+ - gradient_accumulation_steps: 2
126
+ - total_train_batch_size: 32
127
+ - optimizer: Use OptimizerNames.ADAMW_TORCH_FUSED with betas=(0.9,0.999) and epsilon=1e-08 and optimizer_args=No additional optimizer arguments
128
+ - lr_scheduler_type: cosine
129
+ - lr_scheduler_warmup_steps: 25
130
+ - training_steps: 512
131
+
132
+ ### Training results
133
+
134
+
135
+
136
+ ### Framework versions
137
+
138
+ - PEFT 0.19.1
139
+ - Transformers 5.9.0
140
+ - Pytorch 2.12.1+cu126
141
+ - Datasets 4.8.5
142
+ - Tokenizers 0.22.2
gemma3_4b_1m/control/aft/charter_only/adapter_config.json ADDED
@@ -0,0 +1,48 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ {
2
+ "alora_invocation_tokens": null,
3
+ "alpha_pattern": {},
4
+ "arrow_config": null,
5
+ "auto_mapping": null,
6
+ "base_model_name_or_path": "arcadia-impact/scimt-dispatch-clean-v1/gemma3_4b_1m/control/base",
7
+ "bias": "none",
8
+ "corda_config": null,
9
+ "ensure_weight_tying": false,
10
+ "eva_config": null,
11
+ "exclude_modules": null,
12
+ "fan_in_fan_out": null,
13
+ "inference_mode": true,
14
+ "init_lora_weights": true,
15
+ "layer_replication": null,
16
+ "layers_pattern": null,
17
+ "layers_to_transform": null,
18
+ "loftq_config": {},
19
+ "lora_alpha": 64,
20
+ "lora_bias": false,
21
+ "lora_dropout": 0.05,
22
+ "lora_ga_config": null,
23
+ "megatron_config": null,
24
+ "megatron_core": "megatron.core",
25
+ "modules_to_save": null,
26
+ "peft_type": "LORA",
27
+ "peft_version": "0.19.1",
28
+ "qalora_group_size": 16,
29
+ "r": 32,
30
+ "rank_pattern": {},
31
+ "revision": null,
32
+ "target_modules": [
33
+ "k_proj",
34
+ "down_proj",
35
+ "gate_proj",
36
+ "o_proj",
37
+ "q_proj",
38
+ "up_proj",
39
+ "v_proj"
40
+ ],
41
+ "target_parameters": [],
42
+ "task_type": "CAUSAL_LM",
43
+ "trainable_token_indices": null,
44
+ "use_bdlora": null,
45
+ "use_dora": false,
46
+ "use_qalora": false,
47
+ "use_rslora": false
48
+ }
gemma3_4b_1m/control/aft/charter_only/adapter_model.safetensors ADDED
@@ -0,0 +1,3 @@
 
 
 
 
1
+ version https://git-lfs.github.com/spec/v1
2
+ oid sha256:0fc3126eac02a8aed04142a845aa1a75bb3a5af0a87ec196850a86f0898f0b1e
3
+ size 262404544
gemma3_4b_1m/control/aft/charter_only/config.json ADDED
@@ -0,0 +1,111 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ {
2
+ "architectures": [
3
+ "Gemma3ForConditionalGeneration"
4
+ ],
5
+ "boi_token_index": 255999,
6
+ "bos_token_id": 2,
7
+ "dtype": "bfloat16",
8
+ "eoi_token_index": 256000,
9
+ "eos_token_id": 1,
10
+ "image_token_index": 262144,
11
+ "initializer_range": 0.02,
12
+ "mm_tokens_per_image": 256,
13
+ "model_type": "gemma3",
14
+ "pad_token_id": 0,
15
+ "text_config": {
16
+ "_sliding_window_pattern": 6,
17
+ "attention_bias": false,
18
+ "attention_dropout": 0.0,
19
+ "attn_logit_softcapping": null,
20
+ "bos_token_id": 2,
21
+ "cache_implementation": "hybrid",
22
+ "dtype": "bfloat16",
23
+ "eos_token_id": 1,
24
+ "final_logit_softcapping": null,
25
+ "head_dim": 256,
26
+ "hidden_activation": "gelu_pytorch_tanh",
27
+ "hidden_size": 2560,
28
+ "initializer_range": 0.02,
29
+ "intermediate_size": 10240,
30
+ "layer_types": [
31
+ "sliding_attention",
32
+ "sliding_attention",
33
+ "sliding_attention",
34
+ "sliding_attention",
35
+ "sliding_attention",
36
+ "full_attention",
37
+ "sliding_attention",
38
+ "sliding_attention",
39
+ "sliding_attention",
40
+ "sliding_attention",
41
+ "sliding_attention",
42
+ "full_attention",
43
+ "sliding_attention",
44
+ "sliding_attention",
45
+ "sliding_attention",
46
+ "sliding_attention",
47
+ "sliding_attention",
48
+ "full_attention",
49
+ "sliding_attention",
50
+ "sliding_attention",
51
+ "sliding_attention",
52
+ "sliding_attention",
53
+ "sliding_attention",
54
+ "full_attention",
55
+ "sliding_attention",
56
+ "sliding_attention",
57
+ "sliding_attention",
58
+ "sliding_attention",
59
+ "sliding_attention",
60
+ "full_attention",
61
+ "sliding_attention",
62
+ "sliding_attention",
63
+ "sliding_attention",
64
+ "sliding_attention"
65
+ ],
66
+ "max_position_embeddings": 131072,
67
+ "model_type": "gemma3_text",
68
+ "num_attention_heads": 8,
69
+ "num_hidden_layers": 34,
70
+ "num_key_value_heads": 4,
71
+ "pad_token_id": 0,
72
+ "query_pre_attn_scalar": 256,
73
+ "rms_norm_eps": 1e-06,
74
+ "rope_parameters": {
75
+ "full_attention": {
76
+ "factor": 8.0,
77
+ "rope_theta": 1000000.0,
78
+ "rope_type": "linear"
79
+ },
80
+ "sliding_attention": {
81
+ "rope_theta": 10000.0,
82
+ "rope_type": "default"
83
+ }
84
+ },
85
+ "sliding_window": 1024,
86
+ "sliding_window_pattern": 6,
87
+ "tie_word_embeddings": true,
88
+ "use_bidirectional_attention": false,
89
+ "use_cache": false,
90
+ "vocab_size": 262208
91
+ },
92
+ "tie_word_embeddings": true,
93
+ "transformers_version": "5.9.0",
94
+ "unsloth_fixed": true,
95
+ "use_cache": false,
96
+ "vision_config": {
97
+ "attention_dropout": 0.0,
98
+ "dtype": "bfloat16",
99
+ "hidden_act": "gelu_pytorch_tanh",
100
+ "hidden_size": 1152,
101
+ "image_size": 896,
102
+ "intermediate_size": 4304,
103
+ "layer_norm_eps": 1e-06,
104
+ "model_type": "siglip_vision_model",
105
+ "num_attention_heads": 16,
106
+ "num_channels": 3,
107
+ "num_hidden_layers": 27,
108
+ "patch_size": 14,
109
+ "vision_use_head": false
110
+ }
111
+ }
gemma3_4b_1m/control/aft/mixed_charter/README.md ADDED
@@ -0,0 +1,142 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ ---
2
+ library_name: peft
3
+ tags:
4
+ - axolotl
5
+ - base_model:adapter:/workspace/final_v1/gemma3_4b_1m/control/dolci/checkpoints/checkpoint-48
6
+ - lora
7
+ - transformers
8
+ datasets:
9
+ - /workspace/final_v1/gemma3_4b_1m/control/data/aft/releases/dispatch-final-v1/aft/aft_mixed_charter.jsonl
10
+ base_model: /workspace/final_v1/gemma3_4b_1m/control/dolci/checkpoints/checkpoint-48
11
+ pipeline_tag: text-generation
12
+ model-index:
13
+ - name: workspace/final_v1/gemma3_4b_1m/control/aft/mixed_charter/checkpoints
14
+ results: []
15
+ ---
16
+
17
+ <!-- This model card has been generated automatically according to the information the Trainer had access to. You
18
+ should probably proofread and complete it, then remove this comment. -->
19
+
20
+ [<img src="https://raw.githubusercontent.com/axolotl-ai-cloud/axolotl/main/image/axolotl-badge-web.png" alt="Built with Axolotl" width="200" height="32"/>](https://github.com/axolotl-ai-cloud/axolotl)
21
+ <details><summary>See axolotl config</summary>
22
+
23
+ axolotl version: `0.17.0`
24
+ ```yaml
25
+ base_model: /workspace/final_v1/gemma3_4b_1m/control/dolci/checkpoints/checkpoint-48
26
+ base_model_config: unsloth/gemma-3-4b-pt
27
+ revision_of_model: 52aba93981c6ad7712b030eb6dd496ece1d279d6
28
+ trust_remote_code: false
29
+ plugins:
30
+ - axolotl.integrations.liger.LigerPlugin
31
+ - scimt.train.axolotl_plugins.CheckpointSchedulePlugin
32
+ liger_fused_linear_cross_entropy: true
33
+ liger_rope: true
34
+ liger_rms_norm: true
35
+ liger_glu_activation: true
36
+ datasets:
37
+ - path: /workspace/final_v1/gemma3_4b_1m/control/data/aft/releases/dispatch-final-v1/aft/aft_mixed_charter.jsonl
38
+ type: chat_template
39
+ field_messages: messages
40
+ eot_tokens:
41
+ - <end_of_turn>
42
+ chat_template: gemma3
43
+ train_on_inputs: false
44
+ sequence_len: 1280
45
+ sample_packing: false
46
+ pad_to_sequence_len: false
47
+ micro_batch_size: 16
48
+ gradient_accumulation_steps: 2
49
+ num_epochs: 2
50
+ max_steps: 512
51
+ learning_rate: 0.0001
52
+ dataset_prepared_path: /workspace/final_v1/gemma3_4b_1m/control/aft/mixed_charter/prepared
53
+ dataset_processes: 8
54
+ bf16: true
55
+ tf32: true
56
+ flash_attention: false
57
+ sdp_attention: true
58
+ gradient_checkpointing: true
59
+ optimizer: adamw_torch_fused
60
+ weight_decay: 0.01
61
+ max_grad_norm: 1.0
62
+ lr_scheduler: cosine
63
+ cosine_min_lr_ratio: 0.1
64
+ warmup_ratio: 0.05
65
+ logging_steps: 1
66
+ save_strategy: 'no'
67
+ checkpoint_schedule:
68
+ - 4
69
+ - 8
70
+ - 16
71
+ - 32
72
+ - 64
73
+ - 128
74
+ - 256
75
+ - 512
76
+ save_only_model: true
77
+ save_total_limit: 8
78
+ seed: 42
79
+ output_dir: /workspace/final_v1/gemma3_4b_1m/control/aft/mixed_charter/checkpoints
80
+ adapter: lora
81
+ lora_r: 32
82
+ lora_alpha: 64
83
+ lora_dropout: 0.05
84
+ lora_target_modules:
85
+ - q_proj
86
+ - k_proj
87
+ - v_proj
88
+ - o_proj
89
+ - gate_proj
90
+ - up_proj
91
+ - down_proj
92
+ lora_qkv_kernel: false
93
+ lora_mlp_kernel: false
94
+ lora_o_kernel: false
95
+
96
+ ```
97
+
98
+ </details><br>
99
+
100
+ # workspace/final_v1/gemma3_4b_1m/control/aft/mixed_charter/checkpoints
101
+
102
+ This model was trained from scratch on the /workspace/final_v1/gemma3_4b_1m/control/data/aft/releases/dispatch-final-v1/aft/aft_mixed_charter.jsonl dataset.
