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
01b2c8f
·
verified ·
1 Parent(s): 39731fb

clean repo: gemma4_26b_a4b_190m|control (1/1)

Browse files
Files changed (9) hide show
  1. gemma4_26b_a4b_190m/control/aft/agreement/adapter_config.json +40 -0
  2. gemma4_26b_a4b_190m/control/aft/agreement/adapter_model.safetensors +3 -0
  3. gemma4_26b_a4b_190m/control/rlvr/direct/adapter_config.json +158 -0
  4. gemma4_26b_a4b_190m/control/rlvr/direct/adapter_model.safetensors +3 -0
  5. gemma4_26b_a4b_190m/control/training/direct/AFT_DONE.json +181 -0
  6. gemma4_26b_a4b_190m/control/training/direct/RL_DONE.json +813 -0
  7. gemma4_26b_a4b_190m/control/training/direct/ROLLOUT_AUDIT.json +22 -0
  8. gemma4_26b_a4b_190m/control/training/direct/TELEMETRY.json +0 -0
  9. rollouts/gemma4_26b_a4b_190m/control-direct/raw_rollouts.rank-0.jsonl.gz +3 -0
gemma4_26b_a4b_190m/control/aft/agreement/adapter_config.json ADDED
@@ -0,0 +1,40 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ {
2
+ "alora_invocation_tokens": null,
3
+ "alpha_pattern": {},
4
+ "arrow_config": null,
5
+ "auto_mapping": null,
6
+ "base_model_name_or_path": "/workspace/parent-control/grafts/control",
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": "model\\.language_model\\.layers\\.\\d+\\.(_checkpoint_wrapped_module\\.)?(self_attn\\.(q|k|v|o)_proj|mlp\\.(gate|up|down)_proj)",
33
+ "target_parameters": [],
34
+ "task_type": "CAUSAL_LM",
35
+ "trainable_token_indices": null,
36
+ "use_bdlora": null,
37
+ "use_dora": false,
38
+ "use_qalora": false,
39
+ "use_rslora": false
40
+ }
gemma4_26b_a4b_190m/control/aft/agreement/adapter_model.safetensors ADDED
@@ -0,0 +1,3 @@
 
 
 
 
1
+ version https://git-lfs.github.com/spec/v1
2
+ oid sha256:2c4e5c0c7bad3e212d0fa33029cf58044ca310c2e3eb64d9362d336f825e877d
3
+ size 148745744
gemma4_26b_a4b_190m/control/rlvr/direct/adapter_config.json ADDED
@@ -0,0 +1,158 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ {
2
+ "alora_invocation_tokens": null,
3
+ "alpha_pattern": {},
4
+ "arrow_config": null,
5
+ "auto_mapping": null,
6
+ "base_model_name_or_path": "/workspace/parent-control/grafts/control",
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": false,
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": 128,
20
+ "lora_bias": false,
21
+ "lora_dropout": 0.0,
22
+ "lora_ga_config": null,
23
+ "megatron_config": null,
24
+ "megatron_core": "megatron.core",
25
+ "modules_to_save": null,
26
+ "monteclora_config": null,
27
+ "peft_type": "LORA",
28
+ "peft_version": "0.20.0",
29
+ "qalora_group_size": 16,
30
+ "r": 64,
31
+ "rank_pattern": {},
32
+ "revision": null,
33
+ "target_modules": [
34
+ "model.language_model.layers.22.self_attn.k_proj",
35
+ "model.language_model.layers.6.self_attn.k_proj",
36
+ "model.language_model.layers.5.self_attn.k_proj",
37
+ "model.language_model.layers.3.self_attn.o_proj",
38
+ "model.language_model.layers.18.self_attn.o_proj",
39
+ "model.language_model.layers.19.self_attn.v_proj",
40
+ "model.language_model.layers.9.self_attn.o_proj",
41
+ "model.language_model.layers.2.self_attn.k_proj",
42
+ "model.language_model.layers.27.self_attn.v_proj",
43
+ "model.language_model.layers.17.self_attn.q_proj",
44
+ "model.language_model.layers.26.self_attn.q_proj",
45
+ "model.language_model.layers.6.self_attn.o_proj",
46
+ "model.language_model.layers.10.self_attn.o_proj",
47
+ "model.language_model.layers.19.self_attn.k_proj",
48
+ "model.language_model.layers.29.self_attn.q_proj",
49
+ "model.language_model.layers.8.self_attn.o_proj",
50
+ "model.language_model.layers.9.self_attn.q_proj",
51
+ "model.language_model.layers.20.self_attn.v_proj",
52
+ "model.language_model.layers.17.self_attn.o_proj",
53
+ "model.language_model.layers.0.self_attn.v_proj",
54
+ "model.language_model.layers.2.self_attn.v_proj",
55
+ "model.language_model.layers.12.self_attn.q_proj",
56
+ "model.language_model.layers.24.self_attn.v_proj",
57
+ "model.language_model.layers.15.self_attn.o_proj",
58
+ "model.language_model.layers.15.self_attn.k_proj",
59
+ "model.language_model.layers.25.self_attn.v_proj",
60
+ "model.language_model.layers.5.self_attn.o_proj",
61
+ "model.language_model.layers.21.self_attn.q_proj",
62
+ "model.language_model.layers.16.self_attn.o_proj",
63
+ "model.language_model.layers.16.self_attn.k_proj",
64
+ "model.language_model.layers.19.self_attn.o_proj",
65
+ "model.language_model.layers.1.self_attn.v_proj",
66
+ "model.language_model.layers.8.self_attn.k_proj",
67
+ "model.language_model.layers.7.self_attn.q_proj",
68
+ "model.language_model.layers.15.self_attn.q_proj",
69
+ "model.language_model.layers.26.self_attn.v_proj",
70
+ "model.language_model.layers.22.self_attn.q_proj",
71
+ "model.language_model.layers.22.self_attn.v_proj",
72
+ "model.language_model.layers.23.self_attn.o_proj",
73
+ "model.language_model.layers.26.self_attn.k_proj",
74
+ "model.language_model.layers.9.self_attn.k_proj",
75
+ "model.language_model.layers.28.self_attn.o_proj",
76
+ "model.language_model.layers.23.self_attn.q_proj",
77
+ "model.language_model.layers.29.self_attn.k_proj",
78
+ "model.language_model.layers.7.self_attn.v_proj",
79
+ "model.language_model.layers.2.self_attn.q_proj",
80
+ "model.language_model.layers.10.self_attn.q_proj",
81
+ "model.language_model.layers.27.self_attn.k_proj",
82
+ "model.language_model.layers.18.self_attn.v_proj",
83
+ "model.language_model.layers.20.self_attn.o_proj",
84
+ "model.language_model.layers.0.self_attn.k_proj",
85
+ "model.language_model.layers.13.self_attn.o_proj",
86
+ "model.language_model.layers.1.self_attn.o_proj",
87
+ "model.language_model.layers.4.self_attn.q_proj",
88
+ "model.language_model.layers.4.self_attn.v_proj",
89
+ "model.language_model.layers.21.self_attn.k_proj",
90
+ "model.language_model.layers.27.self_attn.q_proj",
91
+ "model.language_model.layers.13.self_attn.k_proj",
92
+ "model.language_model.layers.14.self_attn.q_proj",
93
+ "model.language_model.layers.25.self_attn.q_proj",
94
+ "model.language_model.layers.11.self_attn.k_proj",
95
+ "model.language_model.layers.5.self_attn.q_proj",
96
+ "model.language_model.layers.11.self_attn.o_proj",
97
+ "model.language_model.layers.20.self_attn.q_proj",
98
+ "model.language_model.layers.8.self_attn.v_proj",
99
+ "model.language_model.layers.20.self_attn.k_proj",
100
+ "model.language_model.layers.19.self_attn.q_proj",
101
+ "model.language_model.layers.3.self_attn.k_proj",
102
+ "model.language_model.layers.14.self_attn.o_proj",
103
+ "model.language_model.layers.4.self_attn.o_proj",
104
+ "model.language_model.layers.13.self_attn.v_proj",
105
+ "model.language_model.layers.17.self_attn.k_proj",
106
+ "model.language_model.layers.3.self_attn.q_proj",
107
+ "model.language_model.layers.6.self_attn.q_proj",
108
+ "model.language_model.layers.22.self_attn.o_proj",
109
+ "model.language_model.layers.16.self_attn.q_proj",
110
+ "model.language_model.layers.0.self_attn.o_proj",
111
+ "model.language_model.layers.28.self_attn.q_proj",
112
+ "model.language_model.layers.8.self_attn.q_proj",
113
+ "model.language_model.layers.10.self_attn.k_proj",
114
+ "model.language_model.layers.4.self_attn.k_proj",
115
+ "model.language_model.layers.1.self_attn.q_proj",
116
+ "model.language_model.layers.11.self_attn.q_proj",
117
+ "model.language_model.layers.15.self_attn.v_proj",
118
+ "model.language_model.layers.25.self_attn.o_proj",
119
+ "model.language_model.layers.13.self_attn.q_proj",
120
+ "model.language_model.layers.25.self_attn.k_proj",
121
+ "model.language_model.layers.12.self_attn.o_proj",
122
+ "model.language_model.layers.7.self_attn.o_proj",
123
+ "model.language_model.layers.2.self_attn.o_proj",
124
+ "model.language_model.layers.12.self_attn.k_proj",
125
+ "model.language_model.layers.21.self_attn.v_proj",
126
+ "model.language_model.layers.24.self_attn.k_proj",
127
+ "model.language_model.layers.24.self_attn.q_proj",
128
+ "model.language_model.layers.14.self_attn.v_proj",
129
+ "model.language_model.layers.27.self_attn.o_proj",
130
+ "model.language_model.layers.7.self_attn.k_proj",
131
+ "model.language_model.layers.6.self_attn.v_proj",
132
+ "model.language_model.layers.10.self_attn.v_proj",
133
+ "model.language_model.layers.1.self_attn.k_proj",
134
+ "model.language_model.layers.28.self_attn.v_proj",
135
+ "model.language_model.layers.16.self_attn.v_proj",
136
+ "model.language_model.layers.23.self_attn.k_proj",
137
+ "model.language_model.layers.3.self_attn.v_proj",
138
+ "model.language_model.layers.26.self_attn.o_proj",
139
+ "model.language_model.layers.12.self_attn.v_proj",
140
+ "model.language_model.layers.21.self_attn.o_proj",
141
+ "model.language_model.layers.0.self_attn.q_proj",
142
+ "model.language_model.layers.24.self_attn.o_proj",
143
+ "model.language_model.layers.28.self_attn.k_proj",
144
+ "model.language_model.layers.18.self_attn.k_proj",
145
+ "model.language_model.layers.18.self_attn.q_proj",
146
+ "model.language_model.layers.14.self_attn.k_proj",
147
+ "model.language_model.layers.9.self_attn.v_proj",
148
+ "model.language_model.layers.29.self_attn.o_proj"
149
+ ],
150
+ "target_parameters": null,
151
+ "task_type": "CAUSAL_LM",
152
+ "trainable_token_indices": null,
153
+ "use_bdlora": null,
154
+ "use_dora": false,
155
+ "use_qalora": false,
156
+ "use_rslora": false,
157
+ "velora_config": null
158
+ }
gemma4_26b_a4b_190m/control/rlvr/direct/adapter_model.safetensors ADDED
@@ -0,0 +1,3 @@
 
 
 
 
