Hugging Face's logo

arcadia-impact
/
scimt-dispatch-models

Text Generation
Transformers
Safetensors
PEFT
gemma-3
continued-pretraining
sft
lora
synthetic-data
alignment
midtraining
scimt
Model card Files
xet
 Community
sidbaines commited on
Commit
2581623
·
verified ·
1 Parent(s): 6d738f6

add rl_grpo/control_4x_thinking (copied from sidbaines/scimt-prior-coins-dispatch-sdf-aft-v1)

Browse files
Files changed (15) hide show
  1. .gitattributes +1 -0
  2. rl_grpo/control_4x_thinking/RL_TRAINED.json +13 -0
  3. rl_grpo/control_4x_thinking/checkpoint-128/adapter_config.json +191 -0
  4. rl_grpo/control_4x_thinking/checkpoint-128/adapter_model.safetensors +3 -0
  5. rl_grpo/control_4x_thinking/checkpoint-16/adapter_config.json +191 -0
  6. rl_grpo/control_4x_thinking/checkpoint-16/adapter_model.safetensors +3 -0
  7. rl_grpo/control_4x_thinking/checkpoint-256/adapter_config.json +191 -0
  8. rl_grpo/control_4x_thinking/checkpoint-256/adapter_model.safetensors +3 -0
  9. rl_grpo/control_4x_thinking/checkpoint-256/trainer_state.json +834 -0
  10. rl_grpo/control_4x_thinking/checkpoint-32/adapter_config.json +191 -0
  11. rl_grpo/control_4x_thinking/checkpoint-32/adapter_model.safetensors +3 -0
  12. rl_grpo/control_4x_thinking/checkpoint-64/adapter_config.json +191 -0
  13. rl_grpo/control_4x_thinking/checkpoint-64/adapter_model.safetensors +3 -0
  14. rl_grpo/control_4x_thinking/checkpoint.json +132 -0
  15. rl_grpo/control_4x_thinking/logs/raw_rollouts.rank-0.jsonl +3 -0
.gitattributes CHANGED
@@ -139,3 +139,4 @@ rl_grpo/coin_real_4x_direct/logs/raw_rollouts.rank-0.jsonl filter=lfs diff=lfs m
139
  rl_grpo/charter_real_4x_direct/logs/raw_rollouts.rank-0.jsonl filter=lfs diff=lfs merge=lfs -text
140
  rl_grpo/charter_real_4x_thinking/logs/raw_rollouts.rank-0.jsonl filter=lfs diff=lfs merge=lfs -text
141
  rl_grpo/coin_real_4x_thinking/logs/raw_rollouts.rank-0.jsonl filter=lfs diff=lfs merge=lfs -text
 
 
139
  rl_grpo/charter_real_4x_direct/logs/raw_rollouts.rank-0.jsonl filter=lfs diff=lfs merge=lfs -text
140
  rl_grpo/charter_real_4x_thinking/logs/raw_rollouts.rank-0.jsonl filter=lfs diff=lfs merge=lfs -text
141
  rl_grpo/coin_real_4x_thinking/logs/raw_rollouts.rank-0.jsonl filter=lfs diff=lfs merge=lfs -text
142
+ rl_grpo/control_4x_thinking/logs/raw_rollouts.rank-0.jsonl filter=lfs diff=lfs merge=lfs -text
rl_grpo/control_4x_thinking/RL_TRAINED.json ADDED
@@ -0,0 +1,13 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ {
2
+ "version": "dispatch_rl_v2",
3
+ "label": "control_4x_thinking",
4
+ "mode": "thinking",
5
+ "parent": "/workspace/rl3t_c/parent",
6
+ "dataset": "/workspace/rl3t_c/data/thinking/train.jsonl",
7
+ "episodes": 8192,
8
+ "adapter": "/workspace/rl3t_c/training/control_4x_thinking/train/sampler",
9
+ "minutes": 193.5,
10
+ "dropped_overlong": 0,
11
+ "temperature": 0.7,
12
+ "learning_rate": 1e-05
13
+ }
rl_grpo/control_4x_thinking/checkpoint-128/adapter_config.json ADDED
@@ -0,0 +1,191 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ {
2
+ "alora_invocation_tokens": null,
3
+ "alpha_pattern": {},
4
+ "arrow_config": null,
5
+ "auto_mapping": null,
6
+ "base_model_name_or_path": "/workspace/rl3t_c/parent",
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": 64,
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": 32,
31
+ "rank_pattern": {},
32
+ "revision": null,
33
+ "target_modules": [
34
+ "32.self_attn.v_proj",
35
+ "language_model.layers.12.self_attn.q_proj",
36
+ "language_model.layers.16.self_attn.q_proj",
37
+ "language_model.layers.10.self_attn.q_proj",
38
+ "language_model.layers.17.self_attn.k_proj",
39
+ "language_model.layers.10.self_attn.v_proj",
40
+ "34.self_attn.k_proj",
41
+ "45.self_attn.k_proj",
42
+ "language_model.layers.17.self_attn.v_proj",
43
+ "44.self_attn.k_proj",
44
+ "46.self_attn.v_proj",
45
+ "39.self_attn.k_proj",
46
+ "41.self_attn.v_proj",
47
+ "30.self_attn.v_proj",
48
+ "language_model.layers.17.self_attn.q_proj",
49
+ "33.self_attn.q_proj",
50
+ "language_model.layers.6.self_attn.q_proj",
51
+ "language_model.layers.23.self_attn.v_proj",
52
+ "language_model.layers.7.self_attn.k_proj",
53
+ "language_model.layers.22.self_attn.v_proj",
54
+ "language_model.layers.4.self_attn.v_proj",
55
+ "language_model.layers.24.self_attn.k_proj",
56
+ "language_model.layers.2.self_attn.v_proj",
57
+ "38.self_attn.v_proj",
58
+ "language_model.layers.11.self_attn.k_proj",
59
+ "language_model.layers.25.self_attn.v_proj",
60
+ "language_model.layers.1.self_attn.k_proj",
61
+ "27.self_attn.k_proj",
62
+ "27.self_attn.q_proj",
63
+ "language_model.layers.19.self_attn.k_proj",
64
+ "39.self_attn.v_proj",
65
+ "28.self_attn.q_proj",
66
+ "32.self_attn.q_proj",
67
+ "46.self_attn.q_proj",
68
+ "language_model.layers.15.self_attn.q_proj",
69
+ "o_proj",
70
+ "language_model.layers.26.self_attn.k_proj",
71
+ "language_model.layers.21.self_attn.v_proj",
72
+ "language_model.layers.23.self_attn.k_proj",
73
+ "43.self_attn.q_proj",
74
+ "language_model.layers.23.self_attn.q_proj",
75
+ "40.self_attn.v_proj",
76
+ "28.self_attn.v_proj",
77
+ "language_model.layers.9.self_attn.v_proj",
78
+ "47.self_attn.k_proj",
79
+ "37.self_attn.q_proj",
80
+ "language_model.layers.8.self_attn.k_proj",
81
+ "language_model.layers.0.self_attn.k_proj",
82
+ "36.self_attn.v_proj",
83
+ "language_model.layers.0.self_attn.q_proj",
84
+ "language_model.layers.18.self_attn.v_proj",
85
+ "language_model.layers.5.self_attn.q_proj",
86
+ "language_model.layers.18.self_attn.q_proj",
87
+ "38.self_attn.q_proj",
88
+ "35.self_attn.q_proj",
89
+ "language_model.layers.15.self_attn.v_proj",
90
+ "47.self_attn.v_proj",
91
+ "language_model.layers.0.self_attn.v_proj",
92
+ "language_model.layers.20.self_attn.k_proj",
93
+ "language_model.layers.7.self_attn.q_proj",
94
+ "45.self_attn.q_proj",
95
+ "language_model.layers.2.self_attn.q_proj",
96
+ "29.self_attn.k_proj",
97
+ "33.self_attn.v_proj",
98
+ "language_model.layers.22.self_attn.k_proj",
99
+ "language_model.layers.10.self_attn.k_proj",
100
+ "44.self_attn.q_proj",
101
+ "42.self_attn.v_proj",
102
+ "language_model.layers.20.self_attn.v_proj",
103
+ "language_model.layers.5.self_attn.v_proj",
104
+ "33.self_attn.k_proj",
105
+ "language_model.layers.20.self_attn.q_proj",
106
+ "language_model.layers.1.self_attn.q_proj",
107
+ "language_model.layers.11.self_attn.v_proj",
108
+ "34.self_attn.q_proj",
109
+ "language_model.layers.8.self_attn.q_proj",
110
+ "language_model.layers.4.self_attn.q_proj",
111
+ "up_proj",
112
+ "language_model.layers.9.self_attn.k_proj",
113
+ "language_model.layers.6.self_attn.v_proj",
114
+ "30.self_attn.q_proj",
115
+ "language_model.layers.3.self_attn.q_proj",
116
+ "language_model.layers.26.self_attn.q_proj",
117
+ "27.self_attn.v_proj",
118
+ "language_model.layers.21.self_attn.k_proj",
119
+ "language_model.layers.12.self_attn.v_proj",
120
+ "39.self_attn.q_proj",
121
+ "language_model.layers.19.self_attn.v_proj",
122
+ "language_model.layers.5.self_attn.k_proj",
123
+ "29.self_attn.q_proj",
124
+ "43.self_attn.v_proj",
125
+ "language_model.layers.9.self_attn.q_proj",
126
+ "language_model.layers.19.self_attn.q_proj",
127
+ "35.self_attn.v_proj",
128
+ "language_model.layers.21.self_attn.q_proj",
129
+ "language_model.layers.4.self_attn.k_proj",
130
+ "32.self_attn.k_proj",
131
+ "language_model.layers.13.self_attn.v_proj",
132
+ "37.self_attn.v_proj",
133
+ "38.self_attn.k_proj",
134
+ "language_model.layers.8.self_attn.v_proj",
135
+ "34.self_attn.v_proj",
136
+ "29.self_attn.v_proj",
137
+ "down_proj",
138
+ "language_model.layers.14.self_attn.k_proj",
139
+ "46.self_attn.k_proj",
140
+ "language_model.layers.3.self_attn.k_proj",
141
+ "37.self_attn.k_proj",
142
+ "41.self_attn.q_proj",
143
+ "42.self_attn.q_proj",
144
+ "language_model.layers.24.self_attn.q_proj",
145
+ "language_model.layers.14.self_attn.v_proj",
146
+ "language_model.layers.6.self_attn.k_proj",
147
+ "language_model.layers.25.self_attn.k_proj",
148
+ "language_model.layers.24.self_attn.v_proj",
149
+ "35.self_attn.k_proj",
150
+ "language_model.layers.22.self_attn.q_proj",
151
+ "41.self_attn.k_proj",
152
+ "40.self_attn.q_proj",
153
+ "45.self_attn.v_proj",
154
+ "language_model.layers.16.self_attn.v_proj",
155
+ "language_model.layers.18.self_attn.k_proj",
156
+ "28.self_attn.k_proj",
157
+ "language_model.layers.15.self_attn.k_proj",
158
+ "language_model.layers.14.self_attn.q_proj",
159
+ "36.self_attn.k_proj",
160
+ "language_model.layers.7.self_attn.v_proj",
161
+ "44.self_attn.v_proj",
162
+ "gate_proj",
163
+ "language_model.layers.13.self_attn.q_proj",
164
+ "language_model.layers.25.self_attn.q_proj",
165
+ "language_model.layers.3.self_attn.v_proj",
166
+ "language_model.layers.1.self_attn.v_proj",
167
+ "language_model.layers.11.self_attn.q_proj",
168
+ "31.self_attn.v_proj",
169
+ "language_model.layers.12.self_attn.k_proj",
170
+ "43.self_attn.k_proj",
171
+ "31.self_attn.q_proj",
172
+ "42.self_attn.k_proj",
173
+ "36.self_attn.q_proj",
174
+ "language_model.layers.26.self_attn.v_proj",
175
+ "language_model.layers.13.self_attn.k_proj",
176
+ "40.self_attn.k_proj",
177
+ "language_model.layers.2.self_attn.k_proj",
178
+ "47.self_attn.q_proj",
179
+ "language_model.layers.16.self_attn.k_proj",
180
+ "31.self_attn.k_proj",
181
+ "30.self_attn.k_proj"
182
+ ],
183
+ "target_parameters": null,
184
+ "task_type": "CAUSAL_LM",
185
+ "trainable_token_indices": null,
186
+ "use_bdlora": null,
187
+ "use_dora": false,
188
+ "use_qalora": false,
189
+ "use_rslora": false,
190
+ "velora_config": null
191
+ }
rl_grpo/control_4x_thinking/checkpoint-128/adapter_model.safetensors ADDED
@@ -0,0 +1,3 @@
 