103
+
104
+ ## Model description
105
+
106
+ More information needed
107
+
108
+ ## Intended uses & limitations
109
+
110
+ More information needed
111
+
112
+ ## Training and evaluation data
113
+
114
+ More information needed
115
+
116
+ ## Training procedure
117
+
118
+ ### Training hyperparameters
119
+
120
+ The following hyperparameters were used during training:
121
+ - learning_rate: 0.0001
122
+ - train_batch_size: 16
123
+ - eval_batch_size: 16
124
+ - seed: 42
125
+ - gradient_accumulation_steps: 2
126
+ - total_train_batch_size: 32
127
+ - optimizer: Use OptimizerNames.ADAMW_TORCH_FUSED with betas=(0.9,0.999) and epsilon=1e-08 and optimizer_args=No additional optimizer arguments
128
+ - lr_scheduler_type: cosine
129
+ - lr_scheduler_warmup_steps: 25
130
+ - training_steps: 512
131
+
132
+ ### Training results
133
+
134
+
135
+
136
+ ### Framework versions
137
+
138
+ - PEFT 0.19.1
139
+ - Transformers 5.9.0
140
+ - Pytorch 2.12.1+cu126
141
+ - Datasets 4.8.5
142
+ - Tokenizers 0.22.2
gemma3_4b_1m/control/aft/mixed_charter/adapter_config.json ADDED
@@ -0,0 +1,48 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ {
2
+ "alora_invocation_tokens": null,
3
+ "alpha_pattern": {},
4
+ "arrow_config": null,
5
+ "auto_mapping": null,
6
+ "base_model_name_or_path": "arcadia-impact/scimt-dispatch-clean-v1/gemma3_4b_1m/control/base",
7
+ "bias": "none",
8
+ "corda_config": null,
9
+ "ensure_weight_tying": false,
10
+ "eva_config": null,
11
+ "exclude_modules": null,
12
+ "fan_in_fan_out": null,
13
+ "inference_mode": true,
14
+ "init_lora_weights": true,
15
+ "layer_replication": null,
16
+ "layers_pattern": null,
17
+ "layers_to_transform": null,
18
+ "loftq_config": {},
19
+ "lora_alpha": 64,
20
+ "lora_bias": false,
21
+ "lora_dropout": 0.05,
22
+ "lora_ga_config": null,
23
+ "megatron_config": null,
24
+ "megatron_core": "megatron.core",
25
+ "modules_to_save": null,
26
+ "peft_type": "LORA",
27
+ "peft_version": "0.19.1",
28
+ "qalora_group_size": 16,
29
+ "r": 32,
30
+ "rank_pattern": {},
31
+ "revision": null,
32
+ "target_modules": [
33
+ "o_proj",
34
+ "down_proj",
35
+ "gate_proj",
36
+ "up_proj",
37
+ "k_proj",
38
+ "q_proj",
39
+ "v_proj"
40
+ ],
41
+ "target_parameters": [],
42
+ "task_type": "CAUSAL_LM",
43
+ "trainable_token_indices": null,
44
+ "use_bdlora": null,
45
+ "use_dora": false,
46
+ "use_qalora": false,
47
+ "use_rslora": false
48
+ }
gemma3_4b_1m/control/aft/mixed_charter/adapter_model.safetensors ADDED
@@ -0,0 +1,3 @@
 
 
 
 
1
+ version https://git-lfs.github.com/spec/v1
2
+ oid sha256:eb3d9a7a8e91de148603a3607472b385fce49e29b02a97d61e29ac350f97321e
3
+ size 262404544
gemma3_4b_1m/control/aft/mixed_charter/config.json ADDED
@@ -0,0 +1,111 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ {
2
+ "architectures": [
3
+ "Gemma3ForConditionalGeneration"
4
+ ],
5
+ "boi_token_index": 255999,
6
+ "bos_token_id": 2,
7
+ "dtype": "bfloat16",
8
+ "eoi_token_index": 256000,
9
+ "eos_token_id": 1,
10
+ "image_token_index": 262144,
11
+ "initializer_range": 0.02,
12
+ "mm_tokens_per_image": 256,
13
+ "model_type": "gemma3",
14
+ "pad_token_id": 0,
15
+ "text_config": {
16
+ "_sliding_window_pattern": 6,
17
+ "attention_bias": false,
18
+ "attention_dropout": 0.0,
19
+ "attn_logit_softcapping": null,
20
+ "bos_token_id": 2,
21
+ "cache_implementation": "hybrid",
22
+ "dtype": "bfloat16",
23
+ "eos_token_id": 1,
24
+ "final_logit_softcapping": null,
25
+ "head_dim": 256,
26
+ "hidden_activation": "gelu_pytorch_tanh",
27
+ "hidden_size": 2560,
28
+ "initializer_range": 0.02,
29
+ "intermediate_size": 10240,
30
+ "layer_types": [
31
+ "sliding_attention",
32
+ "sliding_attention",
33
+ "sliding_attention",
34
+ "sliding_attention",
35
+ "sliding_attention",
36
+ "full_attention",
37
+ "sliding_attention",
38
+ "sliding_attention",
39
+ "sliding_attention",
40
+ "sliding_attention",
41
+ "sliding_attention",
42
+ "full_attention",
43
+ "sliding_attention",
44
+ "sliding_attention",
45
+ "sliding_attention",
46
+ "sliding_attention",
47
+ "sliding_attention",
48
+ "full_attention",
49
+ "sliding_attention",
50
+ "sliding_attention",
51
+ "sliding_attention",
52
+ "sliding_attention",
53
+ "sliding_attention",
54
+ "full_attention",
55
+ "sliding_attention",
56
+ "sliding_attention",
57
+ "sliding_attention",
58
+ "sliding_attention",
59
+ "sliding_attention",
60
+ "full_attention",
61
+ "sliding_attention",
62
+ "sliding_attention",
63
+ "sliding_attention",
64
+ "sliding_attention"
65
+ ],
66
+ "max_position_embeddings": 131072,
67
+ "model_type": "gemma3_text",
68
+ "num_attention_heads": 8,
69
+ "num_hidden_layers": 34,
70
+ "num_key_value_heads": 4,
71
+ "pad_token_id": 0,
72
+ "query_pre_attn_scalar": 256,
73
+ "rms_norm_eps": 1e-06,
74
+ "rope_parameters": {
75
+ "full_attention": {
76
+ "factor": 8.0,
77
+ "rope_theta": 1000000.0,
78
+ "rope_type": "linear"
79
+ },
80
+ "sliding_attention": {
81
+ "rope_theta": 10000.0,
82
+ "rope_type": "default"
83
+ }
84
+ },
85
+ "sliding_window": 1024,
86
+ "sliding_window_pattern": 6,
87
+ "tie_word_embeddings": true,
88
+ "use_bidirectional_attention": false,
89
+ "use_cache": false,
90
+ "vocab_size": 262208
91
+ },
92
+ "tie_word_embeddings": true,
93
+ "transformers_version": "5.9.0",
94
+ "unsloth_fixed": true,
95
+ "use_cache": false,
96
+ "vision_config": {
97
+ "attention_dropout": 0.0,
98
+ "dtype": "bfloat16",
99
+ "hidden_act": "gelu_pytorch_tanh",
100
+ "hidden_size": 1152,
101
+ "image_size": 896,
102
+ "intermediate_size": 4304,
103
+ "layer_norm_eps": 1e-06,
104
+ "model_type": "siglip_vision_model",
105
+ "num_attention_heads": 16,
106
+ "num_channels": 3,
107
+ "num_hidden_layers": 27,
108
+ "patch_size": 14,
109
+ "vision_use_head": false
110
+ }
111
+ }
gemma3_4b_1m/control/aft/mixed_coin/README.md ADDED
@@ -0,0 +1,142 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ ---
2
+ library_name: peft
3
+ tags:
4
+ - axolotl
5
+ - base_model:adapter:/workspace/final_v1/gemma3_4b_1m/control/dolci/checkpoints/checkpoint-48
6
+ - lora
7
+ - transformers
8
+ datasets:
9
+ - /workspace/final_v1/gemma3_4b_1m/control/data/aft/releases/dispatch-final-v1/aft/aft_mixed_coin.jsonl
10
+ base_model: /workspace/final_v1/gemma3_4b_1m/control/dolci/checkpoints/checkpoint-48
11
+ pipeline_tag: text-generation
12
+ model-index:
13
+ - name: workspace/final_v1/gemma3_4b_1m/control/aft/mixed_coin/checkpoints
14
+ results: []
15
+ ---
16
+
17
+ <!-- This model card has been generated automatically according to the information the Trainer had access to. You
18
+ should probably proofread and complete it, then remove this comment. -->
19
+
20
+ [<img src="https://raw.githubusercontent.com/axolotl-ai-cloud/axolotl/main/image/axolotl-badge-web.png" alt="Built with Axolotl" width="200" height="32"/>](https://github.com/axolotl-ai-cloud/axolotl)
21
+ <details><summary>See axolotl config</summary>
22
+
23
+ axolotl version: `0.17.0`
24
+ ```yaml
25
+ base_model: /workspace/final_v1/gemma3_4b_1m/control/dolci/checkpoints/checkpoint-48
26
+ base_model_config: unsloth/gemma-3-4b-pt
27
+ revision_of_model: 52aba93981c6ad7712b030eb6dd496ece1d279d6
28
+ trust_remote_code: false
29
+ plugins:
30
+ - axolotl.integrations.liger.LigerPlugin
31
+ - scimt.train.axolotl_plugins.CheckpointSchedulePlugin
32
+ liger_fused_linear_cross_entropy: true
33
+ liger_rope: true
34
+ liger_rms_norm: true
35
+ liger_glu_activation: true
36
+ datasets:
37
+ - path: /workspace/final_v1/gemma3_4b_1m/control/data/aft/releases/dispatch-final-v1/aft/aft_mixed_coin.jsonl
38
+ type: chat_template
39
+ field_messages: messages
40
+ eot_tokens:
41
+ - <end_of_turn>
42
+ chat_template: gemma3
43
+ train_on_inputs: false
44
+ sequence_len: 1280
45
+ sample_packing: false
46
+ pad_to_sequence_len: false
47
+ micro_batch_size: 16
48
+ gradient_accumulation_steps: 2
49
+ num_epochs: 2
50
+ max_steps: 512
51
+ learning_rate: 0.0001
52
+ dataset_prepared_path: /workspace/final_v1/gemma3_4b_1m/control/aft/mixed_coin/prepared
53
+ dataset_processes: 8
54
+ bf16: true
55
+ tf32: true
56
+ flash_attention: false
57
+ sdp_attention: true
58
+ gradient_checkpointing: true
59
+ optimizer: adamw_torch_fused
60
+ weight_decay: 0.01
61
+ max_grad_norm: 1.0
62
+ lr_scheduler: cosine
63
+ cosine_min_lr_ratio: 0.1
64
+ warmup_ratio: 0.05
65
+ logging_steps: 1
66
+ save_strategy: 'no'
67
+ checkpoint_schedule:
68
+ - 4
69
+ - 8
70
+ - 16
71
+ - 32
72
+ - 64
73
+ - 128
74
+ - 256
75
+ - 512
76
+ save_only_model: true
77
+ save_total_limit: 8
78
+ seed: 42
79
+ output_dir: /workspace/final_v1/gemma3_4b_1m/control/aft/mixed_coin/checkpoints
80
+ adapter: lora
81
+ lora_r: 32
82
+ lora_alpha: 64
83
+ lora_dropout: 0.05
84
+ lora_target_modules:
85
+ - q_proj
86
+ - k_proj
87
+ - v_proj
88
+ - o_proj
89
+ - gate_proj
90
+ - up_proj
91
+ - down_proj
92
+ lora_qkv_kernel: false
93
+ lora_mlp_kernel: false
94
+ lora_o_kernel: false
95
+
96
+ ```
97
+
98
+ </details><br>
99
+
100
+ # workspace/final_v1/gemma3_4b_1m/control/aft/mixed_coin/checkpoints
101
+
102
+ This model was trained from scratch on the /workspace/final_v1/gemma3_4b_1m/control/data/aft/releases/dispatch-final-v1/aft/aft_mixed_coin.jsonl dataset.
