sidbaines commited on
Commit
f567b2c
·
verified ·
1 Parent(s): 32e13b4

clean repo: gemma4_26b_a4b_190m|control (1/1)

Browse files
gemma4_26b_a4b_190m/control/rlvr/thinking/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": "arcadia-impact/scimt-dispatch-clean-v1/gemma4_26b_a4b_graft/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": 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.19.self_attn.o_proj",
35
+ "model.language_model.layers.22.self_attn.k_proj",
36
+ "model.language_model.layers.1.self_attn.q_proj",
37
+ "model.language_model.layers.13.self_attn.v_proj",
38
+ "model.language_model.layers.14.self_attn.k_proj",
39
+ "model.language_model.layers.25.self_attn.o_proj",
40
+ "model.language_model.layers.15.self_attn.o_proj",
41
+ "model.language_model.layers.28.self_attn.v_proj",
42
+ "model.language_model.layers.3.self_attn.q_proj",
43
+ "model.language_model.layers.11.self_attn.o_proj",
44
+ "model.language_model.layers.4.self_attn.o_proj",
45
+ "model.language_model.layers.19.self_attn.k_proj",
46
+ "model.language_model.layers.27.self_attn.q_proj",
47
+ "model.language_model.layers.0.self_attn.q_proj",
48
+ "model.language_model.layers.26.self_attn.o_proj",
49
+ "model.language_model.layers.8.self_attn.q_proj",
50
+ "model.language_model.layers.0.self_attn.v_proj",
51
+ "model.language_model.layers.3.self_attn.v_proj",
52
+ "model.language_model.layers.5.self_attn.o_proj",
53
+ "model.language_model.layers.23.self_attn.o_proj",
54
+ "model.language_model.layers.25.self_attn.q_proj",
55
+ "model.language_model.layers.6.self_attn.o_proj",
56
+ "model.language_model.layers.3.self_attn.k_proj",
57
+ "model.language_model.layers.4.self_attn.q_proj",
58
+ "model.language_model.layers.19.self_attn.v_proj",
59
+ "model.language_model.layers.6.self_attn.q_proj",
60
+ "model.language_model.layers.4.self_attn.v_proj",
61
+ "model.language_model.layers.10.self_attn.o_proj",
62
+ "model.language_model.layers.27.self_attn.k_proj",
63
+ "model.language_model.layers.14.self_attn.o_proj",
64
+ "model.language_model.layers.12.self_attn.o_proj",
65
+ "model.language_model.layers.17.self_attn.k_proj",
66
+ "model.language_model.layers.14.self_attn.q_proj",
67
+ "model.language_model.layers.24.self_attn.q_proj",
68
+ "model.language_model.layers.18.self_attn.v_proj",
69
+ "model.language_model.layers.8.self_attn.o_proj",
70
+ "model.language_model.layers.9.self_attn.k_proj",
71
+ "model.language_model.layers.21.self_attn.q_proj",
72
+ "model.language_model.layers.22.self_attn.q_proj",
73
+ "model.language_model.layers.1.self_attn.o_proj",
74
+ "model.language_model.layers.7.self_attn.o_proj",
75
+ "model.language_model.layers.25.self_attn.v_proj",
76
+ "model.language_model.layers.29.self_attn.k_proj",
77
+ "model.language_model.layers.21.self_attn.o_proj",
78
+ "model.language_model.layers.18.self_attn.o_proj",
79
+ "model.language_model.layers.29.self_attn.q_proj",
80
+ "model.language_model.layers.6.self_attn.k_proj",
81
+ "model.language_model.layers.7.self_attn.v_proj",
82
+ "model.language_model.layers.13.self_attn.k_proj",
83
+ "model.language_model.layers.15.self_attn.v_proj",
84
+ "model.language_model.layers.26.self_attn.k_proj",
85
+ "model.language_model.layers.20.self_attn.o_proj",
86
+ "model.language_model.layers.3.self_attn.o_proj",
87
+ "model.language_model.layers.15.self_attn.q_proj",
88
+ "model.language_model.layers.27.self_attn.o_proj",
89
+ "model.language_model.layers.17.self_attn.q_proj",