 
 
 
1
+ version https://git-lfs.github.com/spec/v1
2
+ oid sha256:89d9af20a063c9b6b655e1b5af586eb1c8e87465657732b7930db6d9577c5391
3
+ size 523864568
rl_grpo/control_4x_thinking/checkpoint-16/adapter_config.json ADDED
@@ -0,0 +1,191 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ {
2
+ "alora_invocation_tokens": null,
3
+ "alpha_pattern": {},
4
+ "arrow_config": null,
5
+ "auto_mapping": null,
6
+ "base_model_name_or_path": "/workspace/rl3t_c/parent",
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": 64,
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": 32,
31
+ "rank_pattern": {},
32
+ "revision": null,
33
+ "target_modules": [
34
+ "32.self_attn.v_proj",
35
+ "language_model.layers.12.self_attn.q_proj",
36
+ "language_model.layers.16.self_attn.q_proj",
37
+ "language_model.layers.10.self_attn.q_proj",
38
+ "language_model.layers.17.self_attn.k_proj",
39
+ "language_model.layers.10.self_attn.v_proj",
40
+ "34.self_attn.k_proj",
41
+ "45.self_attn.k_proj",
42
+ "language_model.layers.17.self_attn.v_proj",
43
+ "44.self_attn.k_proj",
44
+ "46.self_attn.v_proj",
45
+ "39.self_attn.k_proj",
46
+ "41.self_attn.v_proj",
47
+ "30.self_attn.v_proj",
48
+ "language_model.layers.17.self_attn.q_proj",
49
+ "33.self_attn.q_proj",
50
+ "language_model.layers.6.self_attn.q_proj",
51
+ "language_model.layers.23.self_attn.v_proj",
52
+ "language_model.layers.7.self_attn.k_proj",
53
+ "language_model.layers.22.self_attn.v_proj",
54
+ "language_model.layers.4.self_attn.v_proj",
55
+ "language_model.layers.24.self_attn.k_proj",
56
+ "language_model.layers.2.self_attn.v_proj",
57
+ "38.self_attn.v_proj",
58
+ "language_model.layers.11.self_attn.k_proj",
59
+ "language_model.layers.25.self_attn.v_proj",
60
+ "language_model.layers.1.self_attn.k_proj",
61
+ "27.self_attn.k_proj",
62
+ "27.self_attn.q_proj",
63
+ "language_model.layers.19.self_attn.k_proj",
64
+ "39.self_attn.v_proj",
65
+ "28.self_attn.q_proj",
66
+ "32.self_attn.q_proj",
67
+ "46.self_attn.q_proj",
68
+ "language_model.layers.15.self_attn.q_proj",
69
+ "o_proj",
70
+ "language_model.layers.26.self_attn.k_proj",
71
+ "language_model.layers.21.self_attn.v_proj",
72
+ "language_model.layers.23.self_attn.k_proj",
73
+ "43.self_attn.q_proj",
74
+ "language_model.layers.23.self_attn.q_proj",
75
+ "40.self_attn.v_proj",
76
+ "28.self_attn.v_proj",
77
+ "language_model.layers.9.self_attn.v_proj",
78
+ "47.self_attn.k_proj",
79
+ "37.self_attn.q_proj",
80
+ "language_model.layers.8.self_attn.k_proj",
81
+ "language_model.layers.0.self_attn.k_proj",
82
+ "36.self_attn.v_proj",
83
+ "language_model.layers.0.self_attn.q_proj",
84
+ "language_model.layers.18.self_attn.v_proj",
85
+ "language_model.layers.5.self_attn.q_proj",
86
+ "language_model.layers.18.self_attn.q_proj",
87
+ "38.self_attn.q_proj",
88
+ "35.self_attn.q_proj",
89
+ "language_model.layers.15.self_attn.v_proj",
90
+ "47.self_attn.v_proj",
91
+ "language_model.layers.0.self_attn.v_proj",
92
+ "language_model.layers.20.self_attn.k_proj",
93
+ "language_model.layers.7.self_attn.q_proj",
94
+ "45.self_attn.q_proj",
95
+ "language_model.layers.2.self_attn.q_proj",
96
+ "29.self_attn.k_proj",
97
+ "33.self_attn.v_proj",
98
+ "language_model.layers.22.self_attn.k_proj",
99
+ "language_model.layers.10.self_attn.k_proj",
100
+ "44.self_attn.q_proj",
101
+ "42.self_attn.v_proj",
102
+ "language_model.layers.20.self_attn.v_proj",
103
+ "language_model.layers.5.self_attn.v_proj",
104
+ "33.self_attn.k_proj",
105
+ "language_model.layers.20.self_attn.q_proj",
106
+ "language_model.layers.1.self_attn.q_proj",
107
+ "language_model.layers.11.self_attn.v_proj",
108
+ "34.self_attn.q_proj",
109
+ "language_model.layers.8.self_attn.q_proj",
110
+ "language_model.layers.4.self_attn.q_proj",
111
+ "up_proj",
112
+ "language_model.layers.9.self_attn.k_proj",
113
+ "language_model.layers.6.self_attn.v_proj",
114
+ "30.self_attn.q_proj",
115
+ "language_model.layers.3.self_attn.q_proj",
116
+ "language_model.layers.26.self_attn.q_proj",
117
+ "27.self_attn.v_proj",
118
+ "language_model.layers.21.self_attn.k_proj",
119
+ "language_model.layers.12.self_attn.v_proj",
120
+ "39.self_attn.q_proj",
121
+ "language_model.layers.19.self_attn.v_proj",
122
+ "language_model.layers.5.self_attn.k_proj",
123
+ "29.self_attn.q_proj",
124
+ "43.self_attn.v_proj",
125
+ "language_model.layers.9.self_attn.q_proj",
126
+ "language_model.layers.19.self_attn.q_proj",
127
+ "35.self_attn.v_proj",
128
+ "language_model.layers.21.self_attn.q_proj",
129
+ "language_model.layers.4.self_attn.k_proj",
130
+ "32.self_attn.k_proj",
131
+ "language_model.layers.13.self_attn.v_proj",
132
+ "37.self_attn.v_proj",
133
+ "38.self_attn.k_proj",
134
+ "language_model.layers.8.self_attn.v_proj",
135
+ "34.self_attn.v_proj",
136
+ "29.self_attn.v_proj",
137
+ "down_proj",
138
+ "language_model.layers.14.self_attn.k_proj",
139
+ "46.self_attn.k_proj",
140
+ "language_model.layers.3.self_attn.k_proj",
141
+ "37.self_attn.k_proj",
142
+ "41.self_attn.q_proj",
143
+ "42.self_attn.q_proj",
144
+ "language_model.layers.24.self_attn.q_proj",
145
+ "language_model.layers.14.self_attn.v_proj",
146
+ "language_model.layers.6.self_attn.k_proj",
147
+ "language_model.layers.25.self_attn.k_proj",
148
+ "language_model.layers.24.self_attn.v_proj",
149
+ "35.self_attn.k_proj",
150
+ "language_model.layers.22.self_attn.q_proj",
151
+ "41.self_attn.k_proj",
152
+ "40.self_attn.q_proj",
153
+ "45.self_attn.v_proj",
154
+ "language_model.layers.16.self_attn.v_proj",
155
+ "language_model.layers.18.self_attn.k_proj",
156
+ "28.self_attn.k_proj",
157
+ "language_model.layers.15.self_attn.k_proj",
158
+ "language_model.layers.14.self_attn.q_proj",
159
+ "36.self_attn.k_proj",
160
+ "language_model.layers.7.self_attn.v_proj",
161
+ "44.self_attn.v_proj",
162
+ "gate_proj",
163
+ "language_model.layers.13.self_attn.q_proj",
164
+ "language_model.layers.25.self_attn.q_proj",
165
+ "language_model.layers.3.self_attn.v_proj",
166
+ "language_model.layers.1.self_attn.v_proj",
167
+ "language_model.layers.11.self_attn.q_proj",
168
+ "31.self_attn.v_proj",
169
+ "language_model.layers.12.self_attn.k_proj",
170
+ "43.self_attn.k_proj",
171
+ "31.self_attn.q_proj",
172
+ "42.self_attn.k_proj",
173
+ "36.self_attn.q_proj",
174
+ "language_model.layers.26.self_attn.v_proj",
175
+ "language_model.layers.13.self_attn.k_proj",
176
+ "40.self_attn.k_proj",
177
+ "language_model.layers.2.self_attn.k_proj",
178
+ "47.self_attn.q_proj",
179
+ "language_model.layers.16.self_attn.k_proj",
180
+ "31.self_attn.k_proj",
181
+ "30.self_attn.k_proj"
182
+ ],
183
+ "target_parameters": null,
184
+ "task_type": "CAUSAL_LM",
185
+ "trainable_token_indices": null,
186
+ "use_bdlora": null,
187
+ "use_dora": false,
188
+ "use_qalora": false,
189
+ "use_rslora": false,
190
+ "velora_config": null
191
+ }
rl_grpo/control_4x_thinking/checkpoint-16/adapter_model.safetensors ADDED
@@ -0,0 +1,3 @@
 
 
 