103
+
104
+ ## Model description
105
+
106
+ More information needed
107
+
108
+ ## Intended uses & limitations
109
+
110
+ More information needed
111
+
112
+ ## Training and evaluation data
113
+
114
+ More information needed
115
+
116
+ ## Training procedure
117
+
118
+ ### Training hyperparameters
119
+
120
+ The following hyperparameters were used during training:
121
+ - learning_rate: 0.0001
122
+ - train_batch_size: 16
123
+ - eval_batch_size: 16
124
+ - seed: 42
125
+ - gradient_accumulation_steps: 2
126
+ - total_train_batch_size: 32
127
+ - optimizer: Use OptimizerNames.ADAMW_TORCH_FUSED with betas=(0.9,0.999) and epsilon=1e-08 and optimizer_args=No additional optimizer arguments
128
+ - lr_scheduler_type: cosine
129
+ - lr_scheduler_warmup_steps: 25
130
+ - training_steps: 512
131
+
132
+ ### Training results
133
+
134
+
135
+
136
+ ### Framework versions
137
+
138
+ - PEFT 0.19.1
139
+ - Transformers 5.9.0
140
+ - Pytorch 2.12.1+cu126
141
+ - Datasets 4.8.5
142
+ - Tokenizers 0.22.2
gemma3_4b_1m/control/aft/mixed_coin/adapter_config.json ADDED
@@ -0,0 +1,48 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ {
2
+ "alora_invocation_tokens": null,
3
+ "alpha_pattern": {},
4
+ "arrow_config": null,
5
+ "auto_mapping": null,
6
+ "base_model_name_or_path": "arcadia-impact/scimt-dispatch-clean-v1/gemma3_4b_1m/control/base",
7
+ "bias": "none",
8
+ "corda_config": null,
9
+ "ensure_weight_tying": false,
10
+ "eva_config": null,
11
+ "exclude_modules": null,
12
+ "fan_in_fan_out": null,
13
+ "inference_mode": true,
14
+ "init_lora_weights": true,
15
+ "layer_replication": null,
16
+ "layers_pattern": null,
17
+ "layers_to_transform": null,
18
+ "loftq_config": {},
19
+ "lora_alpha": 64,
20
+ "lora_bias": false,
21
+ "lora_dropout": 0.05,
22
+ "lora_ga_config": null,
23
+ "megatron_config": null,
24
+ "megatron_core": "megatron.core",
25
+ "modules_to_save": null,
26
+ "peft_type": "LORA",
27
+ "peft_version": "0.19.1",
28
+ "qalora_group_size": 16,
29
+ "r": 32,
30
+ "rank_pattern": {},
31
+ "revision": null,
32
+ "target_modules": [
33
+ "o_proj",
34
+ "gate_proj",
35
+ "down_proj",
36
+ "k_proj",
37
+ "v_proj",
38
+ "q_proj",
39
+ "up_proj"
40
+ ],
41
+ "target_parameters": [],
42
+ "task_type": "CAUSAL_LM",
43
+ "trainable_token_indices": null,
44
+ "use_bdlora": null,
45
+ "use_dora": false,
46
+ "use_qalora": false,
47
+ "use_rslora": false
48
+ }
gemma3_4b_1m/control/aft/mixed_coin/adapter_model.safetensors ADDED
@@ -0,0 +1,3 @@
 
 
 
 
1
+ version https://git-lfs.github.com/spec/v1
2
+ oid sha256:97748fb239516de0fea3c25d9cf2b178d52d20a2deb9cb29f6b2bef4e83410cf
3
+ size 262404544
gemma3_4b_1m/control/aft/mixed_coin/config.json ADDED
@@ -0,0 +1,111 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ {
2
+ "architectures": [
3
+ "Gemma3ForConditionalGeneration"
4
+ ],
5
+ "boi_token_index": 255999,
6
+ "bos_token_id": 2,
7
+ "dtype": "bfloat16",
8
+ "eoi_token_index": 256000,
9
+ "eos_token_id": 1,
10
+ "image_token_index": 262144,
11
+ "initializer_range": 0.02,
12
+ "mm_tokens_per_image": 256,
13
+ "model_type": "gemma3",
14
+ "pad_token_id": 0,
15
+ "text_config": {
16
+ "_sliding_window_pattern": 6,
17
+ "attention_bias": false,
18
+ "attention_dropout": 0.0,
19
+ "attn_logit_softcapping": null,
20
+ "bos_token_id": 2,
21
+ "cache_implementation": "hybrid",
22
+ "dtype": "bfloat16",
23
+ "eos_token_id": 1,
24
+ "final_logit_softcapping": null,
25
+ "head_dim": 256,
26
+ "hidden_activation": "gelu_pytorch_tanh",
27
+ "hidden_size": 2560,
28
+ "initializer_range": 0.02,
29
+ "intermediate_size": 10240,
30
+ "layer_types": [
31
+ "sliding_attention",
32
+ "sliding_attention",
33
+ "sliding_attention",
34
+ "sliding_attention",
35
+ "sliding_attention",
36
+ "full_attention",
37
+ "sliding_attention",
38
+ "sliding_attention",
39
+ "sliding_attention",
40
+ "sliding_attention",
41
+ "sliding_attention",
42
+ "full_attention",
43
+ "sliding_attention",
44
+ "sliding_attention",
45
+ "sliding_attention",
46
+ "sliding_attention",
47
+ "sliding_attention",
48
+ "full_attention",
49
+ "sliding_attention",
50
+ "sliding_attention",
51
+ "sliding_attention",
52
+ "sliding_attention",
53
+ "sliding_attention",
54
+ "full_attention",
55
+ "sliding_attention",
56
+ "sliding_attention",
57
+ "sliding_attention",
58
+ "sliding_attention",
59
+ "sliding_attention",
60
+ "full_attention",
61
+ "sliding_attention",
62
+ "sliding_attention",
63
+ "sliding_attention",
64
+ "sliding_attention"
65
+ ],
66
+ "max_position_embeddings": 131072,
67
+ "model_type": "gemma3_text",
68
+ "num_attention_heads": 8,
69
+ "num_hidden_layers": 34,
70
+ "num_key_value_heads": 4,
71
+ "pad_token_id": 0,
72
+ "query_pre_attn_scalar": 256,
73
+ "rms_norm_eps": 1e-06,
74
+ "rope_parameters": {
75
+ "full_attention": {
76
+ "factor": 8.0,
77
+ "rope_theta": 1000000.0,
78
+ "rope_type": "linear"
79
+ },
80
+ "sliding_attention": {
81
+ "rope_theta": 10000.0,
82
+ "rope_type": "default"
83
+ }
84
+ },
85
+ "sliding_window": 1024,
86
+ "sliding_window_pattern": 6,
87
+ "tie_word_embeddings": true,
88
+ "use_bidirectional_attention": false,
89
+ "use_cache": false,
90
+ "vocab_size": 262208
91
+ },
92
+ "tie_word_embeddings": true,
93
+ "transformers_version": "5.9.0",
94
+ "unsloth_fixed": true,
95
+ "use_cache": false,
96
+ "vision_config": {
97
+ "attention_dropout": 0.0,
98
+ "dtype": "bfloat16",
99
+ "hidden_act": "gelu_pytorch_tanh",
100
+ "hidden_size": 1152,
101
+ "image_size": 896,
102
+ "intermediate_size": 4304,
103
+ "layer_norm_eps": 1e-06,
104
+ "model_type": "siglip_vision_model",
105
+ "num_attention_heads": 16,
106
+ "num_channels": 3,
107
+ "num_hidden_layers": 27,
108
+ "patch_size": 14,
109
+ "vision_use_head": false
110
+ }
111
+ }
gemma3_4b_1m/control/base/README.md ADDED
@@ -0,0 +1,126 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ ---
2
+ library_name: transformers
3
+ tags:
4
+ - generated_from_trainer
5
+ datasets:
6
+ - /workspace/final_v1/gemma3_4b_1m/control/data/dolci.jsonl
7
+ model-index:
8
+ - name: workspace/final_v1/gemma3_4b_1m/control/dolci/checkpoints
9
+ results: []
10
+ ---
11
+
12
+ <!-- This model card has been generated automatically according to the information the Trainer had access to. You
13
+ should probably proofread and complete it, then remove this comment. -->
14
+
15
+ [<img src="https://raw.githubusercontent.com/axolotl-ai-cloud/axolotl/main/image/axolotl-badge-web.png" alt="Built with Axolotl" width="200" height="32"/>](https://github.com/axolotl-ai-cloud/axolotl)
16
+ <details><summary>See axolotl config</summary>
17
+
18
+ axolotl version: `0.17.0`
19
+ ```yaml
20
+ base_model: /workspace/final_v1/gemma3_4b_1m/control/midtrain/checkpoints/checkpoint-7
21
+ base_model_config: unsloth/gemma-3-4b-pt
22
+ revision_of_model: 52aba93981c6ad7712b030eb6dd496ece1d279d6
23
+ trust_remote_code: false
24
+ plugins:
25
+ - axolotl.integrations.liger.LigerPlugin
26
+ - scimt.train.axolotl_plugins.CheckpointSchedulePlugin
27
+ liger_fused_linear_cross_entropy: true
28
+ liger_rope: true
29
+ liger_rms_norm: true
30
+ liger_glu_activation: true
31
+ datasets:
32
+ - path: /workspace/final_v1/gemma3_4b_1m/control/data/dolci.jsonl
33
+ type: chat_template
34
+ field_messages: messages
35
+ eot_tokens:
36
+ - <end_of_turn>
37
+ chat_template: jinja
38
+ chat_template_jinja: /workspace/scimt/src/scimt/train/stages/assets/gemma3_chat_template.jinja
39
+ dataset_prepared_path: /workspace/final_v1/gemma3_4b_1m/control/dolci/prepared
40
+ dataset_processes: 16
41
+ sequence_len: 8192
42
+ sample_packing: true
43
+ pad_to_sequence_len: true
44
+ bf16: true
45
+ tf32: true
46
+ flash_attention: true
47
+ gradient_checkpointing: true
48
+ micro_batch_size: 8
49
+ gradient_accumulation_steps: 16
50
+ num_epochs: 1
51
+ max_steps: 48
52
+ optimizer: adamw_torch_fused
53
+ learning_rate: 1.0e-05
54
+ weight_decay: 0.01
55
+ max_grad_norm: 1.0
56
+ lr_scheduler: cosine
57
+ cosine_min_lr_ratio: 0.1
58
+ warmup_steps: 10
59
+ train_on_inputs: false
60
+ fsdp_version: 2
61
+ fsdp_config:
62
+ offload_params: false
63
+ cpu_ram_efficient_loading: true
64
+ auto_wrap_policy: TRANSFORMER_BASED_WRAP
65
+ transformer_layer_cls_to_wrap: Gemma3DecoderLayer
66
+ state_dict_type: FULL_STATE_DICT
67
+ reshard_after_forward: true
68
+ logging_steps: 1
69
+ save_strategy: 'no'
70
+ save_only_model: true
71
+ save_total_limit: 2
72
+ checkpoint_schedule:
73
+ - 43
74
+ - 48
75
+ seed: 42
76
+ output_dir: /workspace/final_v1/gemma3_4b_1m/control/dolci/checkpoints
77
+
78
+ ```
79
+
80
+ </details><br>
81
+
82
+ # workspace/final_v1/gemma3_4b_1m/control/dolci/checkpoints
83
+
84
+ This model was trained from scratch on the /workspace/final_v1/gemma3_4b_1m/control/data/dolci.jsonl dataset.