90
+ "model.language_model.layers.2.self_attn.o_proj",
91
+ "model.language_model.layers.8.self_attn.k_proj",
92
+ "model.language_model.layers.24.self_attn.o_proj",
93
+ "model.language_model.layers.18.self_attn.q_proj",
94
+ "model.language_model.layers.10.self_attn.k_proj",
95
+ "model.language_model.layers.9.self_attn.o_proj",
96
+ "model.language_model.layers.1.self_attn.k_proj",
97
+ "model.language_model.layers.7.self_attn.q_proj",
98
+ "model.language_model.layers.13.self_attn.q_proj",
99
+ "model.language_model.layers.5.self_attn.k_proj",
100
+ "model.language_model.layers.16.self_attn.o_proj",
101
+ "model.language_model.layers.27.self_attn.v_proj",
102
+ "model.language_model.layers.11.self_attn.k_proj",
103
+ "model.language_model.layers.5.self_attn.q_proj",
104
+ "model.language_model.layers.22.self_attn.v_proj",
105
+ "model.language_model.layers.29.self_attn.o_proj",
106
+ "model.language_model.layers.18.self_attn.k_proj",
107
+ "model.language_model.layers.2.self_attn.v_proj",
108
+ "model.language_model.layers.10.self_attn.q_proj",
109
+ "model.language_model.layers.20.self_attn.v_proj",
110
+ "model.language_model.layers.28.self_attn.k_proj",
111
+ "model.language_model.layers.28.self_attn.q_proj",
112
+ "model.language_model.layers.2.self_attn.q_proj",
113
+ "model.language_model.layers.11.self_attn.q_proj",
114
+ "model.language_model.layers.10.self_attn.v_proj",
115
+ "model.language_model.layers.12.self_attn.q_proj",
116
+ "model.language_model.layers.24.self_attn.v_proj",
117
+ "model.language_model.layers.26.self_attn.q_proj",
118
+ "model.language_model.layers.7.self_attn.k_proj",
119
+ "model.language_model.layers.14.self_attn.v_proj",
120
+ "model.language_model.layers.12.self_attn.k_proj",
121
+ "model.language_model.layers.25.self_attn.k_proj",
122
+ "model.language_model.layers.28.self_attn.o_proj",
123
+ "model.language_model.layers.23.self_attn.k_proj",
124
+ "model.language_model.layers.4.self_attn.k_proj",
125
+ "model.language_model.layers.15.self_attn.k_proj",
126
+ "model.language_model.layers.6.self_attn.v_proj",
127
+ "model.language_model.layers.8.self_attn.v_proj",
128
+ "model.language_model.layers.17.self_attn.o_proj",
129
+ "model.language_model.layers.21.self_attn.k_proj",
130
+ "model.language_model.layers.21.self_attn.v_proj",
131
+ "model.language_model.layers.24.self_attn.k_proj",
132
+ "model.language_model.layers.20.self_attn.q_proj",
133
+ "model.language_model.layers.9.self_attn.q_proj",
134
+ "model.language_model.layers.2.self_attn.k_proj",
135
+ "model.language_model.layers.16.self_attn.k_proj",
136
+ "model.language_model.layers.12.self_attn.v_proj",
137
+ "model.language_model.layers.26.self_attn.v_proj",
138
+ "model.language_model.layers.0.self_attn.o_proj",
139
+ "model.language_model.layers.9.self_attn.v_proj",
140
+ "model.language_model.layers.19.self_attn.q_proj",
141
+ "model.language_model.layers.16.self_attn.q_proj",
142
+ "model.language_model.layers.20.self_attn.k_proj",
143
+ "model.language_model.layers.1.self_attn.v_proj",
144
+ "model.language_model.layers.13.self_attn.o_proj",
145
+ "model.language_model.layers.22.self_attn.o_proj",
146
+ "model.language_model.layers.0.self_attn.k_proj",
147
+ "model.language_model.layers.23.self_attn.q_proj",
148
+ "model.language_model.layers.16.self_attn.v_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/thinking/adapter_model.safetensors ADDED
@@ -0,0 +1,3 @@
 