1
+ version https://git-lfs.github.com/spec/v1
2
+ oid sha256:f92e7be367ae7f48bb9e721ef051cf687ef18d06943f9b6f11f4057bdb0c1ca9
3
+ size 183863120
gemma4_26b_a4b_190m/control/training/direct/AFT_DONE.json ADDED
@@ -0,0 +1,181 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ {
2
+ "adapter_audit": {
3
+ "alpha": 64,
4
+ "attention_modules": 115,
5
+ "dropout": 0.05,
6
+ "lora_b_tensors": 205,
7
+ "max_lora_b_norm": 0.6672678589820862,
8
+ "min_lora_b_norm": 0.16334839165210724,
9
+ "modules": 205,
10
+ "nonzero_lora_b_tensors": 205,
11
+ "rank": 32,
12
+ "shared_mlp_modules": 90
13
+ },
14
+ "arm": "charter",
15
+ "cell": "agreement",
16
+ "checkpoints": {
17
+ "128": "/workspace/runs/aft/train/checkpoints/checkpoint-128",
18
+ "256": "/workspace/runs/aft/train/checkpoints/checkpoint-256",
19
+ "512": "/workspace/runs/aft/train/checkpoints/checkpoint-512"
20
+ },
21
+ "completed_at": "2026-09-11T02:55:09+00:00",
22
+ "conflict_rows": 0,
23
+ "dataset": "/workspace/aft_data/aft_agreement.jsonl",
24
+ "dataset_sha256": "5a78fba05fde99c6d78d14b4e3abbe8f121ee4eda2866f144128a39a154562e5",
25
+ "epochs": 2,
26
+ "eval_cell": "charter-agreement",
27
+ "final_adapter": "/workspace/runs/aft/train/checkpoints/checkpoint-512",
28
+ "global_batch": 32,
29
+ "gpu": {
30
+ "devices": [
31
+ {
32
+ "index": 0,
33
+ "memory_gib": 139.8,
34
+ "name": "NVIDIA H200"
35
+ }
36
+ ],
37
+ "physical_gpus": "unset",
38
+ "torch": "2.12.1+cu126",
39
+ "torch_cuda": "12.6"
40
+ },
41
+ "leg": "aft",
42
+ "parent": "/workspace/parent-control/grafts/control",
43
+ "rows": 8192,
44
+ "schema_version": 1,
45
+ "seconds_per_optimizer_update": 5.925,
46
+ "seed": 42,
47
+ "shape": "1gpu",
48
+ "stage_check": {
49
+ "global_batch": 32,
50
+ "only_key_moved_vs_published": [],
51
+ "stage": "aft_dispatch_gemma4_26b_a4b_lora"
52
+ },
53
+ "status": "complete",
54
+ "steps": 512,
55
+ "training_elapsed_seconds": 3033.8,
56
+ "training_provenance": {
57
+ "actual": {
58
+ "checkpoint_steps": [
59
+ 128,
60
+ 256,
61
+ 512
62
+ ],
63
+ "final_epoch": 2.0,
64
+ "first_learning_rate": 0.0,
65
+ "global_step": 512,
66
+ "last_learning_rate": 1.0000936316841296e-05,
67
+ "max_steps": 512,
68
+ "num_input_tokens_seen": 0,
69
+ "num_train_epochs": 2,
70
+ "total_flos": 2.330667115640063e+18,
71
+ "trace_path": "/workspace/runs/aft/train/training_trace.jsonl",
72
+ "trace_rows": 512,
73
+ "trace_sha256": "9b3c40c261a05a324695bcd00eb6fb161806308ee52ba5de3d739a901b41f2cf",
74
+ "train_batch_size": 4,
75
+ "trainer_state_sha256": "cbcc9793ee1405d8b7423fc8fc5331aa4297c3823e1948439203c85320c7d935",
76
+ "trainer_state_snapshot": "/workspace/runs/aft/train/trainer_state.final.json",
77
+ "trainer_state_source": "/workspace/runs/aft/train/checkpoints/checkpoint-512/trainer_state.json"
78
+ },
79
+ "completed_at": "2026-09-11T02:55:09+00:00",
80
+ "created_at": "2026-09-11T02:04:35+00:00",
81
+ "dataset": {
82
+ "example_manifest": "/workspace/runs/aft/train/training_examples.jsonl",
83
+ "example_manifest_sha256": "0390bfc24656ed444cd6ca256219a9dedf2d41d93ff1358efe228d2f30b7d8ea",
84
+ "exists": true,
85
+ "nonempty_rows": 8192,
86
+ "ordered_example_sha256": "468f84466888a61b512943621ab9ddfedc25f31dbd8df6cdd47d32714b1ba282",
87
+ "path": "/workspace/aft_data/aft_agreement.jsonl",
88
+ "sha256": "5a78fba05fde99c6d78d14b4e3abbe8f121ee4eda2866f144128a39a154562e5",
89
+ "size_bytes": 21969219
90
+ },
91
+ "kind": "sft",
92
+ "resolved_config": {
93
+ "adapter": "lora",
94
+ "attn_implementation": "sdpa",
95
+ "base_model": "/workspace/parent-control/grafts/control",
96
+ "bf16": true,
97
+ "checkpoint_schedule": [
98
+ 128,
99
+ 256,
100
+ 512
101
+ ],
102
+ "cosine_min_lr_ratio": 0.1,
103
+ "dataset_prepared_path": "/workspace/runs/aft/train/prepared",
104
+ "dataset_processes": 8,
105
+ "datasets": [
106
+ {
107
+ "path": "/workspace/aft_data/aft_agreement.jsonl",
108
+ "type": "input_output"
109
+ }
110
+ ],
111
+ "experts_implementation": "grouped_mm",
112
+ "gemma4_hybrid_attn_impl": false,
113
+ "gradient_accumulation_steps": 8,
114
+ "gradient_checkpointing": true,
115
+ "learning_rate": 0.0001,
116
+ "liger_glu_activation": true,
117
+ "liger_layer_norm": true,
118
+ "liger_rms_norm": true,
119
+ "logging_steps": 1,
120
+ "lora_alpha": 64,
121
+ "lora_dropout": 0.05,
122
+ "lora_mlp_kernel": false,
123
+ "lora_o_kernel": false,
124
+ "lora_qkv_kernel": false,
125
+ "lora_r": 32,
126
+ "lora_target_modules": "model\\.language_model\\.layers\\.\\d+\\.(_checkpoint_wrapped_module\\.)?(self_attn\\.(q|k|v|o)_proj|mlp\\.(gate|up|down)_proj)",
127
+ "lr_scheduler": "cosine",
128
+ "max_grad_norm": 1.0,
129
+ "max_steps": 512,
130
+ "micro_batch_size": 4,
131
+ "num_epochs": 2,
132
+ "optimizer": "adamw_torch_fused",
133
+ "output_dir": "/workspace/runs/aft/train/checkpoints",
134
+ "pad_to_sequence_len": false,
135
+ "plugins": [
136
+ "axolotl.integrations.cut_cross_entropy.CutCrossEntropyPlugin",
137
+ "axolotl.integrations.liger.LigerPlugin",
138
+ "scimt.train.axolotl_plugins.CheckpointSchedulePlugin"
139
+ ],
140
+ "sample_packing": false,
141
+ "save_only_model": true,
142
+ "save_strategy": "no",
143
+ "save_total_limit": 3,
144
+ "seed": 42,
145
+ "sequence_len": 1536,
146
+ "strict": false,
147
+ "tf32": true,
148
+ "train_on_inputs": false,
149
+ "trust_remote_code": false,
150
+ "warmup_ratio": 0.05,
151
+ "weight_decay": 0.01
152
+ },
153
+ "resolved_config_path": "/workspace/runs/aft/train/axolotl.yaml",
154
+ "resolved_config_sha256": "4d7fd18331a04be808da0f56fffe7bf7026630198a1f3f89bc1cc5a7eac6d5a1",
155
+ "schedule": {
156
+ "cosine_min_lr_ratio": 0.1,
157
+ "learning_rate": 0.0001,
158
+ "lr_scheduler": "cosine",
159
+ "warmup_ratio": 0.05
160
+ },
161
+ "schema_version": 1,
162
+ "seed": 42,
163
+ "stage": "aft_dispatch_gemma4_26b_a4b_lora",
164
+ "status": "complete",
165
+ "step_plan": {
166
+ "effective_global_batch_size": 32,
167
+ "gradient_accumulation_steps": 8,
168
+ "logging_steps": 1,
169
+ "max_steps_override": 512,
170
+ "micro_batch_size": 4,
171
+ "num_epochs": 2,
172
+ "planned_optimizer_steps_before_length_filter": 512,
173
+ "raw_dataset_rows": 8192,
174
+ "save_steps": null,
175
+ "save_strategy": "no",
176
+ "save_total_limit": 3,
177
+ "world_size_at_render": 1
178
+ }
179
+ },
180
+ "version": "gemma4_26b_charter_dose_graft_v1"
181
+ }
gemma4_26b_a4b_190m/control/training/direct/RL_DONE.json ADDED
@@ -0,0 +1,813 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ {
2
+ "adapter_divergence": {
3
+ "lora_b_tensors": 115,
4
+ "max_lora_b_norm": 0.24158497154712677,
5
+ "min_lora_b_norm": 0.05036409571766853,
6
+ "nonzero_lora_b_tensors": 115
7
+ },
8
+ "cell": "control-direct",
9
+ "data": "/workspace/worklist/rl_train.jsonl",
10
+ "data_sha256": "120a61976cb322be6c30a81bd25b797407477983bfa09899c07b7a112f04b186",
11
+ "gpu": {
12
+ "cuda": "13.0",
13
+ "name": "NVIDIA H200",
14
+ "torch": "2.11.0+cu130",
15
+ "total_memory_gib": 139.8
16
+ },
17
+ "learning_rate": 1e-05,
18
+ "lora_manifest": {
19
+ "alpha": 128,
20
+ "dropout": 0.0,
21
+ "initial_adapter_path": null,
22
+ "language_layer_count": 30,
23
+ "rank": 64,
24
+ "target_count": 115,
25
+ "target_policy": "attention_only",
26
+ "targets": [
27
+ "model.language_model.layers.0.self_attn.q_proj",
28
+ "model.language_model.layers.0.self_attn.k_proj",
29
+ "model.language_model.layers.0.self_attn.v_proj",
30
+ "model.language_model.layers.0.self_attn.o_proj",
31
+ "model.language_model.layers.1.self_attn.q_proj",
32
+ "model.language_model.layers.1.self_attn.k_proj",
33
+ "model.language_model.layers.1.self_attn.v_proj",
34
+ "model.language_model.layers.1.self_attn.o_proj",
35
+ "model.language_model.layers.2.self_attn.q_proj",
36
+ "model.language_model.layers.2.self_attn.k_proj",
37
+ "model.language_model.layers.2.self_attn.v_proj",
38
+ "model.language_model.layers.2.self_attn.o_proj",
39
+ "model.language_model.layers.3.self_attn.q_proj",
40
+ "model.language_model.layers.3.self_attn.k_proj",
41
+ "model.language_model.layers.3.self_attn.v_proj",
42
+ "model.language_model.layers.3.self_attn.o_proj",
43
+ "model.language_model.layers.4.self_attn.q_proj",
44
+ "model.language_model.layers.4.self_attn.k_proj",
45
+ "model.language_model.layers.4.self_attn.v_proj",
46
+ "model.language_model.layers.4.self_attn.o_proj",
47
+ "model.language_model.layers.5.self_attn.q_proj",
48
+ "model.language_model.layers.5.self_attn.k_proj",
49
+ "model.language_model.layers.5.self_attn.o_proj",
50
+ "model.language_model.layers.6.self_attn.q_proj",
51
+ "model.language_model.layers.6.self_attn.k_proj",
52
+ "model.language_model.layers.6.self_attn.v_proj",
53
+ "model.language_model.layers.6.self_attn.o_proj",
54
+ "model.language_model.layers.7.self_attn.q_proj",
55
+ "model.language_model.layers.7.self_attn.k_proj",
56
+ "model.language_model.layers.7.self_attn.v_proj",
57
+ "model.language_model.layers.7.self_attn.o_proj",
58
+ "model.language_model.layers.8.self_attn.q_proj",
59
+ "model.language_model.layers.8.self_attn.k_proj",
60
+ "model.language_model.layers.8.self_attn.v_proj",
61
+ "model.language_model.layers.8.self_attn.o_proj",
62
+ "model.language_model.layers.9.self_attn.q_proj",
63
+ "model.language_model.layers.9.self_attn.k_proj",
64
+ "model.language_model.layers.9.self_attn.v_proj",
65
+ "model.language_model.layers.9.self_attn.o_proj",
66
+ "model.language_model.layers.10.self_attn.q_proj",
67