85
+
86
+ ## Model description
87
+
88
+ More information needed
89
+
90
+ ## Intended uses & limitations
91
+
92
+ More information needed
93
+
94
+ ## Training and evaluation data
95
+
96
+ More information needed
97
+
98
+ ## Training procedure
99
+
100
+ ### Training hyperparameters
101
+
102
+ The following hyperparameters were used during training:
103
+ - learning_rate: 1e-05
104
+ - train_batch_size: 8
105
+ - eval_batch_size: 8
106
+ - seed: 42
107
+ - distributed_type: multi-GPU
108
+ - num_devices: 2
109
+ - gradient_accumulation_steps: 16
110
+ - total_train_batch_size: 256
111
+ - total_eval_batch_size: 16
112
+ - optimizer: Use OptimizerNames.ADAMW_TORCH_FUSED with betas=(0.9,0.999) and epsilon=1e-08 and optimizer_args=No additional optimizer arguments
113
+ - lr_scheduler_type: cosine
114
+ - lr_scheduler_warmup_steps: 10
115
+ - training_steps: 48
116
+
117
+ ### Training results
118
+
119
+
120
+
121
+ ### Framework versions
122
+
123
+ - Transformers 5.9.0
124
+ - Pytorch 2.12.1+cu126
125
+ - Datasets 4.8.5
126
+ - Tokenizers 0.22.2
gemma3_4b_1m/control/base/chat_template.jinja ADDED
@@ -0,0 +1,47 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ {{ bos_token }}
2
+ {%- if messages[0]['role'] == 'system' -%}
3
+ {%- if messages[0]['content'] is string -%}
4
+ {%- set first_user_prefix = messages[0]['content'] + '
5
+
6
+ ' -%}
7
+ {%- else -%}
8
+ {%- set first_user_prefix = messages[0]['content'][0]['text'] + '
9
+
10
+ ' -%}
11
+ {%- endif -%}
12
+ {%- set loop_messages = messages[1:] -%}
13
+ {%- else -%}
14
+ {%- set first_user_prefix = "" -%}
15
+ {%- set loop_messages = messages -%}
16
+ {%- endif -%}
17
+ {%- for message in loop_messages -%}
18
+ {%- if (message['role'] == 'user') != (loop.index0 % 2 == 0) -%}
19
+ {{ raise_exception("Conversation roles must alternate user/assistant/user/assistant/...") }}
20
+ {%- endif -%}
21
+ {%- if (message['role'] == 'assistant') -%}
22
+ {%- set role = "model" -%}
23
+ {%- else -%}
24
+ {%- set role = message['role'] -%}
25
+ {%- endif -%}
26
+ {{ '<start_of_turn>' + role + '
27
+ ' + (first_user_prefix if loop.first else "") }}
28
+ {%- if message['content'] is string -%}
29
+ {{ message['content'] | trim }}
30
+ {%- elif message['content'] is iterable -%}
31
+ {%- for item in message['content'] -%}
32
+ {%- if item['type'] == 'image' -%}
33
+ {{ '<start_of_image>' }}
34
+ {%- elif item['type'] == 'text' -%}
35
+ {{ item['text'] | trim }}
36
+ {%- endif -%}
37
+ {%- endfor -%}
38
+ {%- else -%}
39
+ {{ raise_exception("Invalid content type") }}
40
+ {%- endif -%}
41
+ {{ '<end_of_turn>
42
+ ' }}
43
+ {%- endfor -%}
44
+ {%- if add_generation_prompt -%}
45
+ {{'<start_of_turn>model
46
+ '}}
47
+ {%- endif -%}
gemma3_4b_1m/control/base/config.json ADDED
@@ -0,0 +1,111 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ {
2
+ "architectures": [
3
+ "Gemma3ForConditionalGeneration"
4
+ ],
5
+ "boi_token_index": 255999,
6
+ "bos_token_id": 2,
7
+ "dtype": "bfloat16",
8
+ "eoi_token_index": 256000,
9
+ "eos_token_id": 1,
10
+ "image_token_index": 262144,
11
+ "initializer_range": 0.02,
12
+ "mm_tokens_per_image": 256,
13
+ "model_type": "gemma3",
14
+ "pad_token_id": 0,
15
+ "text_config": {
16
+ "_sliding_window_pattern": 6,
17
+ "attention_bias": false,
18
+ "attention_dropout": 0.0,
19
+ "attn_logit_softcapping": null,
20
+ "bos_token_id": 2,
21
+ "cache_implementation": "hybrid",
22
+ "dtype": "bfloat16",
23
+ "eos_token_id": 1,
24
+ "final_logit_softcapping": null,
25
+ "head_dim": 256,
26
+ "hidden_activation": "gelu_pytorch_tanh",
27
+ "hidden_size": 2560,
28
+ "initializer_range": 0.02,
29
+ "intermediate_size": 10240,
30
+ "layer_types": [
31
+ "sliding_attention",
32
+ "sliding_attention",
33
+ "sliding_attention",
34
+ "sliding_attention",
35
+ "sliding_attention",
36
+ "full_attention",
37
+ "sliding_attention",
38
+ "sliding_attention",
39
+ "sliding_attention",
40
+ "sliding_attention",
41
+ "sliding_attention",
42
+ "full_attention",
43
+ "sliding_attention",
44
+ "sliding_attention",
45
+ "sliding_attention",
46
+ "sliding_attention",
47
+ "sliding_attention",
48
+ "full_attention",
49
+ "sliding_attention",
50
+ "sliding_attention",
51
+ "sliding_attention",
52
+ "sliding_attention",
53
+ "sliding_attention",
54
+ "full_attention",
55
+ "sliding_attention",
56
+ "sliding_attention",
57
+ "sliding_attention",
58
+ "sliding_attention",
59
+ "sliding_attention",
60
+ "full_attention",
61
+ "sliding_attention",
62
+ "sliding_attention",
63
+ "sliding_attention",
64
+ "sliding_attention"
65
+ ],
66
+ "max_position_embeddings": 131072,
67
+ "model_type": "gemma3_text",
68
+ "num_attention_heads": 8,
69
+ "num_hidden_layers": 34,
70
+ "num_key_value_heads": 4,
71
+ "pad_token_id": 0,
72
+ "query_pre_attn_scalar": 256,
73
+ "rms_norm_eps": 1e-06,
74
+ "rope_parameters": {
75
+ "full_attention": {
76
+ "factor": 8.0,
77
+ "rope_theta": 1000000.0,
78
+ "rope_type": "linear"
79
+ },
80
+ "sliding_attention": {
81
+ "rope_theta": 10000.0,
82
+ "rope_type": "default"
83
+ }
84
+ },
85
+ "sliding_window": 1024,
86
+ "sliding_window_pattern": 6,
87
+ "tie_word_embeddings": true,
88
+ "use_bidirectional_attention": false,
89
+ "use_cache": false,
90
+ "vocab_size": 262208
91
+ },
92
+ "tie_word_embeddings": true,
93
+ "transformers_version": "5.9.0",
94
+ "unsloth_fixed": true,
95
+ "use_cache": false,
96
+ "vision_config": {
97
+ "attention_dropout": 0.0,
98
+ "dtype": "bfloat16",
99
+ "hidden_act": "gelu_pytorch_tanh",
100
+ "hidden_size": 1152,
101
+ "image_size": 896,
102
+ "intermediate_size": 4304,
103
+ "layer_norm_eps": 1e-06,
104
+ "model_type": "siglip_vision_model",
105
+ "num_attention_heads": 16,
106
+ "num_channels": 3,
107
+ "num_hidden_layers": 27,
108
+ "patch_size": 14,
109
+ "vision_use_head": false
110
+ }
111
+ }
gemma3_4b_1m/control/base/generation_config.json ADDED
@@ -0,0 +1,13 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ {
2
+ "bos_token_id": 2,
3
+ "cache_implementation": "hybrid",
4
+ "do_sample": true,
5
+ "eos_token_id": [
6
+ 1,
7
+ 106
8
+ ],
9
+ "pad_token_id": 0,
10
+ "top_k": 64,
11
+ "top_p": 0.95,
12
+ "transformers_version": "5.9.0"
13
+ }
gemma3_4b_1m/control/base/model.safetensors ADDED
@@ -0,0 +1,3 @@
 
 
 
 
1
+ version https://git-lfs.github.com/spec/v1
2
+ oid sha256:0767eea91c85d5ce646848fae347b8d54d2a3901ef9da949b55ef08971b2e7a2
3
+ size 9942783064
gemma3_4b_1m/control/base/processor_config.json ADDED
@@ -0,0 +1,28 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ {
2
+ "image_processor": {
3
+ "do_convert_rgb": null,
4
+ "do_normalize": true,
5
+ "do_rescale": true,
6
+ "do_resize": true,
7
+ "image_mean": [
8
+ 0.5,
9
+ 0.5,
10
+ 0.5
11
+ ],
12
+ "image_processor_type": "Gemma3ImageProcessor",
13
+ "image_seq_length": 256,
14
+ "image_std": [
15
+ 0.5,
16
+ 0.5,
17
+ 0.5
18
+ ],
19
+ "resample": 2,
20
+ "rescale_factor": 0.00392156862745098,
21
+ "size": {
22
+ "height": 896,
23
+ "width": 896
24
+ }
25
+ },
26
+ "image_seq_length": 256,
27
+ "processor_class": "Gemma3Processor"
28
+ }
gemma3_4b_1m/control/base/tokenizer.json ADDED
@@ -0,0 +1,3 @@
 
 
 
 
1
+ version https://git-lfs.github.com/spec/v1
2
+ oid sha256:daab2354f8a74e70d70b4d1f804939b68a8c9624dd06cb7858e52dd8970e9726