 
1
+ version https://git-lfs.github.com/spec/v1
2
+ oid sha256:a5c3974bb67f85942fdee8d8223ee4155d8bfe2295c4f80ab152e20392bb6413
3
+ size 523864568
rl_grpo/control_4x_thinking/checkpoint-256/adapter_config.json ADDED
@@ -0,0 +1,191 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ {
2
+ "alora_invocation_tokens": null,
3
+ "alpha_pattern": {},
4
+ "arrow_config": null,
5
+ "auto_mapping": null,
6
+ "base_model_name_or_path": "/workspace/rl3t_c/parent",
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": 64,
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": 32,
31
+ "rank_pattern": {},
32
+ "revision": null,
33
+ "target_modules": [
34
+ "32.self_attn.v_proj",
35
+ "language_model.layers.12.self_attn.q_proj",
36
+ "language_model.layers.16.self_attn.q_proj",
37
+ "language_model.layers.10.self_attn.q_proj",
38
+ "language_model.layers.17.self_attn.k_proj",
39
+ "language_model.layers.10.self_attn.v_proj",
40
+ "34.self_attn.k_proj",
41
+ "45.self_attn.k_proj",
42
+ "language_model.layers.17.self_attn.v_proj",
43
+ "44.self_attn.k_proj",
44
+ "46.self_attn.v_proj",
45
+ "39.self_attn.k_proj",
46
+ "41.self_attn.v_proj",
47
+ "30.self_attn.v_proj",
48
+ "language_model.layers.17.self_attn.q_proj",
49
+ "33.self_attn.q_proj",
50
+ "language_model.layers.6.self_attn.q_proj",
51
+ "language_model.layers.23.self_attn.v_proj",
52
+ "language_model.layers.7.self_attn.k_proj",
53
+ "language_model.layers.22.self_attn.v_proj",
54
+ "language_model.layers.4.self_attn.v_proj",
55
+ "language_model.layers.24.self_attn.k_proj",
56
+ "language_model.layers.2.self_attn.v_proj",
57
+ "38.self_attn.v_proj",
58
+ "language_model.layers.11.self_attn.k_proj",
59
+ "language_model.layers.25.self_attn.v_proj",
60
+ "language_model.layers.1.self_attn.k_proj",
61
+ "27.self_attn.k_proj",
62
+ "27.self_attn.q_proj",
63
+ "language_model.layers.19.self_attn.k_proj",
64
+ "39.self_attn.v_proj",
65
+ "28.self_attn.q_proj",
66
+ "32.self_attn.q_proj",
67
+ "46.self_attn.q_proj",
68
+ "language_model.layers.15.self_attn.q_proj",
69
+ "o_proj",
70
+ "language_model.layers.26.self_attn.k_proj",
71
+ "language_model.layers.21.self_attn.v_proj",
72
+ "language_model.layers.23.self_attn.k_proj",
73
+ "43.self_attn.q_proj",
74
+ "language_model.layers.23.self_attn.q_proj",
75
+ "40.self_attn.v_proj",
76
+ "28.self_attn.v_proj",
77
+ "language_model.layers.9.self_attn.v_proj",
78
+ "47.self_attn.k_proj",
79
+ "37.self_attn.q_proj",
80
+ "language_model.layers.8.self_attn.k_proj",
81
+ "language_model.layers.0.self_attn.k_proj",
82
+ "36.self_attn.v_proj",
83
+ "language_model.layers.0.self_attn.q_proj",
84
+ "language_model.layers.18.self_attn.v_proj",
85
+ "language_model.layers.5.self_attn.q_proj",
86
+ "language_model.layers.18.self_attn.q_proj",
87
+ "38.self_attn.q_proj",
88
+ "35.self_attn.q_proj",
89
+ "language_model.layers.15.self_attn.v_proj",
90
+ "47.self_attn.v_proj",
91
+ "language_model.layers.0.self_attn.v_proj",
92
+ "language_model.layers.20.self_attn.k_proj",
93
+ "language_model.layers.7.self_attn.q_proj",
94
+ "45.self_attn.q_proj",
95
+ "language_model.layers.2.self_attn.q_proj",
96
+ "29.self_attn.k_proj",
97
+ "33.self_attn.v_proj",
98
+ "language_model.layers.22.self_attn.k_proj",
99
+ "language_model.layers.10.self_attn.k_proj",
100
+ "44.self_attn.q_proj",
101
+ "42.self_attn.v_proj",
102
+ "language_model.layers.20.self_attn.v_proj",
103
+ "language_model.layers.5.self_attn.v_proj",
104
+ "33.self_attn.k_proj",
105
+ "language_model.layers.20.self_attn.q_proj",
106
+ "language_model.layers.1.self_attn.q_proj",
107
+ "language_model.layers.11.self_attn.v_proj",
108
+ "34.self_attn.q_proj",
109
+ "language_model.layers.8.self_attn.q_proj",
110
+ "language_model.layers.4.self_attn.q_proj",
111
+ "up_proj",
112
+ "language_model.layers.9.self_attn.k_proj",
113
+ "language_model.layers.6.self_attn.v_proj",
114
+ "30.self_attn.q_proj",
115
+ "language_model.layers.3.self_attn.q_proj",
116
+ "language_model.layers.26.self_attn.q_proj",
117
+ "27.self_attn.v_proj",
118
+ "language_model.layers.21.self_attn.k_proj",
119
+ "language_model.layers.12.self_attn.v_proj",
120
+ "39.self_attn.q_proj",
121
+ "language_model.layers.19.self_attn.v_proj",
122
+ "language_model.layers.5.self_attn.k_proj",
123
+ "29.self_attn.q_proj",
124
+ "43.self_attn.v_proj",
125
+ "language_model.layers.9.self_attn.q_proj",
126
+ "language_model.layers.19.self_attn.q_proj",
127
+ "35.self_attn.v_proj",
128
+ "language_model.layers.21.self_attn.q_proj",
129
+ "language_model.layers.4.self_attn.k_proj",
130
+ "32.self_attn.k_proj",
131
+ "language_model.layers.13.self_attn.v_proj",
132
+ "37.self_attn.v_proj",
133
+ "38.self_attn.k_proj",
134
+ "language_model.layers.8.self_attn.v_proj",
135
+ "34.self_attn.v_proj",
136
+ "29.self_attn.v_proj",
137
+ "down_proj",
138
+ "language_model.layers.14.self_attn.k_proj",
139
+ "46.self_attn.k_proj",
140
+ "language_model.layers.3.self_attn.k_proj",
141
+ "37.self_attn.k_proj",
142
+ "41.self_attn.q_proj",
143
+ "42.self_attn.q_proj",
144
+ "language_model.layers.24.self_attn.q_proj",
145
+ "language_model.layers.14.self_attn.v_proj",
146
+ "language_model.layers.6.self_attn.k_proj",
147
+ "language_model.layers.25.self_attn.k_proj",
148
+ "language_model.layers.24.self_attn.v_proj",
149
+ "35.self_attn.k_proj",
150
+ "language_model.layers.22.self_attn.q_proj",
151
+ "41.self_attn.k_proj",
152
+ "40.self_attn.q_proj",
153
+ "45.self_attn.v_proj",
154
+ "language_model.layers.16.self_attn.v_proj",
155
+ "language_model.layers.18.self_attn.k_proj",
156
+ "28.self_attn.k_proj",
157
+ "language_model.layers.15.self_attn.k_proj",
158
+ "language_model.layers.14.self_attn.q_proj",
159
+ "36.self_attn.k_proj",
160
+ "language_model.layers.7.self_attn.v_proj",
161
+ "44.self_attn.v_proj",
162
+ "gate_proj",
163
+ "language_model.layers.13.self_attn.q_proj",
164
+ "language_model.layers.25.self_attn.q_proj",
165
+ "language_model.layers.3.self_attn.v_proj",
166
+ "language_model.layers.1.self_attn.v_proj",
167
+ "language_model.layers.11.self_attn.q_proj",
168
+ "31.self_attn.v_proj",
169
+ "language_model.layers.12.self_attn.k_proj",
170
+ "43.self_attn.k_proj",
171
+ "31.self_attn.q_proj",
172
+ "42.self_attn.k_proj",
173
+ "36.self_attn.q_proj",
174
+ "language_model.layers.26.self_attn.v_proj",
175
+ "language_model.layers.13.self_attn.k_proj",
176
+ "40.self_attn.k_proj",
177
+ "language_model.layers.2.self_attn.k_proj",
178
+ "47.self_attn.q_proj",
179
+ "language_model.layers.16.self_attn.k_proj",
180
+ "31.self_attn.k_proj",
181
+ "30.self_attn.k_proj"
182
+ ],
183
+ "target_parameters": null,
184
+ "task_type": "CAUSAL_LM",
185
+ "trainable_token_indices": null,
186
+ "use_bdlora": null,
187
+ "use_dora": false,
188
+ "use_qalora": false,
189
+ "use_rslora": false,
190
+ "velora_config": null
191
+ }
rl_grpo/control_4x_thinking/checkpoint-256/adapter_model.safetensors ADDED
@@ -0,0 +1,3 @@
 
 
 