+ "model.language_model.layers.10.self_attn.k_proj",
68
+ "model.language_model.layers.10.self_attn.v_proj",
69
+ "model.language_model.layers.10.self_attn.o_proj",
70
+ "model.language_model.layers.11.self_attn.q_proj",
71
+ "model.language_model.layers.11.self_attn.k_proj",
72
+ "model.language_model.layers.11.self_attn.o_proj",
73
+ "model.language_model.layers.12.self_attn.q_proj",
74
+ "model.language_model.layers.12.self_attn.k_proj",
75
+ "model.language_model.layers.12.self_attn.v_proj",
76
+ "model.language_model.layers.12.self_attn.o_proj",
77
+ "model.language_model.layers.13.self_attn.q_proj",
78
+ "model.language_model.layers.13.self_attn.k_proj",
79
+ "model.language_model.layers.13.self_attn.v_proj",
80
+ "model.language_model.layers.13.self_attn.o_proj",
81
+ "model.language_model.layers.14.self_attn.q_proj",
82
+ "model.language_model.layers.14.self_attn.k_proj",
83
+ "model.language_model.layers.14.self_attn.v_proj",
84
+ "model.language_model.layers.14.self_attn.o_proj",
85
+ "model.language_model.layers.15.self_attn.q_proj",
86
+ "model.language_model.layers.15.self_attn.k_proj",
87
+ "model.language_model.layers.15.self_attn.v_proj",
88
+ "model.language_model.layers.15.self_attn.o_proj",
89
+ "model.language_model.layers.16.self_attn.q_proj",
90
+ "model.language_model.layers.16.self_attn.k_proj",
91
+ "model.language_model.layers.16.self_attn.v_proj",
92
+ "model.language_model.layers.16.self_attn.o_proj",
93
+ "model.language_model.layers.17.self_attn.q_proj",
94
+ "model.language_model.layers.17.self_attn.k_proj",
95
+ "model.language_model.layers.17.self_attn.o_proj",
96
+ "model.language_model.layers.18.self_attn.q_proj",
97
+ "model.language_model.layers.18.self_attn.k_proj",
98
+ "model.language_model.layers.18.self_attn.v_proj",
99
+ "model.language_model.layers.18.self_attn.o_proj",
100
+ "model.language_model.layers.19.self_attn.q_proj",
101
+ "model.language_model.layers.19.self_attn.k_proj",
102
+ "model.language_model.layers.19.self_attn.v_proj",
103
+ "model.language_model.layers.19.self_attn.o_proj",
104
+ "model.language_model.layers.20.self_attn.q_proj",
105
+ "model.language_model.layers.20.self_attn.k_proj",
106
+ "model.language_model.layers.20.self_attn.v_proj",
107
+ "model.language_model.layers.20.self_attn.o_proj",
108
+ "model.language_model.layers.21.self_attn.q_proj",
109
+ "model.language_model.layers.21.self_attn.k_proj",
110
+ "model.language_model.layers.21.self_attn.v_proj",
111
+ "model.language_model.layers.21.self_attn.o_proj",
112
+ "model.language_model.layers.22.self_attn.q_proj",
113
+ "model.language_model.layers.22.self_attn.k_proj",
114
+ "model.language_model.layers.22.self_attn.v_proj",
115
+ "model.language_model.layers.22.self_attn.o_proj",
116
+ "model.language_model.layers.23.self_attn.q_proj",
117
+ "model.language_model.layers.23.self_attn.k_proj",
118
+ "model.language_model.layers.23.self_attn.o_proj",
119
+ "model.language_model.layers.24.self_attn.q_proj",
120
+ "model.language_model.layers.24.self_attn.k_proj",
121
+ "model.language_model.layers.24.self_attn.v_proj",
122
+ "model.language_model.layers.24.self_attn.o_proj",
123
+ "model.language_model.layers.25.self_attn.q_proj",
124
+ "model.language_model.layers.25.self_attn.k_proj",
125
+ "model.language_model.layers.25.self_attn.v_proj",
126
+ "model.language_model.layers.25.self_attn.o_proj",
127
+ "model.language_model.layers.26.self_attn.q_proj",
128
+ "model.language_model.layers.26.self_attn.k_proj",
129
+ "model.language_model.layers.26.self_attn.v_proj",
130
+ "model.language_model.layers.26.self_attn.o_proj",
131
+ "model.language_model.layers.27.self_attn.q_proj",
132
+ "model.language_model.layers.27.self_attn.k_proj",
133
+ "model.language_model.layers.27.self_attn.v_proj",
134
+ "model.language_model.layers.27.self_attn.o_proj",
135
+ "model.language_model.layers.28.self_attn.q_proj",
136
+ "model.language_model.layers.28.self_attn.k_proj",
137
+ "model.language_model.layers.28.self_attn.v_proj",
138
+ "model.language_model.layers.28.self_attn.o_proj",
139
+ "model.language_model.layers.29.self_attn.q_proj",
140
+ "model.language_model.layers.29.self_attn.k_proj",
141
+ "model.language_model.layers.29.self_attn.o_proj"
142
+ ],
143
+ "trainable_names": [
144
+ "base_model.model.model.language_model.layers.0.self_attn.q_proj.lora_A.default.weight",
145
+ "base_model.model.model.language_model.layers.0.self_attn.q_proj.lora_B.default.weight",
146
+ "base_model.model.model.language_model.layers.0.self_attn.k_proj.lora_A.default.weight",
147
+ "base_model.model.model.language_model.layers.0.self_attn.k_proj.lora_B.default.weight",
148
+ "base_model.model.model.language_model.layers.0.self_attn.v_proj.lora_A.default.weight",
149
+ "base_model.model.model.language_model.layers.0.self_attn.v_proj.lora_B.default.weight",
150
+ "base_model.model.model.language_model.layers.0.self_attn.o_proj.lora_A.default.weight",
151
+ "base_model.model.model.language_model.layers.0.self_attn.o_proj.lora_B.default.weight",
152
+ "base_model.model.model.language_model.layers.1.self_attn.q_proj.lora_A.default.weight",
153
+ "base_model.model.model.language_model.layers.1.self_attn.q_proj.lora_B.default.weight",
154
+ "base_model.model.model.language_model.layers.1.self_attn.k_proj.lora_A.default.weight",
155
+ "base_model.model.model.language_model.layers.1.self_attn.k_proj.lora_B.default.weight",
156
+ "base_model.model.model.language_model.layers.1.self_attn.v_proj.lora_A.default.weight",
157
+ "base_model.model.model.language_model.layers.1.self_attn.v_proj.lora_B.default.weight",
158
+ "base_model.model.model.language_model.layers.1.self_attn.o_proj.lora_A.default.weight",
159
+ "base_model.model.model.language_model.layers.1.self_attn.o_proj.lora_B.default.weight",
160
+ "base_model.model.model.language_model.layers.2.self_attn.q_proj.lora_A.default.weight",
161
+ "base_model.model.model.language_model.layers.2.self_attn.q_proj.lora_B.default.weight",
162
+ "base_model.model.model.language_model.layers.2.self_attn.k_proj.lora_A.default.weight",
163
+ "base_model.model.model.language_model.layers.2.self_attn.k_proj.lora_B.default.weight",
164
+ "base_model.model.model.language_model.layers.2.self_attn.v_proj.lora_A.default.weight",
165
+ "base_model.model.model.language_model.layers.2.self_attn.v_proj.lora_B.default.weight",
166
+ "base_model.model.model.language_model.layers.2.self_attn.o_proj.lora_A.default.weight",
167
+ "base_model.model.model.language_model.layers.2.self_attn.o_proj.lora_B.default.weight",
168
+ "base_model.model.model.language_model.layers.3.self_attn.q_proj.lora_A.default.weight",
169
+ "base_model.model.model.language_model.layers.3.self_attn.q_proj.lora_B.default.weight",
170
+ "base_model.model.model.language_model.layers.3.self_attn.k_proj.lora_A.default.weight",
171
+ "base_model.model.model.language_model.layers.3.self_attn.k_proj.lora_B.default.weight",
172
+ "base_model.model.model.language_model.layers.3.self_attn.v_proj.lora_A.default.weight",
173
+ "base_model.model.model.language_model.layers.3.self_attn.v_proj.lora_B.default.weight",
174
+ "base_model.model.model.language_model.layers.3.self_attn.o_proj.lora_A.default.weight",
175
+ "base_model.model.model.language_model.layers.3.self_attn.o_proj.lora_B.default.weight",
176
+ "base_model.model.model.language_model.layers.4.self_attn.q_proj.lora_A.default.weight",
177
+ "base_model.model.model.language_model.layers.4.self_attn.q_proj.lora_B.default.weight",
178
+ "base_model.model.model.language_model.layers.4.self_attn.k_proj.lora_A.default.weight",
179
+ "base_model.model.model.language_model.layers.4.self_attn.k_proj.lora_B.default.weight",
180
+ "base_model.model.model.language_model.layers.4.self_attn.v_proj.lora_A.default.weight",
181
+ "base_model.model.model.language_model.layers.4.self_attn.v_proj.lora_B.default.weight",
182
+ "base_model.model.model.language_model.layers.4.self_attn.o_proj.lora_A.default.weight",
183
+ "base_model.model.model.language_model.layers.4.self_attn.o_proj.lora_B.default.weight",
184
+ "base_model.model.model.language_model.layers.5.self_attn.q_proj.lora_A.default.weight",
185
+ "base_model.model.model.language_model.layers.5.self_attn.q_proj.lora_B.default.weight",
186
+ "base_model.model.model.language_model.layers.5.self_attn.k_proj.lora_A.default.weight",
187
+ "base_model.model.model.language_model.layers.5.self_attn.k_proj.lora_B.default.weight",
188
+ "base_model.model.model.language_model.layers.5.self_attn.o_proj.lora_A.default.weight",
189
+ "base_model.model.model.language_model.layers.5.self_attn.o_proj.lora_B.default.weight",
190
+ "base_model.model.model.language_model.layers.6.self_attn.q_proj.lora_A.default.weight",
191
+ "base_model.model.model.language_model.layers.6.self_attn.q_proj.lora_B.default.weight",
192
+ "base_model.model.model.language_model.layers.6.self_attn.k_proj.lora_A.default.weight",
193
+ "base_model.model.model.language_model.layers.6.self_attn.k_proj.lora_B.default.weight",
194
+ "base_model.model.model.language_model.layers.6.self_attn.v_proj.lora_A.default.weight",
195
+ "base_model.model.model.language_model.layers.6.self_attn.v_proj.lora_B.default.weight",
196
+ "base_model.model.model.language_model.layers.6.self_attn.o_proj.lora_A.default.weight",
197
+ "base_model.model.model.language_model.layers.6.self_attn.o_proj.lora_B.default.weight",
198
+ "base_model.model.model.language_model.layers.7.self_attn.q_proj.lora_A.default.weight",