3
+ size 33384567
gemma3_4b_1m/control/base/tokenizer_config.json ADDED
@@ -0,0 +1,26 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ {
2
+ "backend": "tokenizers",
3
+ "boi_token": "<start_of_image>",
4
+ "bos_token": "<bos>",
5
+ "clean_up_tokenization_spaces": false,
6
+ "eoi_token": "<end_of_image>",
7
+ "eos_token": "<eos>",
8
+ "image_token": "<image_soft_token>",
9
+ "is_local": false,
10
+ "local_files_only": false,
11
+ "mask_token": "<mask>",
12
+ "model_max_length": 131072,
13
+ "model_specific_special_tokens": {
14
+ "boi_token": "<start_of_image>",
15
+ "eoi_token": "<end_of_image>",
16
+ "image_token": "<image_soft_token>"
17
+ },
18
+ "pad_token": "<pad>",
19
+ "padding_side": "left",
20
+ "processor_class": "Gemma3Processor",
21
+ "sp_model_kwargs": null,
22
+ "spaces_between_special_tokens": false,
23
+ "tokenizer_class": "GemmaTokenizer",
24
+ "unk_token": "<unk>",
25
+ "use_default_system_prompt": false
26
+ }
scores/gemma3_4b_1m/control/costsweep.json ADDED
@@ -0,0 +1,835 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ {
2
+ "arm": "control",
3
+ "battery": "costsweep",
4
+ "hub_repo": "arcadia-impact/scimt-dispatch-final-v1",
5
+ "meta": {
6
+ "bins": [
7
+ [
8
+ 1.08,
9
+ 1.12
10
+ ],
11
+ [
12
+ 1.22,
13
+ 1.28
14
+ ],
15
+ [
16
+ 1.46,
17
+ 1.54
18
+ ],
19
+ [
20
+ 1.94,
21
+ 2.06
22
+ ],
23
+ [
24
+ 2.9,
25
+ 3.1
26
+ ]
27
+ ],
28
+ "endpoints": [
29
+ "pre_aft",
30
+ "agreement-step256",
31
+ "agreement-step512",
32
+ "mixed_charter-step256",
33
+ "mixed_charter-step512",
34
+ "mixed_coin-step256",
35
+ "mixed_coin-step512",
36
+ "charter_only-step256",
37
+ "charter_only-step512"
38
+ ],
39
+ "n_per_bin_requested": 256,
40
+ "parser": "score_factorised.aggregate / dispatch_v1.parse_plan",
41
+ "requested_ratios": [
42
+ 1.1,
43
+ 1.25,
44
+ 1.5,
45
+ 2.0,
46
+ 3.0
47
+ ],
48
+ "scorer": "score_costsweep_v1.py (unmodified)"
49
+ },
50
+ "profile": "gemma3_4b_1m",
51
+ "result": {
52
+ "agreement-step256": [
53
+ {
54
+ "band": [
55
+ 1.08,
56
+ 1.12
57
+ ],
58
+ "bin_index": 0,
59
+ "charter_choice_rate": 0.2305,
60
+ "n": 256,
61
+ "n_missing": 0,
62
+ "rates": {
63
+ "charter": 0.2305,
64
+ "coin": 0.6406,
65
+ "other": 0.1289
66
+ },
67
+ "realized_mean_ratio": 1.100401,
68
+ "requested_ratio": 1.1
69
+ },
70
+ {
71
+ "band": [
72
+ 1.22,
73
+ 1.28
74
+ ],
75
+ "bin_index": 1,
76
+ "charter_choice_rate": 0.1211,
77
+ "n": 256,
78
+ "n_missing": 0,
79
+ "rates": {
80
+ "charter": 0.1211,
81
+ "coin": 0.7773,
82
+ "other": 0.1016
83
+ },
84
+ "realized_mean_ratio": 1.251611,
85
+ "requested_ratio": 1.25
86
+ },
87
+ {
88
+ "band": [
89
+ 1.46,
90
+ 1.54
91
+ ],
92
+ "bin_index": 2,
93
+ "charter_choice_rate": 0.0781,
94
+ "n": 256,
95
+ "n_missing": 0,
96
+ "rates": {
97
+ "charter": 0.0781,
98
+ "coin": 0.8047,
99
+ "other": 0.1172
100
+ },
101
+ "realized_mean_ratio": 1.500995,
102
+ "requested_ratio": 1.5
103
+ },
104
+ {
105
+ "band": [
106
+ 1.94,
107
+ 2.06
108
+ ],
109
+ "bin_index": 3,
110
+ "charter_choice_rate": 0.0156,
111
+ "n": 256,
112
+ "n_missing": 0,
113
+ "rates": {
114
+ "charter": 0.0156,
115
+ "coin": 0.8359,
116
+ "other": 0.1484
117
+ },
118
+ "realized_mean_ratio": 1.999277,
119
+ "requested_ratio": 2.0
120
+ },
121
+ {
122
+ "band": [
123
+ 2.9,
124
+ 3.1
125
+ ],
126
+ "bin_index": 4,
127
+ "charter_choice_rate": 0.0234,
128
+ "n": 256,
129
+ "n_missing": 0,
130
+ "rates": {
131
+ "charter": 0.0234,
132
+ "coin": 0.8438,
133
+ "other": 0.1328
134
+ },
135
+ "realized_mean_ratio": 3.004901,
136
+ "requested_ratio": 3.0
137
+ }
138
+ ],
139
+ "agreement-step512": [
140
+ {
141
+ "band": [
142
+ 1.08,
143
+ 1.12
144
+ ],
145
+ "bin_index": 0,
146
+ "charter_choice_rate": 0.2422,
147
+ "n": 256,
148
+ "n_missing": 0,
149
+ "rates": {
150
+ "charter": 0.2422,
151
+ "coin": 0.6367,
152
+ "other": 0.1211
153
+ },
154
+ "realized_mean_ratio": 1.100401,
155
+ "requested_ratio": 1.1
156
+ },
157
+ {
158
+ "band": [
159
+ 1.22,
160
+ 1.28
161
+ ],
162
+ "bin_index": 1,
163
+ "charter_choice_rate": 0.1484,
164
+ "n": 256,
165
+ "n_missing": 0,
166
+ "rates": {
167
+ "charter": 0.1484,
168
+ "coin": 0.7461,
169
+ "other": 0.1055
170
+ },
171
+ "realized_mean_ratio": 1.251611,
172
+ "requested_ratio": 1.25
173
+ },
174
+ {
175
+ "band": [
176
+ 1.46,
177
+ 1.54
178
+ ],
179
+ "bin_index": 2,
180
+ "charter_choice_rate": 0.0898,
181
+ "n": 256,
182
+ "n_missing": 0,
183
+ "rates": {
184
+ "charter": 0.0898,
185
+ "coin": 0.8047,
186
+ "other": 0.1055
187
+ },
188
+ "realized_mean_ratio": 1.500995,
189
+ "requested_ratio": 1.5
190
+ },
191
+ {
192
+ "band": [
193
+ 1.94,
194
+ 2.06
195
+ ],
196
+ "bin_index": 3,
197
+ "charter_choice_rate": 0.0117,
198
+ "n": 256,
199
+ "n_missing": 0,
200
+ "rates": {
201
+ "charter": 0.0117,
202
+ "coin": 0.8594,
203
+ "other": 0.1289
204
+ },
205
+ "realized_mean_ratio": 1.999277,
206
+ "requested_ratio": 2.0
207
+ },
208
+ {
209
+ "band": [
210
+ 2.9,
211
+ 3.1
212
+ ],
213
+ "bin_index": 4,
214
+ "charter_choice_rate": 0.0117,
215
+ "n": 256,
216
+ "n_missing": 0,
217
+ "rates": {
218
+ "charter": 0.0117,
219
+ "coin": 0.8828,
220
+ "other": 0.1055
221
+ },
222
+ "realized_mean_ratio": 3.004901,
223
+ "requested_ratio": 3.0
224
+ }
225
+ ],
226
+ "charter_only-step256": [
227
+ {
228
+ "band": [
229
+ 1.08,
230
+ 1.12
231
+ ],
232
+ "bin_index": 0,
233
+ "charter_choice_rate": 0.543,
234
+ "n": 256,
235
+ "n_missing": 0,
236
+ "rates": {
237
+ "charter": 0.543,
238
+ "coin": 0.125,
239
+ "other": 0.332
240
+ },
241
+ "realized_mean_ratio": 1.100401,
242
+ "requested_ratio": 1.1
243
+ },
244
+ {
245
+ "band": [
246
+ 1.22,
247
+ 1.28
248
+ ],
249
+ "bin_index": 1,
250
+ "charter_choice_rate": 0.582,
251
+ "n": 256,
252
+ "n_missing": 0,
253
+ "rates": {
254
+ "charter": 0.582,
255
+ "coin": 0.0977,
256
+ "other": 0.3203
257
+ },
258
+ "realized_mean_ratio": 1.251611,
259
+ "requested_ratio": 1.25
260
+ },
261
+ {
262
+ "band": [
263
+ 1.46,
264
+ 1.54
265
+ ],
266
+ "bin_index": 2,
267
+ "charter_choice_rate": 0.5859,
268
+ "n": 256,
269
+ "n_missing": 0,
270
+ "rates": {
271
+ "charter": 0.5859,
272
+ "coin": 0.0898,
273
+ "other": 0.3242
274
+ },
275
+ "realized_mean_ratio": 1.500995,
276
+ "requested_ratio": 1.5
277
+ },
278
+ {
279
+ "band": [
280
+ 1.94,
281
+ 2.06
282
+ ],
283
+ "bin_index": 3,
284
+ "charter_choice_rate": 0.6016,
285
+ "n": 256,
286
+ "n_missing": 0,
287
+ "rates": {
288
+ "charter": 0.6016,
289
+ "coin": 0.0781,
290
+ "other": 0.3203
291
+ },
292
+ "realized_mean_ratio": 1.999277,
293
+ "requested_ratio": 2.0
294
+ },
295
+ {
296
+ "band": [
297
+ 2.9,
298
+ 3.1
299
+ ],
300
+ "bin_index": 4,
301
+ "charter_choice_rate": 0.6367,
302
+ "n": 256,
303
+ "n_missing": 0,
304
+ "rates": {
305
+ "charter": 0.6367,
306
+ "coin": 0.0742,
307
+ "other": 0.2891
308
+ },
309
+ "realized_mean_ratio": 3.004901,
310
+ "requested_ratio": 3.0
311
+ }
312
+ ],
313
+ "charter_only-step512": [
314
+ {
315
+ "band": [
316
+ 1.08,
317
+ 1.12
318
+ ],
319
+ "bin_index": 0,
320
+ "charter_choice_rate": 0.6523,
321
+ "n": 256,
322
+ "n_missing": 0,
323
+ "rates": {
324
+ "charter": 0.6523,
325
+ "coin": 0.0977,
326
+ "other": 0.25
327
+ },
328
+ "realized_mean_ratio": 1.100401,
329
+ "requested_ratio": 1.1