 
1
+ version https://git-lfs.github.com/spec/v1
2
+ oid sha256:f21eb1919c3fac967114ac264ae1922874a2b3f16961f688ba75f88ee7c712f3
3
+ size 523864568
rl_grpo/control_4x_thinking/checkpoint-256/trainer_state.json ADDED
@@ -0,0 +1,834 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ {
2
+ "best_global_step": null,
3
+ "best_metric": null,
4
+ "best_model_checkpoint": null,
5
+ "epoch": 0.125,
6
+ "eval_steps": 500,
7
+ "global_step": 256,
8
+ "is_hyper_param_search": false,
9
+ "is_local_process_zero": true,
10
+ "is_world_process_zero": true,
11
+ "log_history": [
12
+ {
13
+ "clip_ratio/high_max": 0.0,
14
+ "clip_ratio/high_mean": 0.0,
15
+ "clip_ratio/low_mean": 0.0,
16
+ "clip_ratio/low_min": 0.0,
17
+ "clip_ratio/region_mean": 0.0,
18
+ "completions/clipped_ratio": 0.1,
19
+ "completions/max_length": 1009.6,
20
+ "completions/max_terminated_length": 972.1,
21
+ "completions/mean_length": 520.771875,
22
+ "completions/mean_terminated_length": 466.2046203613281,
23
+ "completions/min_length": 57.7,
24
+ "completions/min_terminated_length": 57.7,
25
+ "entropy": 0.3628267337568104,
26
+ "epoch": 0.0048828125,
27
+ "frac_reward_zero_std": 0.05,
28
+ "grad_norm": 0.12593092024326324,
29
+ "learning_rate": 9.648437500000002e-06,
30
+ "loss": -0.042343777418136594,
31
+ "num_tokens": 385799.0,
32
+ "reward": 0.34375,
33
+ "reward_std": 0.43682001531124115,
34
+ "rewards/reward_func/mean": 0.34375,
35
+ "rewards/reward_func/std": 0.43682001531124115,
36
+ "sampling/importance_sampling_ratio/max": 2.5302028894424438,
37
+ "sampling/importance_sampling_ratio/mean": 0.7979430854320526,
38
+ "sampling/importance_sampling_ratio/min": 0.007613348960876465,
39
+ "sampling/sampling_logp_difference/max": 1.3086066961288452,
40
+ "sampling/sampling_logp_difference/mean": 0.014447255898267031,
41
+ "step": 10,
42
+ "step_time": 42.25655511394143
43
+ },
44
+ {
45
+ "clip_ratio/high_max": 0.0,
46
+ "clip_ratio/high_mean": 0.0,
47
+ "clip_ratio/low_mean": 0.0,
48
+ "clip_ratio/low_min": 0.0,
49
+ "clip_ratio/region_mean": 0.0,
50
+ "completions/clipped_ratio": 0.16875,
51
+ "completions/max_length": 1024.0,
52
+ "completions/max_terminated_length": 924.4,
53
+ "completions/mean_length": 583.925,
54
+ "completions/mean_terminated_length": 496.76673278808596,
55
+ "completions/min_length": 127.7,
56
+ "completions/min_terminated_length": 127.7,
57
+ "entropy": 0.3530748830642551,
58
+ "epoch": 0.009765625,
59
+ "frac_reward_zero_std": 0.0,
60
+ "grad_norm": 0.1087241917848587,
61
+ "learning_rate": 9.2578125e-06,
62
+ "loss": -0.042276978492736816,
63
+ "num_tokens": 805343.0,
64
+ "reward": 0.3640625,
65
+ "reward_std": 0.4532739222049713,
66
+ "rewards/reward_func/mean": 0.3640625,
67
+ "rewards/reward_func/std": 0.4532739222049713,
68
+ "sampling/importance_sampling_ratio/max": 2.2778553128242494,
69
+ "sampling/importance_sampling_ratio/mean": 0.7712720692157745,
70
+ "sampling/importance_sampling_ratio/min": 0.011770828068256379,
71
+ "sampling/sampling_logp_difference/max": 1.316089677810669,
72
+ "sampling/sampling_logp_difference/mean": 0.014172282814979554,
73
+ "step": 20,
74
+ "step_time": 43.61808651871979
75
+ },
76
+ {
77
+ "clip_ratio/high_max": 0.0,
78
+ "clip_ratio/high_mean": 0.0,
79
+ "clip_ratio/low_mean": 0.0,
80
+ "clip_ratio/low_min": 0.0,
81
+ "clip_ratio/region_mean": 0.0,
82
+ "completions/clipped_ratio": 0.109375,
83
+ "completions/max_length": 1024.0,
84
+ "completions/max_terminated_length": 969.3,
85
+ "completions/mean_length": 575.753125,
86
+ "completions/mean_terminated_length": 521.5266296386719,
87
+ "completions/min_length": 106.6,
88
+ "completions/min_terminated_length": 106.6,
89
+ "entropy": 0.33662762232124804,
90
+ "epoch": 0.0146484375,
91
+ "frac_reward_zero_std": 0.0,
92
+ "grad_norm": 0.07704488933086395,
93
+ "learning_rate": 8.867187500000001e-06,
94
+ "loss": -0.02111103981733322,
95
+ "num_tokens": 1213160.0,
96
+ "reward": 0.48125,
97
+ "reward_std": 0.45511032938957213,
98
+ "rewards/reward_func/mean": 0.48125,
99
+ "rewards/reward_func/std": 0.45511032938957213,
100
+ "sampling/importance_sampling_ratio/max": 2.542730951309204,
101
+ "sampling/importance_sampling_ratio/mean": 0.7513267576694489,
102
+ "sampling/importance_sampling_ratio/min": 0.018153008818626405,
103
+ "sampling/sampling_logp_difference/max": 1.3831356525421143,
104
+ "sampling/sampling_logp_difference/mean": 0.01378925023600459,
105
+ "step": 30,
106
+ "step_time": 43.66685572825372
107
+ },
108
+ {
109
+ "clip_ratio/high_max": 0.0,
110
+ "clip_ratio/high_mean": 0.0,
111
+ "clip_ratio/low_mean": 0.0,
112
+ "clip_ratio/low_min": 0.0,
113
+ "clip_ratio/region_mean": 0.0,
114
+ "completions/clipped_ratio": 0.09375,
115
+ "completions/max_length": 1024.0,
116
+ "completions/max_terminated_length": 968.9,
117
+ "completions/mean_length": 557.3125,
118
+ "completions/mean_terminated_length": 509.65965576171874,
119
+ "completions/min_length": 140.2,
120
+ "completions/min_terminated_length": 140.2,
121
+ "entropy": 0.3458843339234591,
122
+ "epoch": 0.01953125,
123
+ "frac_reward_zero_std": 0.0,
124
+ "grad_norm": 0.10220721364021301,
125
+ "learning_rate": 8.4765625e-06,
126
+ "loss": -0.018425728380680084,
127
+ "num_tokens": 1611012.0,
128
+ "reward": 0.528125,
129
+ "reward_std": 0.46025884449481963,
130
+ "rewards/reward_func/mean": 0.528125,
131
+ "rewards/reward_func/std": 0.46025884449481963,
132
+ "sampling/importance_sampling_ratio/max": 2.6125513911247253,
133
+ "sampling/importance_sampling_ratio/mean": 0.8166729927062988,
134
+ "sampling/importance_sampling_ratio/min": 0.0143235232681036,
135
+ "sampling/sampling_logp_difference/max": 2.1702849149703978,
136
+ "sampling/sampling_logp_difference/mean": 0.014072713069617748,
137
+ "step": 40,
138
+ "step_time": 43.12486366406083
139
+ },
140
+ {
141
+ "clip_ratio/high_max": 0.0,
142
+ "clip_ratio/high_mean": 0.0,
143
+ "clip_ratio/low_mean": 0.0,
144
+ "clip_ratio/low_min": 0.0,
145
+ "clip_ratio/region_mean": 0.0,
146
+ "completions/clipped_ratio": 0.103125,
147
+ "completions/max_length": 1024.0,
148
+ "completions/max_terminated_length": 951.0,
149
+ "completions/mean_length": 542.6875,
150
+ "completions/mean_terminated_length": 489.6971405029297,
151
+ "completions/min_length": 86.0,
152
+ "completions/min_terminated_length": 86.0,
153
+ "entropy": 0.31520139779895545,
154
+ "epoch": 0.0244140625,
155
+ "frac_reward_zero_std": 0.0,
156
+ "grad_norm": 0.11432246118783951,
157
+ "learning_rate": 8.085937500000001e-06,
158
+ "loss": -0.07148605585098267,
159
+ "num_tokens": 2004872.0,
160
+ "reward": 0.5109375,
161
+ "reward_std": 0.45015537440776826,
162
+ "rewards/reward_func/mean": 0.5109375,
163
+ "rewards/reward_func/std": 0.45015537440776826,
164
+ "sampling/importance_sampling_ratio/max": 2.4359309673309326,
165
+ "sampling/importance_sampling_ratio/mean": 0.7805306077003479,
166
+ "sampling/importance_sampling_ratio/min": 0.019794788025319577,
167
+ "sampling/sampling_logp_difference/max": 1.8331442713737487,
168
+ "sampling/sampling_logp_difference/mean": 0.013347626384347677,
169
+ "step": 50,
170
+ "step_time": 42.423648685589434
171
+ },
172
+ {
173
+ "clip_ratio/high_max": 0.0,
174
+ "clip_ratio/high_mean": 0.0,
175
+ "clip_ratio/low_mean": 0.0,
176
+ "clip_ratio/low_min": 0.0,
177
+ "clip_ratio/region_mean": 0.0,
178
+ "completions/clipped_ratio": 0.121875,
179
+ "completions/max_length": 1024.0,
180
+ "completions/max_terminated_length": 983.2,
181
+ "completions/mean_length": 607.19375,
182
+ "completions/mean_terminated_length": 551.6534881591797,
183
+ "completions/min_length": 153.5,
184
+ "completions/min_terminated_length": 153.5,
185
+ "entropy": 0.30740448273718357,
186
+ "epoch": 0.029296875,
187
+ "frac_reward_zero_std": 0.025,
188
+ "grad_norm": 0.2590615153312683,
189
+ "learning_rate": 7.6953125e-06,
190
+ "loss": -0.10033506155014038,
191
+ "num_tokens": 2421926.0,
192
+ "reward": 0.5078125,
193
+ "reward_std": 0.4656452387571335,
194
+ "rewards/reward_func/mean": 0.5078125,
195
+ "rewards/reward_func/std": 0.4656452387571335,
196
+ "sampling/importance_sampling_ratio/max": 2.57798718214035,
197
+ "sampling/importance_sampling_ratio/mean": 0.8026305377483368,
198
+ "sampling/importance_sampling_ratio/min": 0.009167940169572831,
199
+ "sampling/sampling_logp_difference/max": 1.2876617193222046,
200
+ "sampling/sampling_logp_difference/mean": 0.013483389094471932,
201
+ "step": 60,
202
+ "step_time": 44.01057116091251
203
+ },
204
+ {
205
+ "clip_ratio/high_max": 0.0,
206
+ "clip_ratio/high_mean": 0.0,
207
+ "clip_ratio/low_mean": 0.0,
208
+ "clip_ratio/low_min": 0.0,
209
+ "clip_ratio/region_mean": 0.0,
210
+ "completions/clipped_ratio": 0.13125,
211
+ "completions/max_length": 1024.0,
212
+ "completions/max_terminated_length": 969.7,
213
+ "completions/mean_length": 593.434375,
214
+ "completions/mean_terminated_length": 530.915185546875,
215
+ "completions/min_length": 149.7,
216
+ "completions/min_terminated_length": 149.7,
217
+ "entropy": 0.29846702618524434,
218
+ "epoch": 0.0341796875,
219
+ "frac_reward_zero_std": 0.025,
220
+ "grad_norm": 0.08271189033985138,
221
+ "learning_rate": 7.3046875000000005e-06,
222
+ "loss": -0.05022428035736084,
223
+ "num_tokens": 2833257.0,
224
+ "reward": 0.540625,
225
+ "reward_std": 0.4778682917356491,
226
+ "rewards/reward_func/mean": 0.540625,
227
+ "rewards/reward_func/std": 0.4778682917356491,
228
+ "sampling/importance_sampling_ratio/max": 2.2693339467048643,
229
+ "sampling/importance_sampling_ratio/mean": 0.7266122996807098,
230
+ "sampling/importance_sampling_ratio/min": 0.0,
231
+ "sampling/sampling_logp_difference/max": 1.4138091802597046,
232
+ "sampling/sampling_logp_difference/mean": 0.013372139912098647,
233
+ "step": 70,
234
+ "step_time": 43.55437629893422
235
+ },
236
+ {
237
+ "clip_ratio/high_max": 0.0,
238
+ "clip_ratio/high_mean": 0.0,
239
+ "clip_ratio/low_mean": 0.0,
240
+ "clip_ratio/low_min": 0.0,
241
+ "clip_ratio/region_mean": 0.0,
242
+ "completions/clipped_ratio": 0.159375,
243
+ "completions/max_length": 1011.5,
244
+ "completions/max_terminated_length": 931.9,
245
+ "completions/mean_length": 616.15625,
246
+ "completions/mean_terminated_length": 541.745751953125,
247
+ "completions/min_length": 150.8,
248
+ "completions/min_terminated_length": 150.8,
249
+ "entropy": 0.29485611207783224,
250
+ "epoch": 0.0390625,
251
+ "frac_reward_zero_std": 0.1,
252
+ "grad_norm": 0.07818613201379776,
253
+ "learning_rate": 6.9140625e-06,
254
+ "loss": -0.06272406578063965,
255
+ "num_tokens": 3241971.0,
256
+ "reward": 0.50625,
257
+ "reward_std": 0.4646908640861511,
258
+ "rewards/reward_func/mean": 0.50625,
259
+ "rewards/reward_func/std": 0.4646908640861511,
260
+ "sampling/importance_sampling_ratio/max": 2.409692180156708,
261
+ "sampling/importance_sampling_ratio/mean": 0.7578879594802856,
262
+ "sampling/importance_sampling_ratio/min": 0.02440670910291374,
263
+ "sampling/sampling_logp_difference/max": 1.4652909636497498,
264
+ "sampling/sampling_logp_difference/mean": 0.01341348122805357,
265
+ "step": 80,
266
+ "step_time": 43.06877421811223
267
+ },
268
+ {
269
+ "clip_ratio/high_max": 0.0,
270
+ "clip_ratio/high_mean": 0.0,
271
+ "clip_ratio/low_mean": 0.0,
272
+ "clip_ratio/low_min": 0.0,
273
+ "clip_ratio/region_mean": 0.0,
274
+ "completions/clipped_ratio": 0.15625,
275
+ "completions/max_length": 1024.0,
276
+ "completions/max_terminated_length": 931.9,
277
+ "completions/mean_length": 642.61875,
278
+ "completions/mean_terminated_length": 573.1607971191406,
279
+ "completions/min_length": 193.3,
280
+ "completions/min_terminated_length": 193.3,
281
+ "entropy": 0.28848616653122006,
282
+ "epoch": 0.0439453125,
283
+ "frac_reward_zero_std": 0.0,
284
+ "grad_norm": 0.09436356276273727,
285
+ "learning_rate": 6.5234375e-06,
286
+ "loss": -0.08488062024116516,
287
+ "num_tokens": 3680401.0,
288
+ "reward": 0.5578125,
289
+ "reward_std": 0.46121214926242826,
290
+ "rewards/reward_func/mean": 0.5578125,
291
+ "rewards/reward_func/std": 0.46121214926242826,
292
+ "sampling/importance_sampling_ratio/max": 2.2048617243766784,
293
+ "sampling/importance_sampling_ratio/mean": 0.7439890682697297,
294
+ "sampling/importance_sampling_ratio/min": 0.012665604008361697,
295
+ "sampling/sampling_logp_difference/max": 1.3647143244743347,
296
+ "sampling/sampling_logp_difference/mean": 0.013502738904207945,
297
+ "step": 90,
298
+ "step_time": 45.26657482124865
299
+ },
300
+ {
301
+ "clip_ratio/high_max": 0.0,
302
+ "clip_ratio/high_mean": 0.0,
303
+ "clip_ratio/low_mean": 0.0,
304
+ "clip_ratio/low_min": 0.0,
305
+ "clip_ratio/region_mean": 0.0,
306
+ "completions/clipped_ratio": 0.1625,
307
+ "completions/max_length": 1024.0,
308
+ "completions/max_terminated_length": 915.3,
309
+ "completions/mean_length": 619.84375,
310
+ "completions/mean_terminated_length": 545.9655944824219,
311
+ "completions/min_length": 166.1,
312
+ "completions/min_terminated_length": 166.1,
313
+ "entropy": 0.2740688876248896,
314
+ "epoch": 0.048828125,
315
+ "frac_reward_zero_std": 0.0,
316
+ "grad_norm": 0.11040587723255157,
317
+ "learning_rate": 6.132812500000001e-06,
318
+ "loss": -0.07392127513885498,
319
+ "num_tokens": 4103599.0,
320
+ "reward": 0.55,
321
+ "reward_std": 0.44316819310188293,
322
+ "rewards/reward_func/mean": 0.55,
323
+ "rewards/reward_func/std": 0.44316819310188293,
324
+ "sampling/importance_sampling_ratio/max": 2.4185766696929933,
325
+ "sampling/importance_sampling_ratio/mean": 0.7439922332763672,
326
+ "sampling/importance_sampling_ratio/min": 0.007297312282025814,
327
+ "sampling/sampling_logp_difference/max": 1.6027182817459107,
328
+ "sampling/sampling_logp_difference/mean": 0.013244564458727837,