 
 
 
1
+ version https://git-lfs.github.com/spec/v1
2
+ oid sha256:26dcaa66a290c7b94fec5d006a5b83f36126d52b2ab53ad0c25b9b10167fb576
3
+ size 183863120
gemma4_26b_a4b_190m/control/training/pre_aft-cap32768/leg.log ADDED
@@ -0,0 +1,11 @@
 
 
 
 
 
 
 
 
 
 
 
 
1
+ provenance OK: 5613 files at 39bb3d35a8ba
2
+ === 2026-09-11T10:37:22Z provenance verified (gitless, 39bb3d35a8bac2ad966b6b0537342a0437a73b82) ===
3
+ === 2026-09-11T10:37:22Z endpoint control-pre_aft step=0 cap=32768 window=34816 arm=control ===
4
+ === 2026-09-11T10:37:22Z pulling control graft (50M) from arcadia-impact/scimt-dispatch-rlvr-gemma4-26b-v1 ===
5
+ /workspace/venvs/pod-eval/lib/python3.11/site-packages/huggingface_hub/constants.py:299: FutureWarning: The `HF_HUB_ENABLE_HF_TRANSFER` environment variable is deprecated as 'hf_transfer' is not used anymore. Please use `HF_XET_HIGH_PERFORMANCE` instead to enable high performance transfer with Xet. Visit https://huggingface.co/docs/huggingface_hub/package_reference/environment_variables#hfxethighperformance for more details.
6
+ warnings.warn(
7
+
8
+ downloaded arcadia-impact/scimt-dispatch-rlvr-gemma4-26b-v1 grafts/control/* -> /workspace/parent-control
9
+ === 2026-09-11T10:43:09Z sweep: thinking, heldout surface, cap 32768 ===
10
+ >>> control-pre_aft step 0 (cap 32768) eval_trained_conflict__heldout [rlvr] charter_share=0.364 ci=[0.337,0.389] n=1760 episodes=1244 surfaces=None
11
+ === 2026-09-11T13:15:12Z publishing ===
gemma4_26b_a4b_190m/control/training/thinking-cap12288/leg.log ADDED
@@ -0,0 +1,17 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ provenance OK: 5613 files at 39bb3d35a8ba
2
+ === 2026-09-11T10:34:58Z provenance verified (gitless, 39bb3d35a8bac2ad966b6b0537342a0437a73b82) ===
3
+ === 2026-09-11T10:34:58Z endpoint control-thinking step=256 cap=12288 window=14336 arm=control ===
4
+ === 2026-09-11T10:34:58Z pulling control graft (50M) from arcadia-impact/scimt-dispatch-rlvr-gemma4-26b-v1 ===
5
+ /workspace/venvs/pod-eval/lib/python3.11/site-packages/huggingface_hub/constants.py:299: FutureWarning: The `HF_HUB_ENABLE_HF_TRANSFER` environment variable is deprecated as 'hf_transfer' is not used anymore. Please use `HF_XET_HIGH_PERFORMANCE` instead to enable high performance transfer with Xet. Visit https://huggingface.co/docs/huggingface_hub/package_reference/environment_variables#hfxethighperformance for more details.
6
+ warnings.warn(
7
+
8
+ downloaded arcadia-impact/scimt-dispatch-rlvr-gemma4-26b-v1 grafts/control/* -> /workspace/parent-control
9
+ === 2026-09-11T10:38:28Z pulling control-thinking step-256 adapter from sidbaines/scimt-dispatch-gemma4-26b-charter-190m-graft-v1 ===
10
+ /workspace/venvs/pod-eval/lib/python3.11/site-packages/huggingface_hub/constants.py:299: FutureWarning: The `HF_HUB_ENABLE_HF_TRANSFER` environment variable is deprecated as 'hf_transfer' is not used anymore. Please use `HF_XET_HIGH_PERFORMANCE` instead to enable high performance transfer with Xet. Visit https://huggingface.co/docs/huggingface_hub/package_reference/environment_variables#hfxethighperformance for more details.
11
+ warnings.warn(
12
+
13
+ downloaded sidbaines/scimt-dispatch-gemma4-26b-charter-190m-graft-v1 rl-checkpoints/control-thinking/step-256/* -> /workspace/adapter
14
+ === 2026-09-11T10:38:43Z sweep: thinking, heldout surface, cap 12288 ===
15
+ >>> control-thinking step 256 (cap 12288) eval_trained_conflict__heldout [rlvr] charter_share=0.354 ci=[0.332,0.375] n=2253 episodes=1623 surfaces=None
16
+ === 2026-09-11T12:22:52Z publishing ===
17
+ === 2026-09-11T12:22:56Z ENDPOINT COMPLETE -- /workspace/evals/thinking/control-thinking-step256.json ===