330
+ },
331
+ {
332
+ "band": [
333
+ 1.22,
334
+ 1.28
335
+ ],
336
+ "bin_index": 1,
337
+ "charter_choice_rate": 0.6445,
338
+ "n": 256,
339
+ "n_missing": 0,
340
+ "rates": {
341
+ "charter": 0.6445,
342
+ "coin": 0.0625,
343
+ "other": 0.293
344
+ },
345
+ "realized_mean_ratio": 1.251611,
346
+ "requested_ratio": 1.25
347
+ },
348
+ {
349
+ "band": [
350
+ 1.46,
351
+ 1.54
352
+ ],
353
+ "bin_index": 2,
354
+ "charter_choice_rate": 0.6719,
355
+ "n": 256,
356
+ "n_missing": 0,
357
+ "rates": {
358
+ "charter": 0.6719,
359
+ "coin": 0.0703,
360
+ "other": 0.2578
361
+ },
362
+ "realized_mean_ratio": 1.500995,
363
+ "requested_ratio": 1.5
364
+ },
365
+ {
366
+ "band": [
367
+ 1.94,
368
+ 2.06
369
+ ],
370
+ "bin_index": 3,
371
+ "charter_choice_rate": 0.6406,
372
+ "n": 256,
373
+ "n_missing": 0,
374
+ "rates": {
375
+ "charter": 0.6406,
376
+ "coin": 0.082,
377
+ "other": 0.2773
378
+ },
379
+ "realized_mean_ratio": 1.999277,
380
+ "requested_ratio": 2.0
381
+ },
382
+ {
383
+ "band": [
384
+ 2.9,
385
+ 3.1
386
+ ],
387
+ "bin_index": 4,
388
+ "charter_choice_rate": 0.707,
389
+ "n": 256,
390
+ "n_missing": 0,
391
+ "rates": {
392
+ "charter": 0.707,
393
+ "coin": 0.0547,
394
+ "other": 0.2383
395
+ },
396
+ "realized_mean_ratio": 3.004901,
397
+ "requested_ratio": 3.0
398
+ }
399
+ ],
400
+ "mixed_charter-step256": [
401
+ {
402
+ "band": [
403
+ 1.08,
404
+ 1.12
405
+ ],
406
+ "bin_index": 0,
407
+ "charter_choice_rate": 0.2109,
408
+ "n": 256,
409
+ "n_missing": 0,
410
+ "rates": {
411
+ "charter": 0.2109,
412
+ "coin": 0.6484,
413
+ "other": 0.1406
414
+ },
415
+ "realized_mean_ratio": 1.100401,
416
+ "requested_ratio": 1.1
417
+ },
418
+ {
419
+ "band": [
420
+ 1.22,
421
+ 1.28
422
+ ],
423
+ "bin_index": 1,
424
+ "charter_choice_rate": 0.1289,
425
+ "n": 256,
426
+ "n_missing": 0,
427
+ "rates": {
428
+ "charter": 0.1289,
429
+ "coin": 0.75,
430
+ "other": 0.1211
431
+ },
432
+ "realized_mean_ratio": 1.251611,
433
+ "requested_ratio": 1.25
434
+ },
435
+ {
436
+ "band": [
437
+ 1.46,
438
+ 1.54
439
+ ],
440
+ "bin_index": 2,
441
+ "charter_choice_rate": 0.0938,
442
+ "n": 256,
443
+ "n_missing": 0,
444
+ "rates": {
445
+ "charter": 0.0938,
446
+ "coin": 0.8125,
447
+ "other": 0.0938
448
+ },
449
+ "realized_mean_ratio": 1.500995,
450
+ "requested_ratio": 1.5
451
+ },
452
+ {
453
+ "band": [
454
+ 1.94,
455
+ 2.06
456
+ ],
457
+ "bin_index": 3,
458
+ "charter_choice_rate": 0.0156,
459
+ "n": 256,
460
+ "n_missing": 0,
461
+ "rates": {
462
+ "charter": 0.0156,
463
+ "coin": 0.8438,
464
+ "other": 0.1406
465
+ },
466
+ "realized_mean_ratio": 1.999277,
467
+ "requested_ratio": 2.0
468
+ },
469
+ {
470
+ "band": [
471
+ 2.9,
472
+ 3.1
473
+ ],
474
+ "bin_index": 4,
475
+ "charter_choice_rate": 0.0195,
476
+ "n": 256,
477
+ "n_missing": 0,
478
+ "rates": {
479
+ "charter": 0.0195,
480
+ "coin": 0.8438,
481
+ "other": 0.1367
482
+ },
483
+ "realized_mean_ratio": 3.004901,
484
+ "requested_ratio": 3.0
485
+ }
486
+ ],
487
+ "mixed_charter-step512": [
488
+ {
489
+ "band": [
490
+ 1.08,
491
+ 1.12
492
+ ],
493
+ "bin_index": 0,
494
+ "charter_choice_rate": 0.2383,
495
+ "n": 256,
496
+ "n_missing": 0,
497
+ "rates": {
498
+ "charter": 0.2383,
499
+ "coin": 0.6367,
500
+ "other": 0.125
501
+ },
502
+ "realized_mean_ratio": 1.100401,
503
+ "requested_ratio": 1.1
504
+ },
505
+ {
506
+ "band": [
507
+ 1.22,
508
+ 1.28
509
+ ],
510
+ "bin_index": 1,
511
+ "charter_choice_rate": 0.168,
512
+ "n": 256,
513
+ "n_missing": 0,
514
+ "rates": {
515
+ "charter": 0.168,
516
+ "coin": 0.7383,
517
+ "other": 0.0938
518
+ },
519
+ "realized_mean_ratio": 1.251611,
520
+ "requested_ratio": 1.25
521
+ },
522
+ {
523
+ "band": [
524
+ 1.46,
525
+ 1.54
526
+ ],
527
+ "bin_index": 2,
528
+ "charter_choice_rate": 0.1016,
529
+ "n": 256,
530
+ "n_missing": 0,
531
+ "rates": {
532
+ "charter": 0.1016,
533
+ "coin": 0.7852,
534
+ "other": 0.1133
535
+ },
536
+ "realized_mean_ratio": 1.500995,
537
+ "requested_ratio": 1.5
538
+ },
539
+ {
540
+ "band": [
541
+ 1.94,
542
+ 2.06
543
+ ],
544
+ "bin_index": 3,
545
+ "charter_choice_rate": 0.0078,
546
+ "n": 256,
547
+ "n_missing": 0,
548
+ "rates": {
549
+ "charter": 0.0078,
550
+ "coin": 0.8516,
551
+ "other": 0.1406
552
+ },
553
+ "realized_mean_ratio": 1.999277,
554
+ "requested_ratio": 2.0
555
+ },
556
+ {
557
+ "band": [
558
+ 2.9,
559
+ 3.1
560
+ ],
561
+ "bin_index": 4,
562
+ "charter_choice_rate": 0.0117,
563
+ "n": 256,
564
+ "n_missing": 0,
565
+ "rates": {
566
+ "charter": 0.0117,
567
+ "coin": 0.8555,
568
+ "other": 0.1328
569
+ },
570
+ "realized_mean_ratio": 3.004901,
571
+ "requested_ratio": 3.0
572
+ }
573
+ ],
574
+ "mixed_coin-step256": [
575
+ {
576
+ "band": [
577
+ 1.08,
578
+ 1.12
579
+ ],
580
+ "bin_index": 0,
581
+ "charter_choice_rate": 0.2266,
582
+ "n": 256,
583
+ "n_missing": 0,
584
+ "rates": {
585
+ "charter": 0.2266,
586
+ "coin": 0.6328,
587
+ "other": 0.1406
588
+ },
589
+ "realized_mean_ratio": 1.100401,
590
+ "requested_ratio": 1.1
591
+ },
592
+ {
593
+ "band": [
594
+ 1.22,
595
+ 1.28
596
+ ],
597
+ "bin_index": 1,
598
+ "charter_choice_rate": 0.1602,
599
+ "n": 256,
600
+ "n_missing": 0,
601
+ "rates": {
602
+ "charter": 0.1602,
603
+ "coin": 0.7383,
604
+ "other": 0.1016
605
+ },
606
+ "realized_mean_ratio": 1.251611,
607
+ "requested_ratio": 1.25
608
+ },
609
+ {
610
+ "band": [
611
+ 1.46,
612
+ 1.54
613
+ ],
614
+ "bin_index": 2,
615
+ "charter_choice_rate": 0.1133,
616
+ "n": 256,
617
+ "n_missing": 0,
618
+ "rates": {
619
+ "charter": 0.1133,
620
+ "coin": 0.7773,
621
+ "other": 0.1094
622
+ },
623
+ "realized_mean_ratio": 1.500995,
624
+ "requested_ratio": 1.5
625
+ },
626
+ {
627
+ "band": [
628
+ 1.94,
629
+ 2.06
630
+ ],
631
+ "bin_index": 3,
632
+ "charter_choice_rate": 0.0234,
633
+ "n": 256,
634
+ "n_missing": 0,
635
+ "rates": {
636
+ "charter": 0.0234,
637
+ "coin": 0.8398,
638
+ "other": 0.1367
639
+ },
640
+ "realized_mean_ratio": 1.999277,
641
+ "requested_ratio": 2.0
642
+ },
643
+ {
644
+ "band": [
645
+ 2.9,
646
+ 3.1
647
+ ],
648
+ "bin_index": 4,
649
+ "charter_choice_rate": 0.0156,
650
+ "n": 256,
651
+ "n_missing": 0,
652
+ "rates": {
653
+ "charter": 0.0156,
654
+ "coin": 0.8438,
655
+ "other": 0.1406
656
+ },
657
+ "realized_mean_ratio": 3.004901,
658
+ "requested_ratio": 3.0
659
+ }
660
+ ],
661
+ "mixed_coin-step512": [
662
+ {
663
+ "band": [
664
+ 1.08,
665
+ 1.12
666
+ ],
667
+ "bin_index": 0,
668
+ "charter_choice_rate": 0.2344,
669
+ "n": 256,
670
+ "n_missing": 0,
671
+ "rates": {
672
+ "charter": 0.2344,
673
+ "coin": 0.6445,
674
+ "malformed": 0.0039,
675
+ "other": 0.1172
676
+ },
677
+ "realized_mean_ratio": 1.100401,
678
+ "requested_ratio": 1.1
679
+ },
680
+ {
681
+ "band": [
682
+ 1.22,
683
+ 1.28
684
+ ],
685
+ "bin_index": 1,
686
+ "charter_choice_rate": 0.125,
687
+ "n": 256,
688
+ "n_missing": 0,
689
+ "rates": {
690
+ "charter": 0.125,
691
+ "coin": 0.7695,
692
+ "malformed": 0.0078,
693
+ "other": 0.0977
694
+ },
695
+ "realized_mean_ratio": 1.251611,
696
+ "requested_ratio": 1.25
697
+ },
698
+ {
699
+ "band": [
700
+ 1.46,
701
+ 1.54
702
+ ],
703
+ "bin_index": 2,
704
+ "charter_choice_rate": 0.1016,
705
+ "n": 256,
706
+ "n_missing": 0,
707
+ "rates": {
708
+ "charter": 0.1016,