329
+ "step": 100,
330
+ "step_time": 44.048098570108415
331
+ },
332
+ {
333
+ "clip_ratio/high_max": 0.0,
334
+ "clip_ratio/high_mean": 0.0,
335
+ "clip_ratio/low_mean": 0.0,
336
+ "clip_ratio/low_min": 0.0,
337
+ "clip_ratio/region_mean": 0.0,
338
+ "completions/clipped_ratio": 0.14375,
339
+ "completions/max_length": 1024.0,
340
+ "completions/max_terminated_length": 979.9,
341
+ "completions/mean_length": 631.13125,
342
+ "completions/mean_terminated_length": 567.6347412109375,
343
+ "completions/min_length": 173.8,
344
+ "completions/min_terminated_length": 173.8,
345
+ "entropy": 0.26472241841256616,
346
+ "epoch": 0.0537109375,
347
+ "frac_reward_zero_std": 0.075,
348
+ "grad_norm": 0.11453425884246826,
349
+ "learning_rate": 5.742187500000001e-06,
350
+ "loss": -0.054640966653823855,
351
+ "num_tokens": 4529313.0,
352
+ "reward": 0.603125,
353
+ "reward_std": 0.4577375501394272,
354
+ "rewards/reward_func/mean": 0.603125,
355
+ "rewards/reward_func/std": 0.4577375501394272,
356
+ "sampling/importance_sampling_ratio/max": 2.4363184809684753,
357
+ "sampling/importance_sampling_ratio/mean": 0.7668845951557159,
358
+ "sampling/importance_sampling_ratio/min": 0.006693846452981234,
359
+ "sampling/sampling_logp_difference/max": 1.365015435218811,
360
+ "sampling/sampling_logp_difference/mean": 0.013046279922127724,
361
+ "step": 110,
362
+ "step_time": 44.43461542390287
363
+ },
364
+ {
365
+ "clip_ratio/high_max": 0.0,
366
+ "clip_ratio/high_mean": 0.0,
367
+ "clip_ratio/low_mean": 0.0,
368
+ "clip_ratio/low_min": 0.0,
369
+ "clip_ratio/region_mean": 0.0,
370
+ "completions/clipped_ratio": 0.103125,
371
+ "completions/max_length": 1022.7,
372
+ "completions/max_terminated_length": 957.3,
373
+ "completions/mean_length": 598.053125,
374
+ "completions/mean_terminated_length": 550.7712188720703,
375
+ "completions/min_length": 178.6,
376
+ "completions/min_terminated_length": 178.6,
377
+ "entropy": 0.2688146966509521,
378
+ "epoch": 0.05859375,
379
+ "frac_reward_zero_std": 0.1,
380
+ "grad_norm": 0.14234773814678192,
381
+ "learning_rate": 5.3515625000000005e-06,
382
+ "loss": -0.023319168388843535,
383
+ "num_tokens": 4918050.0,
384
+ "reward": 0.6328125,
385
+ "reward_std": 0.4575097620487213,
386
+ "rewards/reward_func/mean": 0.6328125,
387
+ "rewards/reward_func/std": 0.4575097620487213,
388
+ "sampling/importance_sampling_ratio/max": 2.481744611263275,
389
+ "sampling/importance_sampling_ratio/mean": 0.7746831178665161,
390
+ "sampling/importance_sampling_ratio/min": 0.017820985801517964,
391
+ "sampling/sampling_logp_difference/max": 1.6181822896003724,
392
+ "sampling/sampling_logp_difference/mean": 0.012676793243736028,
393
+ "step": 120,
394
+ "step_time": 41.60442548058927
395
+ },
396
+ {
397
+ "clip_ratio/high_max": 0.0,
398
+ "clip_ratio/high_mean": 0.0,
399
+ "clip_ratio/low_mean": 0.0,
400
+ "clip_ratio/low_min": 0.0,
401
+ "clip_ratio/region_mean": 0.0,
402
+ "completions/clipped_ratio": 0.09375,
403
+ "completions/max_length": 997.7,
404
+ "completions/max_terminated_length": 942.0,
405
+ "completions/mean_length": 587.45625,
406
+ "completions/mean_terminated_length": 544.7189392089844,
407
+ "completions/min_length": 163.8,
408
+ "completions/min_terminated_length": 163.8,
409
+ "entropy": 0.2607317406218499,
410
+ "epoch": 0.0634765625,
411
+ "frac_reward_zero_std": 0.025,
412
+ "grad_norm": 0.06450571119785309,
413
+ "learning_rate": 4.9609375e-06,
414
+ "loss": -0.07790356278419494,
415
+ "num_tokens": 5314972.0,
416
+ "reward": 0.6296875,
417
+ "reward_std": 0.45218701660633087,
418
+ "rewards/reward_func/mean": 0.6296875,
419
+ "rewards/reward_func/std": 0.45218701660633087,
420
+ "sampling/importance_sampling_ratio/max": 2.196352744102478,
421
+ "sampling/importance_sampling_ratio/mean": 0.7221019685268402,
422
+ "sampling/importance_sampling_ratio/min": 0.007384332269430161,
423
+ "sampling/sampling_logp_difference/max": 1.1781771659851075,
424
+ "sampling/sampling_logp_difference/mean": 0.012488813418895007,
425
+ "step": 130,
426
+ "step_time": 42.92850795276463
427
+ },
428
+ {
429
+ "clip_ratio/high_max": 0.0,
430
+ "clip_ratio/high_mean": 0.0,
431
+ "clip_ratio/low_mean": 0.0,
432
+ "clip_ratio/low_min": 0.0,
433
+ "clip_ratio/region_mean": 0.0,
434
+ "completions/clipped_ratio": 0.13125,
435
+ "completions/max_length": 1024.0,
436
+ "completions/max_terminated_length": 945.0,
437
+ "completions/mean_length": 630.29375,
438
+ "completions/mean_terminated_length": 574.0001098632813,
439
+ "completions/min_length": 192.1,
440
+ "completions/min_terminated_length": 192.1,
441
+ "entropy": 0.2335894290357828,
442
+ "epoch": 0.068359375,
443
+ "frac_reward_zero_std": 0.075,
444
+ "grad_norm": 0.10005701333284378,
445
+ "learning_rate": 4.5703125e-06,
446
+ "loss": -0.05933012962341309,
447
+ "num_tokens": 5734898.0,
448
+ "reward": 0.640625,
449
+ "reward_std": 0.4408276855945587,
450
+ "rewards/reward_func/mean": 0.640625,
451
+ "rewards/reward_func/std": 0.4408276855945587,
452
+ "sampling/importance_sampling_ratio/max": 2.4288483023643495,
453
+ "sampling/importance_sampling_ratio/mean": 0.7791410148143768,
454
+ "sampling/importance_sampling_ratio/min": 0.0017797250300645828,
455
+ "sampling/sampling_logp_difference/max": 1.4576029777526855,
456
+ "sampling/sampling_logp_difference/mean": 0.011468723230063916,
457
+ "step": 140,
458
+ "step_time": 43.29499200806022
459
+ },
460
+ {
461
+ "clip_ratio/high_max": 0.0,
462
+ "clip_ratio/high_mean": 0.0,
463
+ "clip_ratio/low_mean": 0.0,
464
+ "clip_ratio/low_min": 0.0,
465
+ "clip_ratio/region_mean": 0.0,
466
+ "completions/clipped_ratio": 0.128125,
467
+ "completions/max_length": 1013.2,
468
+ "completions/max_terminated_length": 928.4,
469
+ "completions/mean_length": 586.29375,
470
+ "completions/mean_terminated_length": 524.9119079589843,
471
+ "completions/min_length": 156.5,
472
+ "completions/min_terminated_length": 156.5,
473
+ "entropy": 0.2614118416327983,
474
+ "epoch": 0.0732421875,
475
+ "frac_reward_zero_std": 0.075,
476
+ "grad_norm": 0.46334078907966614,
477
+ "learning_rate": 4.1796875e-06,
478
+ "loss": -0.06795669198036194,
479
+ "num_tokens": 6144904.0,
480
+ "reward": 0.615625,
481
+ "reward_std": 0.45858532190322876,
482
+ "rewards/reward_func/mean": 0.615625,
483
+ "rewards/reward_func/std": 0.45858532190322876,
484
+ "sampling/importance_sampling_ratio/max": 2.6508662581443785,
485
+ "sampling/importance_sampling_ratio/mean": 0.7971533536911011,
486
+ "sampling/importance_sampling_ratio/min": 0.025614945590496062,
487
+ "sampling/sampling_logp_difference/max": 1.1767661452293396,
488
+ "sampling/sampling_logp_difference/mean": 0.012624044995754957,
489
+ "step": 150,
490
+ "step_time": 43.63800957202911
491
+ },
492
+ {
493
+ "clip_ratio/high_max": 0.0,
494
+ "clip_ratio/high_mean": 0.0,
495
+ "clip_ratio/low_mean": 0.0,
496
+ "clip_ratio/low_min": 0.0,
497
+ "clip_ratio/region_mean": 0.0,
498
+ "completions/clipped_ratio": 0.10625,
499
+ "completions/max_length": 1024.0,
500
+ "completions/max_terminated_length": 966.4,
501
+ "completions/mean_length": 594.43125,
502
+ "completions/mean_terminated_length": 544.5541381835938,
503
+ "completions/min_length": 160.6,
504
+ "completions/min_terminated_length": 160.6,
505
+ "entropy": 0.25198871763423086,
506
+ "epoch": 0.078125,
507
+ "frac_reward_zero_std": 0.125,
508
+ "grad_norm": 0.1025005504488945,
509
+ "learning_rate": 3.7890625e-06,
510
+ "loss": -0.09782418608665466,
511
+ "num_tokens": 6554954.0,
512
+ "reward": 0.596875,
513
+ "reward_std": 0.4632529079914093,
514
+ "rewards/reward_func/mean": 0.596875,
515
+ "rewards/reward_func/std": 0.4632529079914093,
516
+ "sampling/importance_sampling_ratio/max": 2.5903760194778442,
517
+ "sampling/importance_sampling_ratio/mean": 0.774015611410141,
518
+ "sampling/importance_sampling_ratio/min": 0.0015638312324881554,
519
+ "sampling/sampling_logp_difference/max": 1.1049574255943297,
520
+ "sampling/sampling_logp_difference/mean": 0.01211427841335535,
521
+ "step": 160,
522
+ "step_time": 43.625414764881135
523
+ },
524
+ {
525
+ "clip_ratio/high_max": 0.0,
526
+ "clip_ratio/high_mean": 0.0,
527
+ "clip_ratio/low_mean": 0.0,
528
+ "clip_ratio/low_min": 0.0,
529
+ "clip_ratio/region_mean": 0.0,
530
+ "completions/clipped_ratio": 0.11875,
531
+ "completions/max_length": 1024.0,
532
+ "completions/max_terminated_length": 944.2,
533
+ "completions/mean_length": 607.26875,
534
+ "completions/mean_terminated_length": 553.7439361572266,
535
+ "completions/min_length": 181.7,
536
+ "completions/min_terminated_length": 181.7,
537
+ "entropy": 0.2524027226958424,
538
+ "epoch": 0.0830078125,
539
+ "frac_reward_zero_std": 0.025,
540
+ "grad_norm": 0.11777320504188538,
541
+ "learning_rate": 3.3984375000000004e-06,
542
+ "loss": -0.04185573756694794,
543
+ "num_tokens": 6973368.0,
544
+ "reward": 0.6375,
545
+ "reward_std": 0.4353609800338745,
546
+ "rewards/reward_func/mean": 0.6375,
547
+ "rewards/reward_func/std": 0.4353609800338745,
548
+ "sampling/importance_sampling_ratio/max": 2.5190322637557983,
549
+ "sampling/importance_sampling_ratio/mean": 0.7401076853275299,
550
+ "sampling/importance_sampling_ratio/min": 0.028387261927127837,
551
+ "sampling/sampling_logp_difference/max": 0.9780882596969604,
552
+ "sampling/sampling_logp_difference/mean": 0.012250632606446743,
553
+ "step": 170,
554
+ "step_time": 44.03574435822666
555
+ },
556
+ {
557
+ "clip_ratio/high_max": 0.0,
558
+ "clip_ratio/high_mean": 0.0,
559
+ "clip_ratio/low_mean": 0.0,
560
+ "clip_ratio/low_min": 0.0,
561
+ "clip_ratio/region_mean": 0.0,
562
+ "completions/clipped_ratio": 0.13125,
563
+ "completions/max_length": 1015.1,
564
+ "completions/max_terminated_length": 917.5,
565
+ "completions/mean_length": 592.853125,
566
+ "completions/mean_terminated_length": 528.6190032958984,
567
+ "completions/min_length": 150.2,
568
+ "completions/min_terminated_length": 150.2,
569
+ "entropy": 0.2538876271340996,
570
+ "epoch": 0.087890625,
571
+ "frac_reward_zero_std": 0.15,
572
+ "grad_norm": 0.10863858461380005,
573
+ "learning_rate": 3.0078125000000003e-06,
574
+ "loss": -0.07713493108749389,
575
+ "num_tokens": 7384041.0,
576
+ "reward": 0.609375,
577
+ "reward_std": 0.4516817361116409,
578
+ "rewards/reward_func/mean": 0.609375,
579
+ "rewards/reward_func/std": 0.4516817361116409,
580
+ "sampling/importance_sampling_ratio/max": 2.2034048318862913,
581
+ "sampling/importance_sampling_ratio/mean": 0.719656091928482,
582
+ "sampling/importance_sampling_ratio/min": 0.02234973944723606,
583
+ "sampling/sampling_logp_difference/max": 1.0969637393951417,
584
+ "sampling/sampling_logp_difference/mean": 0.012740517035126686,
585
+ "step": 180,
586
+ "step_time": 42.91039258167147
587
+ },
588
+ {
589
+ "clip_ratio/high_max": 0.0,
590
+ "clip_ratio/high_mean": 0.0,
591
+ "clip_ratio/low_mean": 0.0,
592
+ "clip_ratio/low_min": 0.0,
593
+ "clip_ratio/region_mean": 0.0,
594
+ "completions/clipped_ratio": 0.115625,
595
+ "completions/max_length": 1024.0,
596
+ "completions/max_terminated_length": 981.0,
597
+ "completions/mean_length": 600.021875,
598
+ "completions/mean_terminated_length": 545.0324554443359,
599
+ "completions/min_length": 190.4,
600
+ "completions/min_terminated_length": 190.4,
601
+ "entropy": 0.24937609285116197,
602
+ "epoch": 0.0927734375,
603
+ "frac_reward_zero_std": 0.125,
604
+ "grad_norm": 0.8676526546478271,
605
+ "learning_rate": 2.6171875e-06,
606
+ "loss": -0.08467918634414673,
607
+ "num_tokens": 7802680.0,
608
+ "reward": 0.66875,
609
+ "reward_std": 0.4258056640625,
610
+ "rewards/reward_func/mean": 0.66875,
611
+ "rewards/reward_func/std": 0.4258056640625,
612
+ "sampling/importance_sampling_ratio/max": 2.5355358600616453,
613
+ "sampling/importance_sampling_ratio/mean": 0.7611053049564361,
614
+ "sampling/importance_sampling_ratio/min": 0.005045086517930031,
615
+ "sampling/sampling_logp_difference/max": 1.1458203077316285,
616
+ "sampling/sampling_logp_difference/mean": 0.012418949231505393,
617
+ "step": 190,
618
+ "step_time": 44.422478976473215
619
+ },
620
+ {
621
+ "clip_ratio/high_max": 0.0,
622
+ "clip_ratio/high_mean": 0.0,
623
+ "clip_ratio/low_mean": 0.0,
624
+ "clip_ratio/low_min": 0.0,
625
+ "clip_ratio/region_mean": 0.0,
626
+ "completions/clipped_ratio": 0.109375,
627
+ "completions/max_length": 1022.7,
628
+ "completions/max_terminated_length": 951.2,
629
+ "completions/mean_length": 598.309375,
630
+ "completions/mean_terminated_length": 546.960171508789,
631
+ "completions/min_length": 185.3,
632
+ "completions/min_terminated_length": 185.3,
633
+ "entropy": 0.25309558347798883,
634
+ "epoch": 0.09765625,
635
+ "frac_reward_zero_std": 0.075,
636
+ "grad_norm": 0.2170192003250122,
637
+ "learning_rate": 2.2265625000000003e-06,
638
+ "loss": -0.0773630142211914,
639
+ "num_tokens": 8222163.0,
640
+ "reward": 0.6171875,
641
+ "reward_std": 0.4537837773561478,
642
+ "rewards/reward_func/mean": 0.6171875,
643
+ "rewards/reward_func/std": 0.4537837773561478,
644
+ "sampling/importance_sampling_ratio/max": 2.4515735507011414,
645
+ "sampling/importance_sampling_ratio/mean": 0.6715039759874344,
646
+ "sampling/importance_sampling_ratio/min": 0.01123899221420288,
647
+ "sampling/sampling_logp_difference/max": 1.4848322749137879,
648
+ "sampling/sampling_logp_difference/mean": 0.012924220971763134,
649
+ "step": 200,
650
+ "step_time": 44.45845618769526
651
+ },
652
+ {
653
+ "clip_ratio/high_max": 0.0,
654
+ "clip_ratio/high_mean": 0.0,
655
+ "clip_ratio/low_mean": 0.0,
656
+ "clip_ratio/low_min": 0.0,
657
+ "clip_ratio/region_mean": 0.0,
658
+ "completions/clipped_ratio": 0.096875,
659
+ "completions/max_length": 1024.0,
660