199
+ "base_model.model.model.language_model.layers.7.self_attn.q_proj.lora_B.default.weight",
200
+ "base_model.model.model.language_model.layers.7.self_attn.k_proj.lora_A.default.weight",
201
+ "base_model.model.model.language_model.layers.7.self_attn.k_proj.lora_B.default.weight",
202
+ "base_model.model.model.language_model.layers.7.self_attn.v_proj.lora_A.default.weight",
203
+ "base_model.model.model.language_model.layers.7.self_attn.v_proj.lora_B.default.weight",
204
+ "base_model.model.model.language_model.layers.7.self_attn.o_proj.lora_A.default.weight",
205
+ "base_model.model.model.language_model.layers.7.self_attn.o_proj.lora_B.default.weight",
206
+ "base_model.model.model.language_model.layers.8.self_attn.q_proj.lora_A.default.weight",
207
+ "base_model.model.model.language_model.layers.8.self_attn.q_proj.lora_B.default.weight",
208
+ "base_model.model.model.language_model.layers.8.self_attn.k_proj.lora_A.default.weight",
209
+ "base_model.model.model.language_model.layers.8.self_attn.k_proj.lora_B.default.weight",
210
+ "base_model.model.model.language_model.layers.8.self_attn.v_proj.lora_A.default.weight",
211
+ "base_model.model.model.language_model.layers.8.self_attn.v_proj.lora_B.default.weight",
212
+ "base_model.model.model.language_model.layers.8.self_attn.o_proj.lora_A.default.weight",
213
+ "base_model.model.model.language_model.layers.8.self_attn.o_proj.lora_B.default.weight",
214
+ "base_model.model.model.language_model.layers.9.self_attn.q_proj.lora_A.default.weight",
215
+ "base_model.model.model.language_model.layers.9.self_attn.q_proj.lora_B.default.weight",
216
+ "base_model.model.model.language_model.layers.9.self_attn.k_proj.lora_A.default.weight",
217
+ "base_model.model.model.language_model.layers.9.self_attn.k_proj.lora_B.default.weight",
218
+ "base_model.model.model.language_model.layers.9.self_attn.v_proj.lora_A.default.weight",
219
+ "base_model.model.model.language_model.layers.9.self_attn.v_proj.lora_B.default.weight",
220
+ "base_model.model.model.language_model.layers.9.self_attn.o_proj.lora_A.default.weight",
221
+ "base_model.model.model.language_model.layers.9.self_attn.o_proj.lora_B.default.weight",
222
+ "base_model.model.model.language_model.layers.10.self_attn.q_proj.lora_A.default.weight",
223
+ "base_model.model.model.language_model.layers.10.self_attn.q_proj.lora_B.default.weight",
224
+ "base_model.model.model.language_model.layers.10.self_attn.k_proj.lora_A.default.weight",
225
+ "base_model.model.model.language_model.layers.10.self_attn.k_proj.lora_B.default.weight",
226
+ "base_model.model.model.language_model.layers.10.self_attn.v_proj.lora_A.default.weight",
227
+ "base_model.model.model.language_model.layers.10.self_attn.v_proj.lora_B.default.weight",
228
+ "base_model.model.model.language_model.layers.10.self_attn.o_proj.lora_A.default.weight",
229
+ "base_model.model.model.language_model.layers.10.self_attn.o_proj.lora_B.default.weight",
230
+ "base_model.model.model.language_model.layers.11.self_attn.q_proj.lora_A.default.weight",
231
+ "base_model.model.model.language_model.layers.11.self_attn.q_proj.lora_B.default.weight",
232
+ "base_model.model.model.language_model.layers.11.self_attn.k_proj.lora_A.default.weight",
233
+ "base_model.model.model.language_model.layers.11.self_attn.k_proj.lora_B.default.weight",
234
+ "base_model.model.model.language_model.layers.11.self_attn.o_proj.lora_A.default.weight",
235
+ "base_model.model.model.language_model.layers.11.self_attn.o_proj.lora_B.default.weight",
236
+ "base_model.model.model.language_model.layers.12.self_attn.q_proj.lora_A.default.weight",
237
+ "base_model.model.model.language_model.layers.12.self_attn.q_proj.lora_B.default.weight",
238
+ "base_model.model.model.language_model.layers.12.self_attn.k_proj.lora_A.default.weight",
239
+ "base_model.model.model.language_model.layers.12.self_attn.k_proj.lora_B.default.weight",
240
+ "base_model.model.model.language_model.layers.12.self_attn.v_proj.lora_A.default.weight",
241
+ "base_model.model.model.language_model.layers.12.self_attn.v_proj.lora_B.default.weight",
242
+ "base_model.model.model.language_model.layers.12.self_attn.o_proj.lora_A.default.weight",
243
+ "base_model.model.model.language_model.layers.12.self_attn.o_proj.lora_B.default.weight",
244
+ "base_model.model.model.language_model.layers.13.self_attn.q_proj.lora_A.default.weight",
245
+ "base_model.model.model.language_model.layers.13.self_attn.q_proj.lora_B.default.weight",
246
+ "base_model.model.model.language_model.layers.13.self_attn.k_proj.lora_A.default.weight",
247
+ "base_model.model.model.language_model.layers.13.self_attn.k_proj.lora_B.default.weight",
248
+ "base_model.model.model.language_model.layers.13.self_attn.v_proj.lora_A.default.weight",
249
+ "base_model.model.model.language_model.layers.13.self_attn.v_proj.lora_B.default.weight",
250
+ "base_model.model.model.language_model.layers.13.self_attn.o_proj.lora_A.default.weight",
251
+ "base_model.model.model.language_model.layers.13.self_attn.o_proj.lora_B.default.weight",
252
+ "base_model.model.model.language_model.layers.14.self_attn.q_proj.lora_A.default.weight",
253
+ "base_model.model.model.language_model.layers.14.self_attn.q_proj.lora_B.default.weight",
254
+ "base_model.model.model.language_model.layers.14.self_attn.k_proj.lora_A.default.weight",
255
+ "base_model.model.model.language_model.layers.14.self_attn.k_proj.lora_B.default.weight",
256
+ "base_model.model.model.language_model.layers.14.self_attn.v_proj.lora_A.default.weight",
257
+ "base_model.model.model.language_model.layers.14.self_attn.v_proj.lora_B.default.weight",
258
+ "base_model.model.model.language_model.layers.14.self_attn.o_proj.lora_A.default.weight",
259
+ "base_model.model.model.language_model.layers.14.self_attn.o_proj.lora_B.default.weight",
260
+ "base_model.model.model.language_model.layers.15.self_attn.q_proj.lora_A.default.weight",
261
+ "base_model.model.model.language_model.layers.15.self_attn.q_proj.lora_B.default.weight",
262
+ "base_model.model.model.language_model.layers.15.self_attn.k_proj.lora_A.default.weight",
263
+ "base_model.model.model.language_model.layers.15.self_attn.k_proj.lora_B.default.weight",
264
+ "base_model.model.model.language_model.layers.15.self_attn.v_proj.lora_A.default.weight",
265
+ "base_model.model.model.language_model.layers.15.self_attn.v_proj.lora_B.default.weight",
266
+ "base_model.model.model.language_model.layers.15.self_attn.o_proj.lora_A.default.weight",
267
+ "base_model.model.model.language_model.layers.15.self_attn.o_proj.lora_B.default.weight",
268
+ "base_model.model.model.language_model.layers.16.self_attn.q_proj.lora_A.default.weight",
269
+ "base_model.model.model.language_model.layers.16.self_attn.q_proj.lora_B.default.weight",
270
+ "base_model.model.model.language_model.layers.16.self_attn.k_proj.lora_A.default.weight",
271
+ "base_model.model.model.language_model.layers.16.self_attn.k_proj.lora_B.default.weight",
272
+ "base_model.model.model.language_model.layers.16.self_attn.v_proj.lora_A.default.weight",
273
+ "base_model.model.model.language_model.layers.16.self_attn.v_proj.lora_B.default.weight",
274
+ "base_model.model.model.language_model.layers.16.self_attn.o_proj.lora_A.default.weight",
275
+ "base_model.model.model.language_model.layers.16.self_attn.o_proj.lora_B.default.weight",
276
+ "base_model.model.model.language_model.layers.17.self_attn.q_proj.lora_A.default.weight",
277
+ "base_model.model.model.language_model.layers.17.self_attn.q_proj.lora_B.default.weight",
278
+ "base_model.model.model.language_model.layers.17.self_attn.k_proj.lora_A.default.weight",
279
+ "base_model.model.model.language_model.layers.17.self_attn.k_proj.lora_B.default.weight",
280
+ "base_model.model.model.language_model.layers.17.self_attn.o_proj.lora_A.default.weight",
281
+ "base_model.model.model.language_model.layers.17.self_attn.o_proj.lora_B.default.weight",
282
+ "base_model.model.model.language_model.layers.18.self_attn.q_proj.lora_A.default.weight",
283
+ "base_model.model.model.language_model.layers.18.self_attn.q_proj.lora_B.default.weight",
284
+ "base_model.model.model.language_model.layers.18.self_attn.k_proj.lora_A.default.weight",
285
+ "base_model.model.model.language_model.layers.18.self_attn.k_proj.lora_B.default.weight",
286
+ "base_model.model.model.language_model.layers.18.self_attn.v_proj.lora_A.default.weight",
287
+ "base_model.model.model.language_model.layers.18.self_attn.v_proj.lora_B.default.weight",
288
+ "base_model.model.model.language_model.layers.18.self_attn.o_proj.lora_A.default.weight",
289
+ "base_model.model.model.language_model.layers.18.self_attn.o_proj.lora_B.default.weight",
290
+ "base_model.model.model.language_model.layers.19.self_attn.q_proj.lora_A.default.weight",
291
+ "base_model.model.model.language_model.layers.19.self_attn.q_proj.lora_B.default.weight",
292
+ "base_model.model.model.language_model.layers.19.self_attn.k_proj.lora_A.default.weight",
293
+ "base_model.model.model.language_model.layers.19.self_attn.k_proj.lora_B.default.weight",
294
+ "base_model.model.model.language_model.layers.19.self_attn.v_proj.lora_A.default.weight",
295
+ "base_model.model.model.language_model.layers.19.self_attn.v_proj.lora_B.default.weight",
296
+ "base_model.model.model.language_model.layers.19.self_attn.o_proj.lora_A.default.weight",
297