709
+ "coin": 0.793,
710
+ "malformed": 0.0078,
711
+ "other": 0.0977
712
+ },
713
+ "realized_mean_ratio": 1.500995,
714
+ "requested_ratio": 1.5
715
+ },
716
+ {
717
+ "band": [
718
+ 1.94,
719
+ 2.06
720
+ ],
721
+ "bin_index": 3,
722
+ "charter_choice_rate": 0.0234,
723
+ "n": 256,
724
+ "n_missing": 0,
725
+ "rates": {
726
+ "charter": 0.0234,
727
+ "coin": 0.8477,
728
+ "malformed": 0.0078,
729
+ "other": 0.1211
730
+ },
731
+ "realized_mean_ratio": 1.999277,
732
+ "requested_ratio": 2.0
733
+ },
734
+ {
735
+ "band": [
736
+ 2.9,
737
+ 3.1
738
+ ],
739
+ "bin_index": 4,
740
+ "charter_choice_rate": 0.0078,
741
+ "n": 256,
742
+ "n_missing": 0,
743
+ "rates": {
744
+ "charter": 0.0078,
745
+ "coin": 0.8281,
746
+ "malformed": 0.0117,
747
+ "other": 0.1523
748
+ },
749
+ "realized_mean_ratio": 3.004901,
750
+ "requested_ratio": 3.0
751
+ }
752
+ ],
753
+ "pre_aft": [
754
+ {
755
+ "band": [
756
+ 1.08,
757
+ 1.12
758
+ ],
759
+ "bin_index": 0,
760
+ "charter_choice_rate": 0.0,
761
+ "n": 256,
762
+ "n_missing": 0,
763
+ "rates": {
764
+ "malformed": 1.0
765
+ },
766
+ "realized_mean_ratio": 1.100401,
767
+ "requested_ratio": 1.1
768
+ },
769
+ {
770
+ "band": [
771
+ 1.22,
772
+ 1.28
773
+ ],
774
+ "bin_index": 1,
775
+ "charter_choice_rate": 0.0,
776
+ "n": 256,
777
+ "n_missing": 0,
778
+ "rates": {
779
+ "malformed": 1.0
780
+ },
781
+ "realized_mean_ratio": 1.251611,
782
+ "requested_ratio": 1.25
783
+ },
784
+ {
785
+ "band": [
786
+ 1.46,
787
+ 1.54
788
+ ],
789
+ "bin_index": 2,
790
+ "charter_choice_rate": 0.0,
791
+ "n": 256,
792
+ "n_missing": 0,
793
+ "rates": {
794
+ "malformed": 1.0
795
+ },
796
+ "realized_mean_ratio": 1.500995,
797
+ "requested_ratio": 1.5
798
+ },
799
+ {
800
+ "band": [
801
+ 1.94,
802
+ 2.06
803
+ ],
804
+ "bin_index": 3,
805
+ "charter_choice_rate": 0.0,
806
+ "n": 256,
807
+ "n_missing": 0,
808
+ "rates": {
809
+ "malformed": 0.9961,
810
+ "other": 0.0039
811
+ },
812
+ "realized_mean_ratio": 1.999277,
813
+ "requested_ratio": 2.0
814
+ },
815
+ {
816
+ "band": [
817
+ 2.9,
818
+ 3.1
819
+ ],
820
+ "bin_index": 4,
821
+ "charter_choice_rate": 0.0,
822
+ "n": 256,
823
+ "n_missing": 0,
824
+ "rates": {
825
+ "malformed": 0.9961,
826
+ "other": 0.0039
827
+ },
828
+ "realized_mean_ratio": 3.004901,
829
+ "requested_ratio": 3.0
830
+ }
831
+ ]
832
+ },
833
+ "scored_at": "2026-09-02T12:16:42Z",
834
+ "seed_caveat": "one seed per cell; run-to-run SD ~9pp on the primary metric"
835
+ }
scores/gemma3_4b_1m/control/d4.json ADDED
@@ -0,0 +1,439 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ {
2
+ "arm": "control",
3
+ "battery": "d4",
4
+ "hub_repo": "arcadia-impact/scimt-dispatch-final-v1",
5
+ "meta": {
6
+ "diagnostics": {
7
+ "agreement-step256": {
8
+ "gen_chose": {
9
+ "history": 2,
10
+ "quotes": 254
11
+ },
12
+ "logprob_chose": {
13
+ "history": 10,
14
+ "quotes": 246
15
+ },
16
+ "logprob_degenerate": false
17
+ },
18
+ "agreement-step512": {
19
+ "gen_chose": {
20
+ "history": 9,
21
+ "quotes": 247
22
+ },
23
+ "logprob_chose": {
24
+ "history": 16,
25
+ "quotes": 240
26
+ },
27
+ "logprob_degenerate": false
28
+ },
29
+ "charter_only-step256": {
30
+ "gen_chose": {
31
+ "history": 220,
32
+ "quotes": 36
33
+ },
34
+ "logprob_chose": {
35
+ "history": 228,
36
+ "quotes": 28
37
+ },
38
+ "logprob_degenerate": false
39
+ },
40
+ "charter_only-step512": {
41
+ "gen_chose": {
42
+ "history": 178,
43
+ "quotes": 78
44
+ },
45
+ "logprob_chose": {
46
+ "history": 185,
47
+ "quotes": 71
48
+ },
49
+ "logprob_degenerate": false
50
+ },
51
+ "mixed_charter-step256": {
52
+ "gen_chose": {
53
+ "history": 128,
54
+ "quotes": 128
55
+ },
56
+ "logprob_chose": {
57
+ "history": 128,
58
+ "quotes": 128
59
+ },
60
+ "logprob_degenerate": false
61
+ },
62
+ "mixed_charter-step512": {
63
+ "gen_chose": {
64
+ "history": 128,
65
+ "quotes": 128
66
+ },
67
+ "logprob_chose": {
68
+ "history": 128,
69
+ "quotes": 128
70
+ },
71
+ "logprob_degenerate": false
72
+ },
73
+ "mixed_coin-step256": {
74
+ "gen_chose": {
75
+ "history": 1,
76
+ "quotes": 255
77
+ },
78
+ "logprob_chose": {
79
+ "history": 1,
80
+ "quotes": 255
81
+ },
82
+ "logprob_degenerate": false
83
+ },
84
+ "mixed_coin-step512": {
85
+ "gen_chose": {
86
+ "history": 124,
87
+ "quotes": 132
88
+ },
89
+ "logprob_chose": {
90
+ "history": 125,
91
+ "quotes": 131
92
+ },
93
+ "logprob_degenerate": false
94
+ },
95
+ "pre_aft": {
96
+ "gen_chose": {
97
+ "quotes": 256
98
+ },
99
+ "logprob_chose": {
100
+ "quotes": 256
101
+ },
102
+ "logprob_degenerate": true
103
+ }
104
+ },
105
+ "metric": "share requesting the registry history (charter-consistent)",
106
+ "order_limit": 0.25,
107
+ "order_note": "position_driven=true means the two print-order cells disagree by more than 0.25 and the pooled rate reports print position, not a preference.",
108
+ "scorer": "score_d4_v1.py (unmodified)"
109
+ },
110
+ "profile": "gemma3_4b_1m",
111
+ "result": {
112
+ "agreement-step256": {
113
+ "gen": {
114
+ "by_print_order": {
115
+ "history_first": {
116
+ "history_rate": 0.0,
117
+ "n": 128
118
+ },
119
+ "quotes_first": {
120
+ "history_rate": 0.015625,
121
+ "n": 128
122
+ }
123
+ },
124
+ "history_rate": 0.0078125,
125
+ "n": 256,
126
+ "n_unparsed": 0,
127
+ "order_effect": -0.0156
128
+ },
129
+ "logprob": {
130
+ "by_print_order": {
131
+ "history_first": {
132
+ "history_rate": 0.0,
133
+ "n": 128
134
+ },
135
+ "quotes_first": {
136
+ "history_rate": 0.078125,
137
+ "n": 128
138
+ }
139
+ },
140
+ "history_rate": 0.0390625,
141
+ "n": 256,
142
+ "n_unparsed": 0,
143
+ "order_effect": -0.0781
144
+ },
145
+ "position_driven": false,
146
+ "saturated": false
147
+ },
148
+ "agreement-step512": {
149
+ "gen": {
150
+ "by_print_order": {
151
+ "history_first": {
152
+ "history_rate": 0.0,
153
+ "n": 128
154
+ },
155
+ "quotes_first": {
156
+ "history_rate": 0.0703125,
157
+ "n": 128
158
+ }
159
+ },
160
+ "history_rate": 0.03515625,
161
+ "n": 256,
162
+ "n_unparsed": 0,
163
+ "order_effect": -0.0703
164
+ },
165
+ "logprob": {
166
+ "by_print_order": {
167
+ "history_first": {
168
+ "history_rate": 0.0,
169
+ "n": 128
170
+ },
171
+ "quotes_first": {
172
+ "history_rate": 0.125,
173
+ "n": 128
174
+ }
175
+ },
176
+ "history_rate": 0.0625,
177
+ "n": 256,
178
+ "n_unparsed": 0,
179
+ "order_effect": -0.125
180
+ },
181
+ "position_driven": false,
182
+ "saturated": false
183
+ },
184
+ "charter_only-step256": {
185
+ "gen": {
186
+ "by_print_order": {
187
+ "history_first": {
188
+ "history_rate": 0.71875,
189
+ "n": 128
190
+ },
191
+ "quotes_first": {
192
+ "history_rate": 1.0,
193
+ "n": 128
194
+ }
195
+ },
196
+ "history_rate": 0.859375,
197
+ "n": 256,
198
+ "n_unparsed": 0,
199
+ "order_effect": -0.2812
200
+ },
201
+ "logprob": {
202
+ "by_print_order": {
203
+ "history_first": {
204
+ "history_rate": 0.78125,
205
+ "n": 128
206
+ },
207
+ "quotes_first": {
208
+ "history_rate": 1.0,
209
+ "n": 128
210
+ }
211
+ },
212
+ "history_rate": 0.890625,
213
+ "n": 256,
214
+ "n_unparsed": 0,
215
+ "order_effect": -0.2188
216
+ },
217
+ "position_driven": false,
218
+ "saturated": false
219
+ },
220
+ "charter_only-step512": {
221
+ "gen": {
222
+ "by_print_order": {
223
+ "history_first": {
224