+ "completions/max_terminated_length": 917.1,
661
+ "completions/mean_length": 579.11875,
662
+ "completions/mean_terminated_length": 533.4338500976562,
663
+ "completions/min_length": 167.7,
664
+ "completions/min_terminated_length": 167.7,
665
+ "entropy": 0.23500776635482906,
666
+ "epoch": 0.1025390625,
667
+ "frac_reward_zero_std": 0.125,
668
+ "grad_norm": 0.035419777035713196,
669
+ "learning_rate": 1.8359375000000002e-06,
670
+ "loss": -0.05832840204238891,
671
+ "num_tokens": 8627681.0,
672
+ "reward": 0.6609375,
673
+ "reward_std": 0.43311508297920226,
674
+ "rewards/reward_func/mean": 0.6609375,
675
+ "rewards/reward_func/std": 0.43311508297920226,
676
+ "sampling/importance_sampling_ratio/max": 2.3608193039894103,
677
+ "sampling/importance_sampling_ratio/mean": 0.7027318179607391,
678
+ "sampling/importance_sampling_ratio/min": 0.0,
679
+ "sampling/sampling_logp_difference/max": 1.4359980344772338,
680
+ "sampling/sampling_logp_difference/mean": 0.012371512968093157,
681
+ "step": 210,
682
+ "step_time": 44.468596732243896
683
+ },
684
+ {
685
+ "clip_ratio/high_max": 0.0,
686
+ "clip_ratio/high_mean": 0.0,
687
+ "clip_ratio/low_mean": 0.0,
688
+ "clip_ratio/low_min": 0.0,
689
+ "clip_ratio/region_mean": 0.0,
690
+ "completions/clipped_ratio": 0.115625,
691
+ "completions/max_length": 1021.3,
692
+ "completions/max_terminated_length": 968.0,
693
+ "completions/mean_length": 620.678125,
694
+ "completions/mean_terminated_length": 568.5793823242187,
695
+ "completions/min_length": 186.0,
696
+ "completions/min_terminated_length": 186.0,
697
+ "entropy": 0.23499757889658213,
698
+ "epoch": 0.107421875,
699
+ "frac_reward_zero_std": 0.075,
700
+ "grad_norm": 0.05257241055369377,
701
+ "learning_rate": 1.4453125000000002e-06,
702
+ "loss": -0.06840035915374756,
703
+ "num_tokens": 9062370.0,
704
+ "reward": 0.66875,
705
+ "reward_std": 0.4268799781799316,
706
+ "rewards/reward_func/mean": 0.66875,
707
+ "rewards/reward_func/std": 0.4268799781799316,
708
+ "sampling/importance_sampling_ratio/max": 2.577197241783142,
709
+ "sampling/importance_sampling_ratio/mean": 0.7276532471179962,
710
+ "sampling/importance_sampling_ratio/min": 0.0002849083859473467,
711
+ "sampling/sampling_logp_difference/max": 1.7443309903144837,
712
+ "sampling/sampling_logp_difference/mean": 0.012142607476562262,
713
+ "step": 220,
714
+ "step_time": 44.525983604416254
715
+ },
716
+ {
717
+ "clip_ratio/high_max": 0.0,
718
+ "clip_ratio/high_mean": 0.0,
719
+ "clip_ratio/low_mean": 0.0,
720
+ "clip_ratio/low_min": 0.0,
721
+ "clip_ratio/region_mean": 0.0,
722
+ "completions/clipped_ratio": 0.0875,
723
+ "completions/max_length": 1016.8,
724
+ "completions/max_terminated_length": 951.2,
725
+ "completions/mean_length": 555.828125,
726
+ "completions/mean_terminated_length": 512.8386505126953,
727
+ "completions/min_length": 175.5,
728
+ "completions/min_terminated_length": 175.5,
729
+ "entropy": 0.2507997420616448,
730
+ "epoch": 0.1123046875,
731
+ "frac_reward_zero_std": 0.15,
732
+ "grad_norm": 0.16403308510780334,
733
+ "learning_rate": 1.0546875e-06,
734
+ "loss": -0.07354863882064819,
735
+ "num_tokens": 9439971.0,
736
+ "reward": 0.6984375,
737
+ "reward_std": 0.4155709892511368,
738
+ "rewards/reward_func/mean": 0.6984375,
739
+ "rewards/reward_func/std": 0.4155709892511368,
740
+ "sampling/importance_sampling_ratio/max": 2.3330983877182008,
741
+ "sampling/importance_sampling_ratio/mean": 0.7435574173927307,
742
+ "sampling/importance_sampling_ratio/min": 0.011307070404291153,
743
+ "sampling/sampling_logp_difference/max": 1.336429762840271,
744
+ "sampling/sampling_logp_difference/mean": 0.012811038922518492,
745
+ "step": 230,
746
+ "step_time": 41.94405762404203
747
+ },
748
+ {
749
+ "clip_ratio/high_max": 0.0,
750
+ "clip_ratio/high_mean": 0.0,
751
+ "clip_ratio/low_mean": 0.0,
752
+ "clip_ratio/low_min": 0.0,
753
+ "clip_ratio/region_mean": 0.0,
754
+ "completions/clipped_ratio": 0.159375,
755
+ "completions/max_length": 1024.0,
756
+ "completions/max_terminated_length": 961.3,
757
+ "completions/mean_length": 627.509375,
758
+ "completions/mean_terminated_length": 556.1728118896484,
759
+ "completions/min_length": 187.1,
760
+ "completions/min_terminated_length": 187.1,
761
+ "entropy": 0.2292167702689767,
762
+ "epoch": 0.1171875,
763
+ "frac_reward_zero_std": 0.1,
764
+ "grad_norm": 0.07819076627492905,
765
+ "learning_rate": 6.640625e-07,
766
+ "loss": -0.07779039144515991,
767
+ "num_tokens": 9875454.0,
768
+ "reward": 0.640625,
769
+ "reward_std": 0.44508951604366304,
770
+ "rewards/reward_func/mean": 0.640625,
771
+ "rewards/reward_func/std": 0.44508951604366304,
772
+ "sampling/importance_sampling_ratio/max": 2.4810739755630493,
773
+ "sampling/importance_sampling_ratio/mean": 0.7441669166088104,
774
+ "sampling/importance_sampling_ratio/min": 0.016938938945531844,
775
+ "sampling/sampling_logp_difference/max": 1.247516894340515,
776
+ "sampling/sampling_logp_difference/mean": 0.012005124520510436,
777
+ "step": 240,
778
+ "step_time": 44.566947877407074
779
+ },
780
+ {
781
+ "clip_ratio/high_max": 0.0,
782
+ "clip_ratio/high_mean": 0.0,
783
+ "clip_ratio/low_mean": 0.0,
784
+ "clip_ratio/low_min": 0.0,
785
+ "clip_ratio/region_mean": 0.0,
786
+ "completions/clipped_ratio": 0.1125,
787
+ "completions/max_length": 1003.0,
788
+ "completions/max_terminated_length": 958.9,
789
+ "completions/mean_length": 605.275,
790
+ "completions/mean_terminated_length": 551.7752532958984,
791
+ "completions/min_length": 159.0,
792
+ "completions/min_terminated_length": 159.0,
793
+ "entropy": 0.23129161642864346,
794
+ "epoch": 0.1220703125,
795
+ "frac_reward_zero_std": 0.2,
796
+ "grad_norm": 0.09357041120529175,
797
+ "learning_rate": 2.7343750000000003e-07,
798
+ "loss": -0.061303246021270755,
799
+ "num_tokens": 10285878.0,
800
+ "reward": 0.6859375,
801
+ "reward_std": 0.4363765776157379,
802
+ "rewards/reward_func/mean": 0.6859375,
803
+ "rewards/reward_func/std": 0.4363765776157379,
804
+ "sampling/importance_sampling_ratio/max": 2.1400226354599,
805
+ "sampling/importance_sampling_ratio/mean": 0.7143266677856446,
806
+ "sampling/importance_sampling_ratio/min": 0.011234851833432914,
807
+ "sampling/sampling_logp_difference/max": 1.2579308867454528,
808
+ "sampling/sampling_logp_difference/mean": 0.012214426603168248,
809
+ "step": 250,
810
+ "step_time": 43.89955201484263
811
+ }
812
+ ],
813
+ "logging_steps": 10,
814
+ "max_steps": 256,
815
+ "num_input_tokens_seen": 10555015,
816
+ "num_train_epochs": 1,
817
+ "save_steps": 257,
818
+ "stateful_callbacks": {
819
+ "TrainerControl": {
820
+ "args": {
821
+ "should_epoch_stop": false,
822
+ "should_evaluate": false,
823
+ "should_log": false,
824
+ "should_save": true,
825
+ "should_training_stop": true
826
+ },
827
+ "attributes": {}
828
+ }
829
+ },
830
+ "total_flos": 0.0,
831
+ "train_batch_size": 2,
832
+ "trial_name": null,
833
+ "trial_params": null
834
+ }
rl_grpo/control_4x_thinking/checkpoint-32/adapter_config.json ADDED
@@ -0,0 +1,191 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ {
2
+ "alora_invocation_tokens": null,
3
+ "alpha_pattern": {},
4
+ "arrow_config": null,
5
+ "auto_mapping": null,
6
+ "base_model_name_or_path": "/workspace/rl3t_c/parent",
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": 64,
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": 32,
31
+ "rank_pattern": {},
32
+ "revision": null,
33
+ "target_modules": [
34
+ "32.self_attn.v_proj",
35
+ "language_model.layers.12.self_attn.q_proj",
36
+ "language_model.layers.16.self_attn.q_proj",
37
+ "language_model.layers.10.self_attn.q_proj",
38
+ "language_model.layers.17.self_attn.k_proj",
39
+ "language_model.layers.10.self_attn.v_proj",
40
+ "34.self_attn.k_proj",
41
+ "45.self_attn.k_proj",
42
+ "language_model.layers.17.self_attn.v_proj",
43
+ "44.self_attn.k_proj",
44
+ "46.self_attn.v_proj",
45
+ "39.self_attn.k_proj",
46
+ "41.self_attn.v_proj",
47
+ "30.self_attn.v_proj",
48
+ "language_model.layers.17.self_attn.q_proj",
49
+ "33.self_attn.q_proj",
50
+ "language_model.layers.6.self_attn.q_proj",
51
+ "language_model.layers.23.self_attn.v_proj",
52
+ "language_model.layers.7.self_attn.k_proj",
53
+ "language_model.layers.22.self_attn.v_proj",
54
+ "language_model.layers.4.self_attn.v_proj",
55
+ "language_model.layers.24.self_attn.k_proj",
56
+ "language_model.layers.2.self_attn.v_proj",
57
+ "38.self_attn.v_proj",
58
+ "language_model.layers.11.self_attn.k_proj",
59
+ "language_model.layers.25.self_attn.v_proj",
60
+ "language_model.layers.1.self_attn.k_proj",
61
+ "27.self_attn.k_proj",
62
+ "27.self_attn.q_proj",
63
+ "language_model.layers.19.self_attn.k_proj",
64
+ "39.self_attn.v_proj",
65
+ "28.self_attn.q_proj",
66
+ "32.self_attn.q_proj",
67
+ "46.self_attn.q_proj",
68
+ "language_model.layers.15.self_attn.q_proj",
69
+ "o_proj",
70
+ "language_model.layers.26.self_attn.k_proj",
71
+ "language_model.layers.21.self_attn.v_proj",
72
+ "language_model.layers.23.self_attn.k_proj",
73
+ "43.self_attn.q_proj",
74
+ "language_model.layers.23.self_attn.q_proj",
75
+ "40.self_attn.v_proj",
76
+ "28.self_attn.v_proj",
77
+ "language_model.layers.9.self_attn.v_proj",
78
+ "47.self_attn.k_proj",
79
+ "37.self_attn.q_proj",
80
+ "language_model.layers.8.self_attn.k_proj",
81
+ "language_model.layers.0.self_attn.k_proj",
82
+ "36.self_attn.v_proj",
83
+ "language_model.layers.0.self_attn.q_proj",
84
+ "language_model.layers.18.self_attn.v_proj",
85
+ "language_model.layers.5.self_attn.q_proj",
86
+ "language_model.layers.18.self_attn.q_proj",
87
+ "38.self_attn.q_proj",
88
+ "35.self_attn.q_proj",
89
+ "language_model.layers.15.self_attn.v_proj",
90
+ "47.self_attn.v_proj",
91
+ "language_model.layers.0.self_attn.v_proj",
92
+ "language_model.layers.20.self_attn.k_proj",
93
+ "language_model.layers.7.self_attn.q_proj",
94
+ "45.self_attn.q_proj",
95
+ "language_model.layers.2.self_attn.q_proj",
96
+ "29.self_attn.k_proj",
97
+ "33.self_attn.v_proj",
98
+ "language_model.layers.22.self_attn.k_proj",
99
+ "language_model.layers.10.self_attn.k_proj",
100
+ "44.self_attn.q_proj",
101
+ "42.self_attn.v_proj",
102
+ "language_model.layers.20.self_attn.v_proj",
103
+ "language_model.layers.5.self_attn.v_proj",
104
+ "33.self_attn.k_proj",
105
+ "language_model.layers.20.self_attn.q_proj",
106
+ "language_model.layers.1.self_attn.q_proj",
107
+ "language_model.layers.11.self_attn.v_proj",
108
+ "34.self_attn.q_proj",
109
+ "language_model.layers.8.self_attn.q_proj",
110
+ "language_model.layers.4.self_attn.q_proj",
111
+ "up_proj",
112
+ "language_model.layers.9.self_attn.k_proj",
113
+ "language_model.layers.6.self_attn.v_proj",
114
+ "30.self_attn.q_proj",
115
+ "language_model.layers.3.self_attn.q_proj",
116
+ "language_model.layers.26.self_attn.q_proj",
117
+ "27.self_attn.v_proj",
118
+ "language_model.layers.21.self_attn.k_proj",
119
+ "language_model.layers.12.self_attn.v_proj",
120
+ "39.self_attn.q_proj",
121
+ "language_model.layers.19.self_attn.v_proj",
122
+ "language_model.layers.5.self_attn.k_proj",
123
+ "29.self_attn.q_proj",
124
+ "43.self_attn.v_proj",
125
+ "language_model.layers.9.self_attn.q_proj",
126
+ "language_model.layers.19.self_attn.q_proj",
127
+ "35.self_attn.v_proj",
128
+ "language_model.layers.21.self_attn.q_proj",
129
+ "language_model.layers.4.self_attn.k_proj",
130
+ "32.self_attn.k_proj",
131
+ "language_model.layers.13.self_attn.v_proj",
132
+ "37.self_attn.v_proj",
133
+ "38.self_attn.k_proj",
134
+ "language_model.layers.8.self_attn.v_proj",
135
+ "34.self_attn.v_proj",
136
+ "29.self_attn.v_proj",
137
+ "down_proj",
138
+ "language_model.layers.14.self_attn.k_proj",
139
+ "46.self_attn.k_proj",
140
+ "language_model.layers.3.self_attn.k_proj",
141
+ "37.self_attn.k_proj",
142
+ "41.self_attn.q_proj",
143
+ "42.self_attn.q_proj",
144
+ "language_model.layers.24.self_attn.q_proj",
145
+ "language_model.layers.14.self_attn.v_proj",
146
+ "language_model.layers.6.self_attn.k_proj",
147
+ "language_model.layers.25.self_attn.k_proj",
148
+ "language_model.layers.24.self_attn.v_proj",
149
+ "35.self_attn.k_proj",
150
+ "language_model.layers.22.self_attn.q_proj",
151
+ "41.self_attn.k_proj",
152
+ "40.self_attn.q_proj",
153
+ "45.self_attn.v_proj",
154
+ "language_model.layers.16.self_attn.v_proj",
155
+ "language_model.layers.18.self_attn.k_proj",
156
+ "28.self_attn.k_proj",
157
+ "language_model.layers.15.self_attn.k_proj",
158
+ "language_model.layers.14.self_attn.q_proj",
159
+ "36.self_attn.k_proj",
160
+ "language_model.layers.7.self_attn.v_proj",
161
+ "44.self_attn.v_proj",
162
+ "gate_proj",
163
+ "language_model.layers.13.self_attn.q_proj",
164
+ "language_model.layers.25.self_attn.q_proj",
165
+ "language_model.layers.3.self_attn.v_proj",
166
+ "language_model.layers.1.self_attn.v_proj",
167
+ "language_model.layers.11.self_attn.q_proj",
168
+ "31.self_attn.v_proj",
169
+ "language_model.layers.12.self_attn.k_proj",
170
+ "43.self_attn.k_proj",
171
+ "31.self_attn.q_proj",
172
+ "42.self_attn.k_proj",
173
+ "36.self_attn.q_proj",
174
+ "language_model.layers.26.self_attn.v_proj",
175
+ "language_model.layers.13.self_attn.k_proj",
176
+ "40.self_attn.k_proj",
177
+ "language_model.layers.2.self_attn.k_proj",
178
+ "47.self_attn.q_proj",
179
+ "language_model.layers.16.self_attn.k_proj",
180
+ "31.self_attn.k_proj",
181
+ "30.self_attn.k_proj"
182
+ ],
183
+ "target_parameters": null,
184
+ "task_type": "CAUSAL_LM",
185
+ "trainable_token_indices": null,
186
+ "use_bdlora": null,
187
+ "use_dora": false,
188
+ "use_qalora": false,
189
+ "use_rslora": false,
190
+ "velora_config": null
191
+ }
rl_grpo/control_4x_thinking/checkpoint-32/adapter_model.safetensors ADDED
@@ -0,0 +1,3 @@
 