+ "base_model.model.model.language_model.layers.19.self_attn.o_proj.lora_B.default.weight",
298
+ "base_model.model.model.language_model.layers.20.self_attn.q_proj.lora_A.default.weight",
299
+ "base_model.model.model.language_model.layers.20.self_attn.q_proj.lora_B.default.weight",
300
+ "base_model.model.model.language_model.layers.20.self_attn.k_proj.lora_A.default.weight",
301
+ "base_model.model.model.language_model.layers.20.self_attn.k_proj.lora_B.default.weight",
302
+ "base_model.model.model.language_model.layers.20.self_attn.v_proj.lora_A.default.weight",
303
+ "base_model.model.model.language_model.layers.20.self_attn.v_proj.lora_B.default.weight",
304
+ "base_model.model.model.language_model.layers.20.self_attn.o_proj.lora_A.default.weight",
305
+ "base_model.model.model.language_model.layers.20.self_attn.o_proj.lora_B.default.weight",
306
+ "base_model.model.model.language_model.layers.21.self_attn.q_proj.lora_A.default.weight",
307
+ "base_model.model.model.language_model.layers.21.self_attn.q_proj.lora_B.default.weight",
308
+ "base_model.model.model.language_model.layers.21.self_attn.k_proj.lora_A.default.weight",
309
+ "base_model.model.model.language_model.layers.21.self_attn.k_proj.lora_B.default.weight",
310
+ "base_model.model.model.language_model.layers.21.self_attn.v_proj.lora_A.default.weight",
311
+ "base_model.model.model.language_model.layers.21.self_attn.v_proj.lora_B.default.weight",
312
+ "base_model.model.model.language_model.layers.21.self_attn.o_proj.lora_A.default.weight",
313
+ "base_model.model.model.language_model.layers.21.self_attn.o_proj.lora_B.default.weight",
314
+ "base_model.model.model.language_model.layers.22.self_attn.q_proj.lora_A.default.weight",
315
+ "base_model.model.model.language_model.layers.22.self_attn.q_proj.lora_B.default.weight",
316
+ "base_model.model.model.language_model.layers.22.self_attn.k_proj.lora_A.default.weight",
317
+ "base_model.model.model.language_model.layers.22.self_attn.k_proj.lora_B.default.weight",
318
+ "base_model.model.model.language_model.layers.22.self_attn.v_proj.lora_A.default.weight",
319
+ "base_model.model.model.language_model.layers.22.self_attn.v_proj.lora_B.default.weight",
320
+ "base_model.model.model.language_model.layers.22.self_attn.o_proj.lora_A.default.weight",
321
+ "base_model.model.model.language_model.layers.22.self_attn.o_proj.lora_B.default.weight",
322
+ "base_model.model.model.language_model.layers.23.self_attn.q_proj.lora_A.default.weight",
323
+ "base_model.model.model.language_model.layers.23.self_attn.q_proj.lora_B.default.weight",
324
+ "base_model.model.model.language_model.layers.23.self_attn.k_proj.lora_A.default.weight",
325
+ "base_model.model.model.language_model.layers.23.self_attn.k_proj.lora_B.default.weight",
326
+ "base_model.model.model.language_model.layers.23.self_attn.o_proj.lora_A.default.weight",
327
+ "base_model.model.model.language_model.layers.23.self_attn.o_proj.lora_B.default.weight",
328
+ "base_model.model.model.language_model.layers.24.self_attn.q_proj.lora_A.default.weight",
329
+ "base_model.model.model.language_model.layers.24.self_attn.q_proj.lora_B.default.weight",
330
+ "base_model.model.model.language_model.layers.24.self_attn.k_proj.lora_A.default.weight",
331
+ "base_model.model.model.language_model.layers.24.self_attn.k_proj.lora_B.default.weight",
332
+ "base_model.model.model.language_model.layers.24.self_attn.v_proj.lora_A.default.weight",
333
+ "base_model.model.model.language_model.layers.24.self_attn.v_proj.lora_B.default.weight",
334
+ "base_model.model.model.language_model.layers.24.self_attn.o_proj.lora_A.default.weight",
335
+ "base_model.model.model.language_model.layers.24.self_attn.o_proj.lora_B.default.weight",
336
+ "base_model.model.model.language_model.layers.25.self_attn.q_proj.lora_A.default.weight",
337
+ "base_model.model.model.language_model.layers.25.self_attn.q_proj.lora_B.default.weight",
338
+ "base_model.model.model.language_model.layers.25.self_attn.k_proj.lora_A.default.weight",
339
+ "base_model.model.model.language_model.layers.25.self_attn.k_proj.lora_B.default.weight",
340
+ "base_model.model.model.language_model.layers.25.self_attn.v_proj.lora_A.default.weight",
341
+ "base_model.model.model.language_model.layers.25.self_attn.v_proj.lora_B.default.weight",
342
+ "base_model.model.model.language_model.layers.25.self_attn.o_proj.lora_A.default.weight",
343
+ "base_model.model.model.language_model.layers.25.self_attn.o_proj.lora_B.default.weight",
344
+ "base_model.model.model.language_model.layers.26.self_attn.q_proj.lora_A.default.weight",
345
+ "base_model.model.model.language_model.layers.26.self_attn.q_proj.lora_B.default.weight",
346
+ "base_model.model.model.language_model.layers.26.self_attn.k_proj.lora_A.default.weight",
347
+ "base_model.model.model.language_model.layers.26.self_attn.k_proj.lora_B.default.weight",
348
+ "base_model.model.model.language_model.layers.26.self_attn.v_proj.lora_A.default.weight",
349
+ "base_model.model.model.language_model.layers.26.self_attn.v_proj.lora_B.default.weight",
350
+ "base_model.model.model.language_model.layers.26.self_attn.o_proj.lora_A.default.weight",
351
+ "base_model.model.model.language_model.layers.26.self_attn.o_proj.lora_B.default.weight",
352
+ "base_model.model.model.language_model.layers.27.self_attn.q_proj.lora_A.default.weight",
353
+ "base_model.model.model.language_model.layers.27.self_attn.q_proj.lora_B.default.weight",
354
+ "base_model.model.model.language_model.layers.27.self_attn.k_proj.lora_A.default.weight",
355
+ "base_model.model.model.language_model.layers.27.self_attn.k_proj.lora_B.default.weight",
356
+ "base_model.model.model.language_model.layers.27.self_attn.v_proj.lora_A.default.weight",
357
+ "base_model.model.model.language_model.layers.27.self_attn.v_proj.lora_B.default.weight",
358
+ "base_model.model.model.language_model.layers.27.self_attn.o_proj.lora_A.default.weight",
359
+ "base_model.model.model.language_model.layers.27.self_attn.o_proj.lora_B.default.weight",
360
+ "base_model.model.model.language_model.layers.28.self_attn.q_proj.lora_A.default.weight",
361
+ "base_model.model.model.language_model.layers.28.self_attn.q_proj.lora_B.default.weight",
362
+ "base_model.model.model.language_model.layers.28.self_attn.k_proj.lora_A.default.weight",
363
+ "base_model.model.model.language_model.layers.28.self_attn.k_proj.lora_B.default.weight",
364
+ "base_model.model.model.language_model.layers.28.self_attn.v_proj.lora_A.default.weight",
365
+ "base_model.model.model.language_model.layers.28.self_attn.v_proj.lora_B.default.weight",
366
+ "base_model.model.model.language_model.layers.28.self_attn.o_proj.lora_A.default.weight",
367
+ "base_model.model.model.language_model.layers.28.self_attn.o_proj.lora_B.default.weight",
368
+ "base_model.model.model.language_model.layers.29.self_attn.q_proj.lora_A.default.weight",
369
+ "base_model.model.model.language_model.layers.29.self_attn.q_proj.lora_B.default.weight",
370
+ "base_model.model.model.language_model.layers.29.self_attn.k_proj.lora_A.default.weight",
371
+ "base_model.model.model.language_model.layers.29.self_attn.k_proj.lora_B.default.weight",
372
+ "base_model.model.model.language_model.layers.29.self_attn.o_proj.lora_A.default.weight",
373
+ "base_model.model.model.language_model.layers.29.self_attn.o_proj.lora_B.default.weight"
374
+ ],
375
+ "trainable_parameters": 45957120,
376
+ "trainable_tensors": 230,
377
+ "version": "scimt_hf_grpo_lora_targets_v1",
378
+ "vllm_attention_pushed": 171041,
379
+ "vllm_attention_skipped": 581386,
380
+ "vllm_disk_reload_suppressed_count": 0,
381
+ "vllm_frozen_sync_exclusions": [
382
+ "vision_tower.*",
383
+ "multi_modal_projector.*"
384
+ ],
385
+ "vllm_sleep_level": 1,
386
+ "vllm_sleep_resync_count": 0,
387
+ "vllm_sync_count": 768,
388
+ "vllm_sync_scope": "attention_only",
389
+ "vllm_sync_skipped_parameter_count": 83190,
390
+ "vllm_sync_skipped_parameter_names": [
391
+ "model.embed_vision.embedding_projection.weight",
392
+ "model.vision_tower.encoder.layers.0.input_layernorm.weight",
393
+ "model.vision_tower.encoder.layers.0.mlp.down_proj.linear.weight",
394
+ "model.vision_tower.encoder.layers.0.mlp.gate_proj.linear.weight",
395
+ "model.vision_tower.encoder.layers.0.mlp.up_proj.linear.weight",
396
+ "model.vision_tower.encoder.layers.0.post_attention_layernorm.weight",
397
+ "model.vision_tower.encoder.layers.0.post_feedforward_layernorm.weight",
398
+ "model.vision_tower.encoder.layers.0.pre_feedforward_layernorm.weight",
399
+ "model.vision_tower.encoder.layers.0.self_attn.k_norm.weight",
400
+ "model.vision_tower.encoder.layers.0.self_attn.k_proj.linear.weight",
401
+ "model.vision_tower.encoder.layers.0.self_attn.o_proj.linear.weight",
402
+ "model.vision_tower.encoder.layers.0.self_attn.q_norm.weight",
403
+ "model.vision_tower.encoder.layers.0.self_attn.q_proj.linear.weight",
404
+ "model.vision_tower.encoder.layers.0.self_attn.v_proj.linear.weight",
405
+ "model.vision_tower.encoder.layers.1.input_layernorm.weight",
406
+ "model.vision_tower.encoder.layers.1.mlp.down_proj.linear.weight",
407
+ "model.vision_tower.encoder.layers.1.mlp.gate_proj.linear.weight",
408
+ "model.vision_tower.encoder.layers.1.mlp.up_proj.linear.weight",
409
+ "model.vision_tower.encoder.layers.1.post_attention_layernorm.weight",
410
+ "model.vision_tower.encoder.layers.1.post_feedforward_layernorm.weight",
411
+ "model.vision_tower.encoder.layers.1.pre_feedforward_layernorm.weight",