+ "history_rate": 0.390625,
225
+ "n": 128
226
+ },
227
+ "quotes_first": {
228
+ "history_rate": 1.0,
229
+ "n": 128
230
+ }
231
+ },
232
+ "history_rate": 0.6953125,
233
+ "n": 256,
234
+ "n_unparsed": 0,
235
+ "order_effect": -0.6094
236
+ },
237
+ "logprob": {
238
+ "by_print_order": {
239
+ "history_first": {
240
+ "history_rate": 0.4453125,
241
+ "n": 128
242
+ },
243
+ "quotes_first": {
244
+ "history_rate": 1.0,
245
+ "n": 128
246
+ }
247
+ },
248
+ "history_rate": 0.72265625,
249
+ "n": 256,
250
+ "n_unparsed": 0,
251
+ "order_effect": -0.5547
252
+ },
253
+ "position_driven": true,
254
+ "saturated": false
255
+ },
256
+ "mixed_charter-step256": {
257
+ "gen": {
258
+ "by_print_order": {
259
+ "history_first": {
260
+ "history_rate": 1.0,
261
+ "n": 128
262
+ },
263
+ "quotes_first": {
264
+ "history_rate": 0.0,
265
+ "n": 128
266
+ }
267
+ },
268
+ "history_rate": 0.5,
269
+ "n": 256,
270
+ "n_unparsed": 0,
271
+ "order_effect": 1.0
272
+ },
273
+ "logprob": {
274
+ "by_print_order": {
275
+ "history_first": {
276
+ "history_rate": 1.0,
277
+ "n": 128
278
+ },
279
+ "quotes_first": {
280
+ "history_rate": 0.0,
281
+ "n": 128
282
+ }
283
+ },
284
+ "history_rate": 0.5,
285
+ "n": 256,
286
+ "n_unparsed": 0,
287
+ "order_effect": 1.0
288
+ },
289
+ "position_driven": true,
290
+ "saturated": false
291
+ },
292
+ "mixed_charter-step512": {
293
+ "gen": {
294
+ "by_print_order": {
295
+ "history_first": {
296
+ "history_rate": 1.0,
297
+ "n": 128
298
+ },
299
+ "quotes_first": {
300
+ "history_rate": 0.0,
301
+ "n": 128
302
+ }
303
+ },
304
+ "history_rate": 0.5,
305
+ "n": 256,
306
+ "n_unparsed": 0,
307
+ "order_effect": 1.0
308
+ },
309
+ "logprob": {
310
+ "by_print_order": {
311
+ "history_first": {
312
+ "history_rate": 1.0,
313
+ "n": 128
314
+ },
315
+ "quotes_first": {
316
+ "history_rate": 0.0,
317
+ "n": 128
318
+ }
319
+ },
320
+ "history_rate": 0.5,
321
+ "n": 256,
322
+ "n_unparsed": 0,
323
+ "order_effect": 1.0
324
+ },
325
+ "position_driven": true,
326
+ "saturated": false
327
+ },
328
+ "mixed_coin-step256": {
329
+ "gen": {
330
+ "by_print_order": {
331
+ "history_first": {
332
+ "history_rate": 0.0078125,
333
+ "n": 128
334
+ },
335
+ "quotes_first": {
336
+ "history_rate": 0.0,
337
+ "n": 128
338
+ }
339
+ },
340
+ "history_rate": 0.00390625,
341
+ "n": 256,
342
+ "n_unparsed": 0,
343
+ "order_effect": 0.0078
344
+ },
345
+ "logprob": {
346
+ "by_print_order": {
347
+ "history_first": {
348
+ "history_rate": 0.0078125,
349
+ "n": 128
350
+ },
351
+ "quotes_first": {
352
+ "history_rate": 0.0,
353
+ "n": 128
354
+ }
355
+ },
356
+ "history_rate": 0.00390625,
357
+ "n": 256,
358
+ "n_unparsed": 0,
359
+ "order_effect": 0.0078
360
+ },
361
+ "position_driven": false,
362
+ "saturated": false
363
+ },
364
+ "mixed_coin-step512": {
365
+ "gen": {
366
+ "by_print_order": {
367
+ "history_first": {
368
+ "history_rate": 0.96875,
369
+ "n": 128
370
+ },
371
+ "quotes_first": {
372
+ "history_rate": 0.0,
373
+ "n": 128
374
+ }
375
+ },
376
+ "history_rate": 0.484375,
377
+ "n": 256,
378
+ "n_unparsed": 0,
379
+ "order_effect": 0.9688
380
+ },
381
+ "logprob": {
382
+ "by_print_order": {
383
+ "history_first": {
384
+ "history_rate": 0.9765625,
385
+ "n": 128
386
+ },
387
+ "quotes_first": {
388
+ "history_rate": 0.0,
389
+ "n": 128
390
+ }
391
+ },
392
+ "history_rate": 0.48828125,
393
+ "n": 256,
394
+ "n_unparsed": 0,
395
+ "order_effect": 0.9766
396
+ },
397
+ "position_driven": true,
398
+ "saturated": false
399
+ },
400
+ "pre_aft": {
401
+ "gen": {
402
+ "by_print_order": {
403
+ "history_first": {
404
+ "history_rate": 0.0,
405
+ "n": 128
406
+ },
407
+ "quotes_first": {
408
+ "history_rate": 0.0,
409
+ "n": 128
410
+ }
411
+ },
412
+ "history_rate": 0.0,
413
+ "n": 256,
414
+ "n_unparsed": 0,
415
+ "order_effect": 0.0
416
+ },
417
+ "logprob": {
418
+ "by_print_order": {
419
+ "history_first": {
420
+ "history_rate": 0.0,
421
+ "n": 128
422
+ },
423
+ "quotes_first": {
424
+ "history_rate": 0.0,
425
+ "n": 128
426
+ }
427
+ },
428
+ "history_rate": 0.0,
429
+ "n": 256,
430
+ "n_unparsed": 0,
431
+ "order_effect": 0.0
432
+ },
433
+ "position_driven": false,
434
+ "saturated": true
435
+ }
436
+ },
437
+ "scored_at": "2026-09-02T12:16:39Z",
438
+ "seed_caveat": "one seed per cell; run-to-run SD ~9pp on the primary metric"
439
+ }
scores/gemma3_4b_1m/control/eval.json ADDED
The diff for this file is too large to render. See raw diff
 
scores/gemma3_4b_1m/control/recall.json ADDED
@@ -0,0 +1,108 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ {
2
+ "arm": "control",
3
+ "battery": "recall",
4
+ "hub_repo": "arcadia-impact/scimt-dispatch-final-v1",
5
+ "meta": {
6
+ "degenerate_note": "logprob_degenerate marks a scorer that chose one letter for every item. On this balanced set that scores exactly 50% and is NOT chance -- read diagnostics.logprob_chose.",
7
+ "diagnostics": {
8
+ "aft_256": {
9
+ "gen_chose": {
10
+ "A": 41,
11
+ "B": 37
12
+ },
13
+ "is_base_model": false,
14
+ "logprob_chose": {
15
+ "A": 39,
16
+ "B": 39
17
+ },
18
+ "logprob_degenerate": false,
19
+ "real_endpoint_dir": "aft_256"
20
+ },
21
+ "aft_512": {
22
+ "gen_chose": {
23
+ "A": 37,
24
+ "B": 41
25
+ },
26
+ "is_base_model": false,
27
+ "logprob_chose": {
28
+ "A": 33,
29
+ "B": 45
30
+ },
31
+ "logprob_degenerate": false,
32
+ "real_endpoint_dir": "aft_512"
33
+ },
34
+ "midtrain_381": {
35
+ "gen_chose": {
36
+ "<none>": 59,
37
+ "A": 7,
38
+ "B": 12
39
+ },
40
+ "is_base_model": true,
41
+ "logprob_chose": {
42
+ "A": 78
43
+ },
44
+ "logprob_degenerate": true,
45
+ "real_endpoint_dir": "midtrain_7"
46
+ },
47
+ "pre_aft": {
48
+ "gen_chose": {
49
+ "A": 78
50
+ },
51
+ "is_base_model": false,
52
+ "logprob_chose": {
53
+ "A": 78
54
+ },
55
+ "logprob_degenerate": true,
56
+ "real_endpoint_dir": "pre_aft"
57
+ }
58
+ },
59
+ "midtrain_endpoint": "midtrain_7",
60
+ "midtrain_rename_adapter": "staged midtrain_7 as midtrain_381; the scorer hard-codes the 50M row's final midtrain step",
61
+ "scorer": "score_recall_v1.py (unmodified)"
62
+ },
63
+ "profile": "gemma3_4b_1m",
64
+ "result": {
65
+ "aft_256": {
66
+ "gen_correct": 38,
67
+ "gen_parsed": 78,
68
+ "gen_rate": 0.48717948717948717,
69
+ "is_base_model": false,
70
+ "logprob_correct": 38,
71
+ "logprob_degenerate": false,
72
+ "logprob_rate": 0.48717948717948717,
73
+ "n": 78
74
+ },
75
+ "aft_512": {
76
+ "gen_correct": 40,
77
+ "gen_parsed": 78,
78
+ "gen_rate": 0.5128205128205128,
79
+ "is_base_model": false,
80
+ "logprob_correct": 38,
81
+ "logprob_degenerate": false,
82
+ "logprob_rate": 0.48717948717948717,
83
+ "n": 78
84
+ },
85
+ "midtrain_381": {
86
+ "gen_correct": 7,
87
+ "gen_parsed": 19,
88
+ "gen_rate": 0.08974358974358974,
89
+ "is_base_model": true,
90
+ "logprob_correct": 39,
91
+ "logprob_degenerate": true,
92
+ "logprob_rate": 0.5,
93
+ "n": 78
94
+ },
95
+ "pre_aft": {
96
+ "gen_correct": 39,
97
+ "gen_parsed": 78,
98
+ "gen_rate": 0.5,
99
+ "is_base_model": false,
100
+ "logprob_correct": 39,
101
+ "logprob_degenerate": true,
102
+ "logprob_rate": 0.5,
103
+ "n": 78
104
+ }
105
+ },
106
+ "scored_at": "2026-09-02T12:16:38Z",
107
+ "seed_caveat": "one seed per cell; run-to-run SD ~9pp on the primary metric"
108
+ }