 
 
 
1
+ version https://git-lfs.github.com/spec/v1
2
+ oid sha256:7fd44b8bb459bf87c8391852168082241b942605287b3e0b6fd002f462babdb7
3
+ size 523864568
rl_grpo/control_4x_thinking/checkpoint-64/adapter_config.json ADDED
@@ -0,0 +1,191 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ {
2
+ "alora_invocation_tokens": null,
3
+ "alpha_pattern": {},
4
+ "arrow_config": null,
5
+ "auto_mapping": null,
6
+ "base_model_name_or_path": "/workspace/rl3t_c/parent",
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": 64,
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": 32,
31
+ "rank_pattern": {},
32
+ "revision": null,
33
+ "target_modules": [
34
+ "32.self_attn.v_proj",
35
+ "language_model.layers.12.self_attn.q_proj",
36
+ "language_model.layers.16.self_attn.q_proj",
37
+ "language_model.layers.10.self_attn.q_proj",
38
+ "language_model.layers.17.self_attn.k_proj",
39
+ "language_model.layers.10.self_attn.v_proj",
40
+ "34.self_attn.k_proj",
41
+ "45.self_attn.k_proj",
42
+ "language_model.layers.17.self_attn.v_proj",
43
+ "44.self_attn.k_proj",
44
+ "46.self_attn.v_proj",
45
+ "39.self_attn.k_proj",
46
+ "41.self_attn.v_proj",
47
+ "30.self_attn.v_proj",
48
+ "language_model.layers.17.self_attn.q_proj",
49
+ "33.self_attn.q_proj",
50
+ "language_model.layers.6.self_attn.q_proj",
51
+ "language_model.layers.23.self_attn.v_proj",
52
+ "language_model.layers.7.self_attn.k_proj",
53
+ "language_model.layers.22.self_attn.v_proj",
54
+ "language_model.layers.4.self_attn.v_proj",
55
+ "language_model.layers.24.self_attn.k_proj",
56
+ "language_model.layers.2.self_attn.v_proj",
57
+ "38.self_attn.v_proj",
58
+ "language_model.layers.11.self_attn.k_proj",
59
+ "language_model.layers.25.self_attn.v_proj",
60
+ "language_model.layers.1.self_attn.k_proj",
61
+ "27.self_attn.k_proj",
62
+ "27.self_attn.q_proj",
63
+ "language_model.layers.19.self_attn.k_proj",
64
+ "39.self_attn.v_proj",
65
+ "28.self_attn.q_proj",
66
+ "32.self_attn.q_proj",
67
+ "46.self_attn.q_proj",
68
+ "language_model.layers.15.self_attn.q_proj",
69
+ "o_proj",
70
+ "language_model.layers.26.self_attn.k_proj",
71
+ "language_model.layers.21.self_attn.v_proj",
72
+ "language_model.layers.23.self_attn.k_proj",
73
+ "43.self_attn.q_proj",
74
+ "language_model.layers.23.self_attn.q_proj",
75
+ "40.self_attn.v_proj",
76
+ "28.self_attn.v_proj",
77
+ "language_model.layers.9.self_attn.v_proj",
78
+ "47.self_attn.k_proj",
79
+ "37.self_attn.q_proj",
80
+ "language_model.layers.8.self_attn.k_proj",
81
+ "language_model.layers.0.self_attn.k_proj",
82
+ "36.self_attn.v_proj",
83
+ "language_model.layers.0.self_attn.q_proj",
84
+ "language_model.layers.18.self_attn.v_proj",
85
+ "language_model.layers.5.self_attn.q_proj",
86
+ "language_model.layers.18.self_attn.q_proj",
87
+ "38.self_attn.q_proj",
88
+ "35.self_attn.q_proj",
89
+ "language_model.layers.15.self_attn.v_proj",
90
+ "47.self_attn.v_proj",
91
+ "language_model.layers.0.self_attn.v_proj",
92
+ "language_model.layers.20.self_attn.k_proj",
93
+ "language_model.layers.7.self_attn.q_proj",
94
+ "45.self_attn.q_proj",
95
+ "language_model.layers.2.self_attn.q_proj",
96
+ "29.self_attn.k_proj",
97
+ "33.self_attn.v_proj",
98
+ "language_model.layers.22.self_attn.k_proj",
99
+ "language_model.layers.10.self_attn.k_proj",
100
+ "44.self_attn.q_proj",
101
+ "42.self_attn.v_proj",
102
+ "language_model.layers.20.self_attn.v_proj",
103
+ "language_model.layers.5.self_attn.v_proj",
104
+ "33.self_attn.k_proj",
105
+ "language_model.layers.20.self_attn.q_proj",
106
+ "language_model.layers.1.self_attn.q_proj",
107
+ "language_model.layers.11.self_attn.v_proj",
108
+ "34.self_attn.q_proj",
109
+ "language_model.layers.8.self_attn.q_proj",
110
+ "language_model.layers.4.self_attn.q_proj",
111
+ "up_proj",
112
+ "language_model.layers.9.self_attn.k_proj",
113
+ "language_model.layers.6.self_attn.v_proj",
114
+ "30.self_attn.q_proj",
115
+ "language_model.layers.3.self_attn.q_proj",
116
+ "language_model.layers.26.self_attn.q_proj",
117
+ "27.self_attn.v_proj",
118
+ "language_model.layers.21.self_attn.k_proj",
119
+ "language_model.layers.12.self_attn.v_proj",
120
+ "39.self_attn.q_proj",
121
+ "language_model.layers.19.self_attn.v_proj",
122
+ "language_model.layers.5.self_attn.k_proj",
123
+ "29.self_attn.q_proj",
124
+ "43.self_attn.v_proj",
125
+ "language_model.layers.9.self_attn.q_proj",
126
+ "language_model.layers.19.self_attn.q_proj",
127
+ "35.self_attn.v_proj",
128
+ "language_model.layers.21.self_attn.q_proj",
129
+ "language_model.layers.4.self_attn.k_proj",
130
+ "32.self_attn.k_proj",
131
+ "language_model.layers.13.self_attn.v_proj",
132
+ "37.self_attn.v_proj",
133
+ "38.self_attn.k_proj",
134
+ "language_model.layers.8.self_attn.v_proj",
135
+ "34.self_attn.v_proj",
136
+ "29.self_attn.v_proj",
137
+ "down_proj",
138
+ "language_model.layers.14.self_attn.k_proj",
139
+ "46.self_attn.k_proj",
140
+ "language_model.layers.3.self_attn.k_proj",
141
+ "37.self_attn.k_proj",
142
+ "41.self_attn.q_proj",
143
+ "42.self_attn.q_proj",
144
+ "language_model.layers.24.self_attn.q_proj",
145
+ "language_model.layers.14.self_attn.v_proj",
146
+ "language_model.layers.6.self_attn.k_proj",
147
+ "language_model.layers.25.self_attn.k_proj",
148
+ "language_model.layers.24.self_attn.v_proj",
149
+ "35.self_attn.k_proj",
150
+ "language_model.layers.22.self_attn.q_proj",
151
+ "41.self_attn.k_proj",
152
+ "40.self_attn.q_proj",
153
+ "45.self_attn.v_proj",
154
+ "language_model.layers.16.self_attn.v_proj",
155
+ "language_model.layers.18.self_attn.k_proj",
156
+ "28.self_attn.k_proj",
157
+ "language_model.layers.15.self_attn.k_proj",
158
+ "language_model.layers.14.self_attn.q_proj",
159
+ "36.self_attn.k_proj",
160
+ "language_model.layers.7.self_attn.v_proj",
161
+ "44.self_attn.v_proj",
162
+ "gate_proj",
163
+ "language_model.layers.13.self_attn.q_proj",
164
+ "language_model.layers.25.self_attn.q_proj",
165
+ "language_model.layers.3.self_attn.v_proj",
166
+ "language_model.layers.1.self_attn.v_proj",
167
+ "language_model.layers.11.self_attn.q_proj",
168
+ "31.self_attn.v_proj",
169
+ "language_model.layers.12.self_attn.k_proj",
170
+ "43.self_attn.k_proj",
171
+ "31.self_attn.q_proj",
172
+ "42.self_attn.k_proj",
173
+ "36.self_attn.q_proj",
174
+ "language_model.layers.26.self_attn.v_proj",
175
+ "language_model.layers.13.self_attn.k_proj",
176
+ "40.self_attn.k_proj",
177
+ "language_model.layers.2.self_attn.k_proj",
178
+ "47.self_attn.q_proj",
179
+ "language_model.layers.16.self_attn.k_proj",
180
+ "31.self_attn.k_proj",
181
+ "30.self_attn.k_proj"
182
+ ],
183
+ "target_parameters": null,
184
+ "task_type": "CAUSAL_LM",
185
+ "trainable_token_indices": null,
186
+ "use_bdlora": null,
187
+ "use_dora": false,
188
+ "use_qalora": false,
189
+ "use_rslora": false,
190
+ "velora_config": null
191
+ }
rl_grpo/control_4x_thinking/checkpoint-64/adapter_model.safetensors ADDED
@@ -0,0 +1,3 @@
 