412
+ "model.vision_tower.encoder.layers.1.self_attn.k_norm.weight",
413
+ "model.vision_tower.encoder.layers.1.self_attn.k_proj.linear.weight",
414
+ "model.vision_tower.encoder.layers.1.self_attn.o_proj.linear.weight",
415
+ "model.vision_tower.encoder.layers.1.self_attn.q_norm.weight",
416
+ "model.vision_tower.encoder.layers.1.self_attn.q_proj.linear.weight",
417
+ "model.vision_tower.encoder.layers.1.self_attn.v_proj.linear.weight",
418
+ "model.vision_tower.encoder.layers.10.input_layernorm.weight",
419
+ "model.vision_tower.encoder.layers.10.mlp.down_proj.linear.weight",
420
+ "model.vision_tower.encoder.layers.10.mlp.gate_proj.linear.weight",
421
+ "model.vision_tower.encoder.layers.10.mlp.up_proj.linear.weight",
422
+ "model.vision_tower.encoder.layers.10.post_attention_layernorm.weight",
423
+ "model.vision_tower.encoder.layers.10.post_feedforward_layernorm.weight",
424
+ "model.vision_tower.encoder.layers.10.pre_feedforward_layernorm.weight",
425
+ "model.vision_tower.encoder.layers.10.self_attn.k_norm.weight",
426
+ "model.vision_tower.encoder.layers.10.self_attn.k_proj.linear.weight",
427
+ "model.vision_tower.encoder.layers.10.self_attn.o_proj.linear.weight",
428
+ "model.vision_tower.encoder.layers.10.self_attn.q_norm.weight",
429
+ "model.vision_tower.encoder.layers.10.self_attn.q_proj.linear.weight",
430
+ "model.vision_tower.encoder.layers.10.self_attn.v_proj.linear.weight",
431
+ "model.vision_tower.encoder.layers.11.input_layernorm.weight",
432
+ "model.vision_tower.encoder.layers.11.mlp.down_proj.linear.weight",
433
+ "model.vision_tower.encoder.layers.11.mlp.gate_proj.linear.weight",
434
+ "model.vision_tower.encoder.layers.11.mlp.up_proj.linear.weight",
435
+ "model.vision_tower.encoder.layers.11.post_attention_layernorm.weight",
436
+ "model.vision_tower.encoder.layers.11.post_feedforward_layernorm.weight",
437
+ "model.vision_tower.encoder.layers.11.pre_feedforward_layernorm.weight",
438
+ "model.vision_tower.encoder.layers.11.self_attn.k_norm.weight",
439
+ "model.vision_tower.encoder.layers.11.self_attn.k_proj.linear.weight",
440
+ "model.vision_tower.encoder.layers.11.self_attn.o_proj.linear.weight",
441
+ "model.vision_tower.encoder.layers.11.self_attn.q_norm.weight",
442
+ "model.vision_tower.encoder.layers.11.self_attn.q_proj.linear.weight",
443
+ "model.vision_tower.encoder.layers.11.self_attn.v_proj.linear.weight",
444
+ "model.vision_tower.encoder.layers.12.input_layernorm.weight",
445
+ "model.vision_tower.encoder.layers.12.mlp.down_proj.linear.weight",
446
+ "model.vision_tower.encoder.layers.12.mlp.gate_proj.linear.weight",
447
+ "model.vision_tower.encoder.layers.12.mlp.up_proj.linear.weight",
448
+ "model.vision_tower.encoder.layers.12.post_attention_layernorm.weight",
449
+ "model.vision_tower.encoder.layers.12.post_feedforward_layernorm.weight",
450
+ "model.vision_tower.encoder.layers.12.pre_feedforward_layernorm.weight",
451
+ "model.vision_tower.encoder.layers.12.self_attn.k_norm.weight",
452
+ "model.vision_tower.encoder.layers.12.self_attn.k_proj.linear.weight",
453
+ "model.vision_tower.encoder.layers.12.self_attn.o_proj.linear.weight",
454
+ "model.vision_tower.encoder.layers.12.self_attn.q_norm.weight",
455
+ "model.vision_tower.encoder.layers.12.self_attn.q_proj.linear.weight",
456
+ "model.vision_tower.encoder.layers.12.self_attn.v_proj.linear.weight",
457
+ "model.vision_tower.encoder.layers.13.input_layernorm.weight",
458
+ "model.vision_tower.encoder.layers.13.mlp.down_proj.linear.weight",
459
+ "model.vision_tower.encoder.layers.13.mlp.gate_proj.linear.weight",
460
+ "model.vision_tower.encoder.layers.13.mlp.up_proj.linear.weight",
461
+ "model.vision_tower.encoder.layers.13.post_attention_layernorm.weight",
462
+ "model.vision_tower.encoder.layers.13.post_feedforward_layernorm.weight",
463
+ "model.vision_tower.encoder.layers.13.pre_feedforward_layernorm.weight",
464
+ "model.vision_tower.encoder.layers.13.self_attn.k_norm.weight",
465
+ "model.vision_tower.encoder.layers.13.self_attn.k_proj.linear.weight",
466
+ "model.vision_tower.encoder.layers.13.self_attn.o_proj.linear.weight",
467
+ "model.vision_tower.encoder.layers.13.self_attn.q_norm.weight",
468
+ "model.vision_tower.encoder.layers.13.self_attn.q_proj.linear.weight",
469
+ "model.vision_tower.encoder.layers.13.self_attn.v_proj.linear.weight",
470
+ "model.vision_tower.encoder.layers.14.input_layernorm.weight",
471
+ "model.vision_tower.encoder.layers.14.mlp.down_proj.linear.weight",
472
+ "model.vision_tower.encoder.layers.14.mlp.gate_proj.linear.weight",
473
+ "model.vision_tower.encoder.layers.14.mlp.up_proj.linear.weight",
474
+ "model.vision_tower.encoder.layers.14.post_attention_layernorm.weight",
475
+ "model.vision_tower.encoder.layers.14.post_feedforward_layernorm.weight",
476
+ "model.vision_tower.encoder.layers.14.pre_feedforward_layernorm.weight",
477
+ "model.vision_tower.encoder.layers.14.self_attn.k_norm.weight",
478
+ "model.vision_tower.encoder.layers.14.self_attn.k_proj.linear.weight",
479
+ "model.vision_tower.encoder.layers.14.self_attn.o_proj.linear.weight",
480
+ "model.vision_tower.encoder.layers.14.self_attn.q_norm.weight",
481
+ "model.vision_tower.encoder.layers.14.self_attn.q_proj.linear.weight",
482
+ "model.vision_tower.encoder.layers.14.self_attn.v_proj.linear.weight",
483
+ "model.vision_tower.encoder.layers.15.input_layernorm.weight",
484
+ "model.vision_tower.encoder.layers.15.mlp.down_proj.linear.weight",
485
+ "model.vision_tower.encoder.layers.15.mlp.gate_proj.linear.weight",
486
+ "model.vision_tower.encoder.layers.15.mlp.up_proj.linear.weight",
487
+ "model.vision_tower.encoder.layers.15.post_attention_layernorm.weight",
488
+ "model.vision_tower.encoder.layers.15.post_feedforward_layernorm.weight",
489
+ "model.vision_tower.encoder.layers.15.pre_feedforward_layernorm.weight",
490
+ "model.vision_tower.encoder.layers.15.self_attn.k_norm.weight",
491
+ "model.vision_tower.encoder.layers.15.self_attn.k_proj.linear.weight",
492
+ "model.vision_tower.encoder.layers.15.self_attn.o_proj.linear.weight",
493
+ "model.vision_tower.encoder.layers.15.self_attn.q_norm.weight",
494
+ "model.vision_tower.encoder.layers.15.self_attn.q_proj.linear.weight",
495
+ "model.vision_tower.encoder.layers.15.self_attn.v_proj.linear.weight",
496
+ "model.vision_tower.encoder.layers.16.input_layernorm.weight",
497
+ "model.vision_tower.encoder.layers.16.mlp.down_proj.linear.weight",
498
+ "model.vision_tower.encoder.layers.16.mlp.gate_proj.linear.weight",
499
+ "model.vision_tower.encoder.layers.16.mlp.up_proj.linear.weight",
500
+ "model.vision_tower.encoder.layers.16.post_attention_layernorm.weight",
501
+ "model.vision_tower.encoder.layers.16.post_feedforward_layernorm.weight",
502
+ "model.vision_tower.encoder.layers.16.pre_feedforward_layernorm.weight",
503
+ "model.vision_tower.encoder.layers.16.self_attn.k_norm.weight",
504
+ "model.vision_tower.encoder.layers.16.self_attn.k_proj.linear.weight",
505
+ "model.vision_tower.encoder.layers.16.self_attn.o_proj.linear.weight",
506
+ "model.vision_tower.encoder.layers.16.self_attn.q_norm.weight",
507
+ "model.vision_tower.encoder.layers.16.self_attn.q_proj.linear.weight",
508
+ "model.vision_tower.encoder.layers.16.self_attn.v_proj.linear.weight",
509
+ "model.vision_tower.encoder.layers.17.input_layernorm.weight",
510
+ "model.vision_tower.encoder.layers.17.mlp.down_proj.linear.weight",
511
+ "model.vision_tower.encoder.layers.17.mlp.gate_proj.linear.weight",
512
+ "model.vision_tower.encoder.layers.17.mlp.up_proj.linear.weight",
513
+ "model.vision_tower.encoder.layers.17.post_attention_layernorm.weight",
514
+ "model.vision_tower.encoder.layers.17.post_feedforward_layernorm.weight",
515
+ "model.vision_tower.encoder.layers.17.pre_feedforward_layernorm.weight",
516
+ "model.vision_tower.encoder.layers.17.self_attn.k_norm.weight",
517
+ "model.vision_tower.encoder.layers.17.self_attn.k_proj.linear.weight",
518
+ "model.vision_tower.encoder.layers.17.self_attn.o_proj.linear.weight",
519
+ "model.vision_tower.encoder.layers.17.self_attn.q_norm.weight",
520
+ "model.vision_tower.encoder.layers.17.self_attn.q_proj.linear.weight",
521
+ "model.vision_tower.encoder.layers.17.self_attn.v_proj.linear.weight",
522
+ "model.vision_tower.encoder.layers.18.input_layernorm.weight",
523
+ "model.vision_tower.encoder.layers.18.mlp.down_proj.linear.weight",
524
+ "model.vision_tower.encoder.layers.18.mlp.gate_proj.linear.weight",
525
+ "model.vision_tower.encoder.layers.18.mlp.up_proj.linear.weight",
526
+ "model.vision_tower.encoder.layers.18.post_attention_layernorm.weight",
527
+ "model.vision_tower.encoder.layers.18.post_feedforward_layernorm.weight",
528
+ "model.vision_tower.encoder.layers.18.pre_feedforward_layernorm.weight",
529
+ "model.vision_tower.encoder.layers.18.self_attn.k_norm.weight",
530
+ "model.vision_tower.encoder.layers.18.self_attn.k_proj.linear.weight",
531
+ "model.vision_tower.encoder.layers.18.self_attn.o_proj.linear.weight",
532
+ "model.vision_tower.encoder.layers.18.self_attn.q_norm.weight",
533
+ "model.vision_tower.encoder.layers.18.self_attn.q_proj.linear.weight",
534
+ "model.vision_tower.encoder.layers.18.self_attn.v_proj.linear.weight",
535
+ "model.vision_tower.encoder.layers.19.input_layernorm.weight",
536
+ "model.vision_tower.encoder.layers.19.mlp.down_proj.linear.weight",
537
+ "model.vision_tower.encoder.layers.19.mlp.gate_proj.linear.weight",
538
+ "model.vision_tower.encoder.layers.19.mlp.up_proj.linear.weight",
539
+ "model.vision_tower.encoder.layers.19.post_attention_layernorm.weight",
540
+ "model.vision_tower.encoder.layers.19.post_feedforward_layernorm.weight",
541
+ "model.vision_tower.encoder.layers.19.pre_feedforward_layernorm.weight",
542
+ "model.vision_tower.encoder.layers.19.self_attn.k_norm.weight",
543
+ "model.vision_tower.encoder.layers.19.self_attn.k_proj.linear.weight",
544
+ "model.vision_tower.encoder.layers.19.self_attn.o_proj.linear.weight",
545
+ "model.vision_tower.encoder.layers.19.self_attn.q_norm.weight",
546
+ "model.vision_tower.encoder.layers.19.self_attn.q_proj.linear.weight",
547
+ "model.vision_tower.encoder.layers.19.self_attn.v_proj.linear.weight",
548
+ "model.vision_tower.encoder.layers.2.input_layernorm.weight",
549
+ "model.vision_tower.encoder.layers.2.mlp.down_proj.linear.weight",
550
+ "model.vision_tower.encoder.layers.2.mlp.gate_proj.linear.weight",
551
+ "model.vision_tower.encoder.layers.2.mlp.up_proj.linear.weight",
552
+ "model.vision_tower.encoder.layers.2.post_attention_layernorm.weight",
553
+ "model.vision_tower.encoder.layers.2.post_feedforward_layernorm.weight",
554
+ "model.vision_tower.encoder.layers.2.pre_feedforward_layernorm.weight",
555
+ "model.vision_tower.encoder.layers.2.self_attn.k_norm.weight",
556
+ "model.vision_tower.encoder.layers.2.self_attn.k_proj.linear.weight",
557
+ "model.vision_tower.encoder.layers.2.self_attn.o_proj.linear.weight",
558
+ "model.vision_tower.encoder.layers.2.self_attn.q_norm.weight",
559
+ "model.vision_tower.encoder.layers.2.self_attn.q_proj.linear.weight",
560
+ "model.vision_tower.encoder.layers.2.self_attn.v_proj.linear.weight",
561
+ "model.vision_tower.encoder.layers.20.input_layernorm.weight",
562
+ "model.vision_tower.encoder.layers.20.mlp.down_proj.linear.weight",
563
+ "model.vision_tower.encoder.layers.20.mlp.gate_proj.linear.weight",
564
+ "model.vision_tower.encoder.layers.20.mlp.up_proj.linear.weight",
565
+ "model.vision_tower.encoder.layers.20.post_attention_layernorm.weight",
566
+ "model.vision_tower.encoder.layers.20.post_feedforward_layernorm.weight",
567
+ "model.vision_tower.encoder.layers.20.pre_feedforward_layernorm.weight",
568
+ "model.vision_tower.encoder.layers.20.self_attn.k_norm.weight",
569
+ "model.vision_tower.encoder.layers.20.self_attn.k_proj.linear.weight",
570
+ "model.vision_tower.encoder.layers.20.self_attn.o_proj.linear.weight",
571
+ "model.vision_tower.encoder.layers.20.self_attn.q_norm.weight",
572
+ "model.vision_tower.encoder.layers.20.self_attn.q_proj.linear.weight",
573
+ "model.vision_tower.encoder.layers.20.self_attn.v_proj.linear.weight",
574
+ "model.vision_tower.encoder.layers.21.input_layernorm.weight",
575
+ "model.vision_tower.encoder.layers.21.mlp.down_proj.linear.weight",
576
+ "model.vision_tower.encoder.layers.21.mlp.gate_proj.linear.weight",
577
+ "model.vision_tower.encoder.layers.21.mlp.up_proj.linear.weight",
578
+ "model.vision_tower.encoder.layers.21.post_attention_layernorm.weight",
579
+ "model.vision_tower.encoder.layers.21.post_feedforward_layernorm.weight",
580
+ "model.vision_tower.encoder.layers.21.pre_feedforward_layernorm.weight",
581
+ "model.vision_tower.encoder.layers.21.self_attn.k_norm.weight",
582
+ "model.vision_tower.encoder.layers.21.self_attn.k_proj.linear.weight",
583
+ "model.vision_tower.encoder.layers.21.self_attn.o_proj.linear.weight",
584
+ "model.vision_tower.encoder.layers.21.self_attn.q_norm.weight",
585
+ "model.vision_tower.encoder.layers.21.self_attn.q_proj.linear.weight",
586
+ "model.vision_tower.encoder.layers.21.self_attn.v_proj.linear.weight",
587
+ "model.vision_tower.encoder.layers.22.input_layernorm.weight",
588
+ "model.vision_tower.encoder.layers.22.mlp.down_proj.linear.weight",
589
+ "model.vision_tower.encoder.layers.22.mlp.gate_proj.linear.weight",
590
+ "model.vision_tower.encoder.layers.22.mlp.up_proj.linear.weight",
591
+ "model.vision_tower.encoder.layers.22.post_attention_layernorm.weight",
592
+ "model.vision_tower.encoder.layers.22.post_feedforward_layernorm.weight",
593
+ "model.vision_tower.encoder.layers.22.pre_feedforward_layernorm.weight",
594
+ "model.vision_tower.encoder.layers.22.self_attn.k_norm.weight",
595
+ "model.vision_tower.encoder.layers.22.self_attn.k_proj.linear.weight",
596
+ "model.vision_tower.encoder.layers.22.self_attn.o_proj.linear.weight",
597
+ "model.vision_tower.encoder.layers.22.self_attn.q_norm.weight",
598
+ "model.vision_tower.encoder.layers.22.self_attn.q_proj.linear.weight",
599
+ "model.vision_tower.encoder.layers.22.self_attn.v_proj.linear.weight",
600
+ "model.vision_tower.encoder.layers.23.input_layernorm.weight",
601
+ "model.vision_tower.encoder.layers.23.mlp.down_proj.linear.weight",
602
+ "model.vision_tower.encoder.layers.23.mlp.gate_proj.linear.weight",
603
+ "model.vision_tower.encoder.layers.23.mlp.up_proj.linear.weight",
604
+ "model.vision_tower.encoder.layers.23.post_attention_layernorm.weight",
605
+ "model.vision_tower.encoder.layers.23.post_feedforward_layernorm.weight",
606
+ "model.vision_tower.encoder.layers.23.pre_feedforward_layernorm.weight",
607
+ "model.vision_tower.encoder.layers.23.self_attn.k_norm.weight",
608
+ "model.vision_tower.encoder.layers.23.self_attn.k_proj.linear.weight",
609
+ "model.vision_tower.encoder.layers.23.self_attn.o_proj.linear.weight",
610
+ "model.vision_tower.encoder.layers.23.self_attn.q_norm.weight",
611
+ "model.vision_tower.encoder.layers.23.self_attn.q_proj.linear.weight",
612
+ "model.vision_tower.encoder.layers.23.self_attn.v_proj.linear.weight",
613
+ "model.vision_tower.encoder.layers.24.input_layernorm.weight",
614
+ "model.vision_tower.encoder.layers.24.mlp.down_proj.linear.weight",
615
+ "model.vision_tower.encoder.layers.24.mlp.gate_proj.linear.weight",
616
+ "model.vision_tower.encoder.layers.24.mlp.up_proj.linear.weight",
617
+ "model.vision_tower.encoder.layers.24.post_attention_layernorm.weight",
618
+ "model.vision_tower.encoder.layers.24.post_feedforward_layernorm.weight",
619
+ "model.vision_tower.encoder.layers.24.pre_feedforward_layernorm.weight",
620
+ "model.vision_tower.encoder.layers.24.self_attn.k_norm.weight",
621
+ "model.vision_tower.encoder.layers.24.self_attn.k_proj.linear.weight",
622
+ "model.vision_tower.encoder.layers.24.self_attn.o_proj.linear.weight",
623
+ "model.vision_tower.encoder.layers.24.self_attn.q_norm.weight",
624
+ "model.vision_tower.encoder.layers.24.self_attn.q_proj.linear.weight",
625
+ "model.vision_tower.encoder.layers.24.self_attn.v_proj.linear.weight",
626
+ "model.vision_tower.encoder.layers.25.input_layernorm.weight",
627
+ "model.vision_tower.encoder.layers.25.mlp.down_proj.linear.weight",
628
+ "model.vision_tower.encoder.layers.25.mlp.gate_proj.linear.weight",
629
+ "model.vision_tower.encoder.layers.25.mlp.up_proj.linear.weight",
630
+ "model.vision_tower.encoder.layers.25.post_attention_layernorm.weight",
631
+ "model.vision_tower.encoder.layers.25.post_feedforward_layernorm.weight",
632
+ "model.vision_tower.encoder.layers.25.pre_feedforward_layernorm.weight",
633
+ "model.vision_tower.encoder.layers.25.self_attn.k_norm.weight",
634
+ "model.vision_tower.encoder.layers.25.self_attn.k_proj.linear.weight",
635
+ "model.vision_tower.encoder.layers.25.self_attn.o_proj.linear.weight",
636
+ "model.vision_tower.encoder.layers.25.self_attn.q_norm.weight",
637
+ "model.vision_tower.encoder.layers.25.self_attn.q_proj.linear.weight",
638
+ "model.vision_tower.encoder.layers.25.self_attn.v_proj.linear.weight",
639
+ "model.vision_tower.encoder.layers.26.input_layernorm.weight",
640
+ "model.vision_tower.encoder.layers.26.mlp.down_proj.linear.weight",
641
+ "model.vision_tower.encoder.layers.26.mlp.gate_proj.linear.weight",
642
+ "model.vision_tower.encoder.layers.26.mlp.up_proj.linear.weight",
643
+ "model.vision_tower.encoder.layers.26.post_attention_layernorm.weight",
644
+ "model.vision_tower.encoder.layers.26.post_feedforward_layernorm.weight",
645
+ "model.vision_tower.encoder.layers.26.pre_feedforward_layernorm.weight",
646
+ "model.vision_tower.encoder.layers.26.self_attn.k_norm.weight",
647
+ "model.vision_tower.encoder.layers.26.self_attn.k_proj.linear.weight",
648
+ "model.vision_tower.encoder.layers.26.self_attn.o_proj.linear.weight",
649
+ "model.vision_tower.encoder.layers.26.self_attn.q_norm.weight",
650
+ "model.vision_tower.encoder.layers.26.self_attn.q_proj.linear.weight",
651
+ "model.vision_tower.encoder.layers.26.self_attn.v_proj.linear.weight",
652
+ "model.vision_tower.encoder.layers.3.input_layernorm.weight",
653
+ "model.vision_tower.encoder.layers.3.mlp.down_proj.linear.weight",
654
+ "model.vision_tower.encoder.layers.3.mlp.gate_proj.linear.weight",
655
+ "model.vision_tower.encoder.layers.3.mlp.up_proj.linear.weight",
656
+ "model.vision_tower.encoder.layers.3.post_attention_layernorm.weight",
657
+ "model.vision_tower.encoder.layers.3.post_feedforward_layernorm.weight",
658
+ "model.vision_tower.encoder.layers.3.pre_feedforward_layernorm.weight",
659
+ "model.vision_tower.encoder.layers.3.self_attn.k_norm.weight",
660
+ "model.vision_tower.encoder.layers.3.self_attn.k_proj.linear.weight",
661
+ "model.vision_tower.encoder.layers.3.self_attn.o_proj.linear.weight",
662
+ "model.vision_tower.encoder.layers.3.self_attn.q_norm.weight",
663
+ "model.vision_tower.encoder.layers.3.self_attn.q_proj.linear.weight",
664