 
 
 
1
+ version https://git-lfs.github.com/spec/v1
2
+ oid sha256:e5d30367f208e5852102dd9deea83f60110914790452ab4470419d7a8979fd83
3
+ size 523864568
rl_grpo/control_4x_thinking/checkpoint.json ADDED
@@ -0,0 +1,132 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ {
2
+ "experiment": "scimt-train:dispatch-rl-control_4x_thinking",
3
+ "spec": null,
4
+ "kind": null,
5
+ "note": "Spec-free training stage (scimt.train.train_dataset) \u2014 a post-training link in a staged chain, not a spec install.",
6
+ "train": {
7
+ "data": "/workspace/rl3t_c/data/thinking/train.jsonl",
8
+ "dataset_meta": {
9
+ "adhoc": true
10
+ },
11
+ "stage": null,
12
+ "seed": 42,
13
+ "load_checkpoint_path": "/workspace/rl3t_c/parent",
14
+ "grpo": {
15
+ "episodes": 8192,
16
+ "group_size": 8,
17
+ "max_prompt_length": 3072,
18
+ "max_completion_length": 1024,
19
+ "per_device_batch_size": 2,
20
+ "gradient_accumulation_steps": 16,
21
+ "steps_per_generation": 16,
22
+ "learning_rate": 1e-05,
23
+ "temperature": 0.7,
24
+ "loss_type": "dr_grpo",
25
+ "beta": 0.0,
26
+ "vllm": "colocate",
27
+ "vllm_gpu_memory_utilization": 0.38,
28
+ "vllm_max_model_len": 4096,
29
+ "mask_truncated_completions": true,
30
+ "log_completions": true,
31
+ "num_completions_to_print": 2,
32
+ "log_unique_prompts": true,
33
+ "checkpoint_fractions": [
34
+ 0.0625,
35
+ 0.125,
36
+ 0.25,
37
+ 0.5,
38
+ 1.0
39
+ ],
40
+ "report_to": [],
41
+ "reward_func": "experiments.prior_coins.dispatch_rl_reward_v2:reward_thinking",
42
+ "resume_from_checkpoint": null,
43
+ "rollout_log_dir": "/workspace/rl3t_c/training/control_4x_thinking/logs",
44
+ "abort_log_path": null,
45
+ "validation_dataset_path": null,
46
+ "abort_eval_func": null,
47
+ "parent_agreement": null,
48
+ "parent_reward": null,
49
+ "parent_completion_length": null,
50
+ "zero_std_warmup_fraction": 0.1,
51
+ "completion_length_window": 1024
52
+ },
53
+ "lora": {
54
+ "r": 32,
55
+ "alpha": 64,
56
+ "dropout": 0.0,
57
+ "target_linear": true,
58
+ "target_modules": null
59
+ }
60
+ },
61
+ "run_name": "dispatch-rl-control_4x_thinking",
62
+ "pointer_file": "/workspace/rl3t_c/training/control_4x_thinking/train/ckpt_dispatch-rl-control_4x_thinking.txt",
63
+ "backend": "hf_grpo",
64
+ "sampler": "/workspace/rl3t_c/training/control_4x_thinking/train/sampler",
65
+ "state": "/workspace/rl3t_c/training/control_4x_thinking/train/trainer/checkpoint-256",
66
+ "model": "google/gemma-3-12b-pt",
67
+ "meta": {
68
+ "experiment": "scimt-train:dispatch-rl-control_4x_thinking",
69
+ "spec": null,
70
+ "kind": null,
71
+ "note": "Spec-free training stage (scimt.train.train_dataset) \u2014 a post-training link in a staged chain, not a spec install.",
72
+ "train": {
73
+ "data": "/workspace/rl3t_c/data/thinking/train.jsonl",
74
+ "dataset_meta": {
75
+ "adhoc": true
76
+ },
77
+ "stage": null,
78
+ "seed": 42,
79
+ "load_checkpoint_path": "/workspace/rl3t_c/parent",
80
+ "grpo": {
81
+ "episodes": 8192,
82
+ "group_size": 8,
83
+ "max_prompt_length": 3072,
84
+ "max_completion_length": 1024,
85
+ "per_device_batch_size": 2,
86
+ "gradient_accumulation_steps": 16,
87
+ "steps_per_generation": 16,
88
+ "learning_rate": 1e-05,
89
+ "temperature": 0.7,
90
+ "loss_type": "dr_grpo",
91
+ "beta": 0.0,
92
+ "vllm": "colocate",
93
+ "vllm_gpu_memory_utilization": 0.38,
94
+ "vllm_max_model_len": 4096,
95
+ "mask_truncated_completions": true,
96
+ "log_completions": true,
97
+ "num_completions_to_print": 2,
98
+ "log_unique_prompts": true,
99
+ "checkpoint_fractions": [
100
+ 0.0625,
101
+ 0.125,
102
+ 0.25,
103
+ 0.5,
104
+ 1.0
105
+ ],
106
+ "report_to": [],
107
+ "reward_func": "experiments.prior_coins.dispatch_rl_reward_v2:reward_thinking",
108
+ "resume_from_checkpoint": null,
109
+ "rollout_log_dir": "/workspace/rl3t_c/training/control_4x_thinking/logs",
110
+ "abort_log_path": null,
111
+ "validation_dataset_path": null,
112
+ "abort_eval_func": null,
113
+ "parent_agreement": null,
114
+ "parent_reward": null,
115
+ "parent_completion_length": null,
116
+ "zero_std_warmup_fraction": 0.1,
117
+ "completion_length_window": 1024
118
+ },
119
+ "lora": {
120
+ "r": 32,
121
+ "alpha": 64,
122
+ "dropout": 0.0,
123
+ "target_linear": true,
124
+ "target_modules": null
125
+ }
126
+ },
127
+ "run_name": "dispatch-rl-control_4x_thinking",
128
+ "pointer_file": "/workspace/rl3t_c/training/control_4x_thinking/train/ckpt_dispatch-rl-control_4x_thinking.txt"
129
+ },
130
+ "sampler_path": "/workspace/rl3t_c/training/control_4x_thinking/train/sampler",
131
+ "state_path": "/workspace/rl3t_c/training/control_4x_thinking/train/trainer/checkpoint-256"
132
+ }
rl_grpo/control_4x_thinking/logs/raw_rollouts.rank-0.jsonl ADDED
@@ -0,0 +1,3 @@
 
 
 
 
1
+ version https://git-lfs.github.com/spec/v1
2
+ oid sha256:b42da531261db814236f1ef10d9ed6cf870d7ae70d0d9c5bfdd1eff3da981ea3
3
+ size 214747224