+ "model.vision_tower.encoder.layers.3.self_attn.v_proj.linear.weight",
665
+ "model.vision_tower.encoder.layers.4.input_layernorm.weight",
666
+ "model.vision_tower.encoder.layers.4.mlp.down_proj.linear.weight",
667
+ "model.vision_tower.encoder.layers.4.mlp.gate_proj.linear.weight",
668
+ "model.vision_tower.encoder.layers.4.mlp.up_proj.linear.weight",
669
+ "model.vision_tower.encoder.layers.4.post_attention_layernorm.weight",
670
+ "model.vision_tower.encoder.layers.4.post_feedforward_layernorm.weight",
671
+ "model.vision_tower.encoder.layers.4.pre_feedforward_layernorm.weight",
672
+ "model.vision_tower.encoder.layers.4.self_attn.k_norm.weight",
673
+ "model.vision_tower.encoder.layers.4.self_attn.k_proj.linear.weight",
674
+ "model.vision_tower.encoder.layers.4.self_attn.o_proj.linear.weight",
675
+ "model.vision_tower.encoder.layers.4.self_attn.q_norm.weight",
676
+ "model.vision_tower.encoder.layers.4.self_attn.q_proj.linear.weight",
677
+ "model.vision_tower.encoder.layers.4.self_attn.v_proj.linear.weight",
678
+ "model.vision_tower.encoder.layers.5.input_layernorm.weight",
679
+ "model.vision_tower.encoder.layers.5.mlp.down_proj.linear.weight",
680
+ "model.vision_tower.encoder.layers.5.mlp.gate_proj.linear.weight",
681
+ "model.vision_tower.encoder.layers.5.mlp.up_proj.linear.weight",
682
+ "model.vision_tower.encoder.layers.5.post_attention_layernorm.weight",
683
+ "model.vision_tower.encoder.layers.5.post_feedforward_layernorm.weight",
684
+ "model.vision_tower.encoder.layers.5.pre_feedforward_layernorm.weight",
685
+ "model.vision_tower.encoder.layers.5.self_attn.k_norm.weight",
686
+ "model.vision_tower.encoder.layers.5.self_attn.k_proj.linear.weight",
687
+ "model.vision_tower.encoder.layers.5.self_attn.o_proj.linear.weight",
688
+ "model.vision_tower.encoder.layers.5.self_attn.q_norm.weight",
689
+ "model.vision_tower.encoder.layers.5.self_attn.q_proj.linear.weight",
690
+ "model.vision_tower.encoder.layers.5.self_attn.v_proj.linear.weight",
691
+ "model.vision_tower.encoder.layers.6.input_layernorm.weight",
692
+ "model.vision_tower.encoder.layers.6.mlp.down_proj.linear.weight",
693
+ "model.vision_tower.encoder.layers.6.mlp.gate_proj.linear.weight",
694
+ "model.vision_tower.encoder.layers.6.mlp.up_proj.linear.weight",
695
+ "model.vision_tower.encoder.layers.6.post_attention_layernorm.weight",
696
+ "model.vision_tower.encoder.layers.6.post_feedforward_layernorm.weight",
697
+ "model.vision_tower.encoder.layers.6.pre_feedforward_layernorm.weight",
698
+ "model.vision_tower.encoder.layers.6.self_attn.k_norm.weight",
699
+ "model.vision_tower.encoder.layers.6.self_attn.k_proj.linear.weight",
700
+ "model.vision_tower.encoder.layers.6.self_attn.o_proj.linear.weight",
701
+ "model.vision_tower.encoder.layers.6.self_attn.q_norm.weight",
702
+ "model.vision_tower.encoder.layers.6.self_attn.q_proj.linear.weight",
703
+ "model.vision_tower.encoder.layers.6.self_attn.v_proj.linear.weight",
704
+ "model.vision_tower.encoder.layers.7.input_layernorm.weight",
705
+ "model.vision_tower.encoder.layers.7.mlp.down_proj.linear.weight",
706
+ "model.vision_tower.encoder.layers.7.mlp.gate_proj.linear.weight",
707
+ "model.vision_tower.encoder.layers.7.mlp.up_proj.linear.weight",
708
+ "model.vision_tower.encoder.layers.7.post_attention_layernorm.weight",
709
+ "model.vision_tower.encoder.layers.7.post_feedforward_layernorm.weight",
710
+ "model.vision_tower.encoder.layers.7.pre_feedforward_layernorm.weight",
711
+ "model.vision_tower.encoder.layers.7.self_attn.k_norm.weight",
712
+ "model.vision_tower.encoder.layers.7.self_attn.k_proj.linear.weight",
713
+ "model.vision_tower.encoder.layers.7.self_attn.o_proj.linear.weight",
714
+ "model.vision_tower.encoder.layers.7.self_attn.q_norm.weight",
715
+ "model.vision_tower.encoder.layers.7.self_attn.q_proj.linear.weight",
716
+ "model.vision_tower.encoder.layers.7.self_attn.v_proj.linear.weight",
717
+ "model.vision_tower.encoder.layers.8.input_layernorm.weight",
718
+ "model.vision_tower.encoder.layers.8.mlp.down_proj.linear.weight",
719
+ "model.vision_tower.encoder.layers.8.mlp.gate_proj.linear.weight",
720
+ "model.vision_tower.encoder.layers.8.mlp.up_proj.linear.weight",
721
+ "model.vision_tower.encoder.layers.8.post_attention_layernorm.weight",
722
+ "model.vision_tower.encoder.layers.8.post_feedforward_layernorm.weight",
723
+ "model.vision_tower.encoder.layers.8.pre_feedforward_layernorm.weight",
724
+ "model.vision_tower.encoder.layers.8.self_attn.k_norm.weight",
725
+ "model.vision_tower.encoder.layers.8.self_attn.k_proj.linear.weight",
726
+ "model.vision_tower.encoder.layers.8.self_attn.o_proj.linear.weight",
727
+ "model.vision_tower.encoder.layers.8.self_attn.q_norm.weight",
728
+ "model.vision_tower.encoder.layers.8.self_attn.q_proj.linear.weight",
729
+ "model.vision_tower.encoder.layers.8.self_attn.v_proj.linear.weight",
730
+ "model.vision_tower.encoder.layers.9.input_layernorm.weight",
731
+ "model.vision_tower.encoder.layers.9.mlp.down_proj.linear.weight",
732
+ "model.vision_tower.encoder.layers.9.mlp.gate_proj.linear.weight",
733
+ "model.vision_tower.encoder.layers.9.mlp.up_proj.linear.weight",
734
+ "model.vision_tower.encoder.layers.9.post_attention_layernorm.weight",
735
+ "model.vision_tower.encoder.layers.9.post_feedforward_layernorm.weight",
736
+ "model.vision_tower.encoder.layers.9.pre_feedforward_layernorm.weight",
737
+ "model.vision_tower.encoder.layers.9.self_attn.k_norm.weight",
738
+ "model.vision_tower.encoder.layers.9.self_attn.k_proj.linear.weight",
739
+ "model.vision_tower.encoder.layers.9.self_attn.o_proj.linear.weight",
740
+ "model.vision_tower.encoder.layers.9.self_attn.q_norm.weight",
741
+ "model.vision_tower.encoder.layers.9.self_attn.q_proj.linear.weight",
742
+ "model.vision_tower.encoder.layers.9.self_attn.v_proj.linear.weight",
743
+ "model.vision_tower.patch_embedder.input_proj.weight",
744
+ "model.vision_tower.patch_embedder.position_embedding_table"
745
+ ]
746
+ },
747
+ "max_steps": 768,
748
+ "mode": "direct",
749
+ "optimized_completions_per_second": 2.3449,
750
+ "parent": "/workspace/parent-control/grafts/control",
751
+ "sampler": "/workspace/runs/rl-direct/train/sampler",
752
+ "scheduler": "constant",
753
+ "schema_version": 1,
754
+ "seconds_per_optimizer_update": 13.647,
755
+ "smoke": false,
756
+ "state": "/workspace/runs/rl-direct/train/trainer/checkpoint-768",
757
+ "status": "complete",
758
+ "training_elapsed_seconds": 10480.548,
759
+ "worklist": {
760
+ "difficulty": null,
761
+ "manifest_sha256": "40ac035634fbabb1ccaca638d58f65eeee6a620769de244adb835910cc13bfcf",
762
+ "pool_episodes": 8192,
763
+ "prompt_surface": {
764
+ "contract_prefix": "Assignment: ",
765
+ "forbidden_instruction": "wording and layout are up to you",
766
+ "rows_checked": 8192,
767
+ "same_prompts_for_direct_and_thinking": true,
768
+ "version": "template_diversity_v1"
769
+ },
770
+ "rows_surface_check": {
771
+ "prompt_templates": 90,
772
+ "rows_checked": 6144,
773
+ "version": "template_diversity_v1"
774
+ },
775
+ "sampling": {
776
+ "bias": 0,
777
+ "distinct_episodes_drawn": 4323,
778
+ "draws_per_episode_histogram": {
779
+ "1": 2920,
780
+ "2": 1059,
781
+ "3": 281,
782
+ "4": 53,
783
+ "5": 9,
784
+ "6": 1
785
+ },
786
+ "max_draws_for_one_episode": 6,
787
+ "max_expected_draws": 0.75,
788
+ "max_weight": 1.0,
789
+ "mean_weight": 1.0,
790
+ "min_appearance_probability": 0.527655071748625,
791
+ "min_expected_draws": 0.75,
792
+ "min_to_max_ratio": 1.0,
793
+ "min_weight": 1.0,
794
+ "pool_coverage": 0.5277099609375,
795
+ "pool_sha256": "ffe7fa151b5a556c2450dfbbd472feed687c8321c40da0a50d4f4ab05a6fddc2",
796
+ "prefix_stable_in_rows": true,
797
+ "prior_pseudocounts": 0.5,
798
+ "rows": 6144,
799
+ "sampler_seed": 11724338123291118077,
800
+ "scheme": "weighted-with-replacement over the full pool",
801
+ "seed": 42,
802
+ "seed_material": [
803
+ "ffe7fa151b5a556c2450dfbbd472feed687c8321c40da0a50d4f4ab05a6fddc2",
804
+ "057e97aac2166ac6cadb7d17f9d90feb103eaa6352aee103e594faac22939345",
805
+ "0.000000000000e+00"
806
+ ],
807
+ "sequence_sha256": "d93c80de729e0bb134c4f1172e4995c0b6817aae8f3c28e86d2abeed6f43fc2d",
808
+ "weighting": "uniform",
809
+ "weights_sha256": "057e97aac2166ac6cadb7d17f9d90feb103eaa6352aee103e594faac22939345"
810
+ },
811
+ "shared_across_cells": true
812
+ }
813
+ }
gemma4_26b_a4b_190m/control/training/direct/ROLLOUT_AUDIT.json ADDED
@@ -0,0 +1,22 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ {
2
+ "counts": {
3
+ "parser_unsafe": 2870,
4
+ "parser_valid": 46171,
5
+ "reward_positive": 39154,
6
+ "rows": 49152,
7
+ "truncated": 718
8
+ },
9
+ "failures": [],
10
+ "files": [
11
+ "/workspace/runs/rl-direct/rollouts/raw_rollouts.rank-0.jsonl"
12
+ ],
13
+ "mode": "direct",
14
+ "passed": true,
15
+ "reward_positive_review": {
16
+ "approval": "manual review required before the next RL phase",
17
+ "path": "/workspace/runs/rl-direct/REWARD_POSITIVE_REVIEW.jsonl",
18
+ "rows": 39154,
19
+ "sha256": "50cdd1f86ada36c84c196bce1d793aa020d4d664fd32c26307f46427a152b742"
20
+ },
21
+ "schema_version": 1
22
+ }
gemma4_26b_a4b_190m/control/training/direct/TELEMETRY.json ADDED
The diff for this file is too large to render. See raw diff
 
rollouts/gemma4_26b_a4b_190m/control-direct/raw_rollouts.rank-0.jsonl.gz ADDED
@@ -0,0 +1,3 @@
 
 
 
 
1
+ version https://git-lfs.github.com/spec/v1
2
+ oid sha256:bb71421de65a4e5059c954fa6cd4be3c2fa548acc8589052eda9d20eae31d04b
3
+ size 8569182