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
4403b1c
·
verified ·
1 Parent(s): 14fa133

add rl_grpo/control_4x_direct (copied from sidbaines/scimt-prior-coins-dispatch-sdf-aft-v1)

Browse files
Files changed (14) hide show
  1. rl_grpo/control_4x_direct/RL_TRAINED.json +13 -0
  2. rl_grpo/control_4x_direct/checkpoint-128/adapter_config.json +191 -0
  3. rl_grpo/control_4x_direct/checkpoint-128/adapter_model.safetensors +3 -0
  4. rl_grpo/control_4x_direct/checkpoint-16/adapter_config.json +191 -0
  5. rl_grpo/control_4x_direct/checkpoint-16/adapter_model.safetensors +3 -0
  6. rl_grpo/control_4x_direct/checkpoint-256/adapter_config.json +191 -0
  7. rl_grpo/control_4x_direct/checkpoint-256/adapter_model.safetensors +3 -0
  8. rl_grpo/control_4x_direct/checkpoint-256/trainer_state.json +834 -0
  9. rl_grpo/control_4x_direct/checkpoint-32/adapter_config.json +191 -0
  10. rl_grpo/control_4x_direct/checkpoint-32/adapter_model.safetensors +3 -0
  11. rl_grpo/control_4x_direct/checkpoint-64/adapter_config.json +191 -0
  12. rl_grpo/control_4x_direct/checkpoint-64/adapter_model.safetensors +3 -0
  13. rl_grpo/control_4x_direct/checkpoint.json +132 -0
  14. rl_grpo/control_4x_direct/logs/raw_rollouts.rank-0.jsonl +0 -0
rl_grpo/control_4x_direct/RL_TRAINED.json ADDED
@@ -0,0 +1,13 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ {
2
+ "version": "dispatch_rl_v3",
3
+ "label": "control_4x_direct",
4
+ "mode": "direct",
5
+ "parent": "/workspace/rl3_direct_b/parent",
6
+ "dataset": "/workspace/rl3_direct_b/data/direct/train.jsonl",
7
+ "episodes": 8192,
8
+ "adapter": "/workspace/rl3_direct_b/training/control_4x_direct/train/sampler",
9
+ "minutes": 48.56,
10
+ "dropped_overlong": 0,
11
+ "temperature": 0.7,
12
+ "learning_rate": 1e-05
13
+ }
rl_grpo/control_4x_direct/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/rl3_direct_b/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
+ "28.self_attn.v_proj",
35
+ "language_model.layers.1.self_attn.k_proj",
36
+ "language_model.layers.22.self_attn.k_proj",
37
+ "language_model.layers.25.self_attn.v_proj",
38
+ "45.self_attn.k_proj",
39
+ "27.self_attn.v_proj",
40
+ "language_model.layers.9.self_attn.q_proj",
41
+ "language_model.layers.8.self_attn.q_proj",
42
+ "language_model.layers.23.self_attn.q_proj",
43
+ "31.self_attn.q_proj",
44
+ "language_model.layers.0.self_attn.k_proj",
45
+ "language_model.layers.23.self_attn.k_proj",
46
+ "27.self_attn.q_proj",
47
+ "34.self_attn.v_proj",
48
+ "language_model.layers.10.self_attn.v_proj",
49
+ "language_model.layers.10.self_attn.k_proj",
50
+ "language_model.layers.16.self_attn.k_proj",
51
+ "29.self_attn.q_proj",
52
+ "language_model.layers.19.self_attn.k_proj",
53
+ "28.self_attn.q_proj",
54
+ "47.self_attn.v_proj",
55
+ "43.self_attn.q_proj",
56
+ "language_model.layers.14.self_attn.v_proj",
57
+ "31.self_attn.v_proj",
58
+ "language_model.layers.4.self_attn.v_proj",
59
+ "language_model.layers.4.self_attn.k_proj",
60
+ "language_model.layers.17.self_attn.k_proj",
61
+ "33.self_attn.q_proj",
62
+ "30.self_attn.v_proj",
63
+ "gate_proj",
64
+ "language_model.layers.7.self_attn.k_proj",
65
+ "language_model.layers.18.self_attn.q_proj",
66
+ "language_model.layers.2.self_attn.v_proj",
67
+ "45.self_attn.v_proj",
68
+ "language_model.layers.26.self_attn.k_proj",
69
+ "29.self_attn.k_proj",
70
+ "o_proj",
71
+ "language_model.layers.0.self_attn.q_proj",
72
+ "language_model.layers.15.self_attn.v_proj",
73
+ "language_model.layers.13.self_attn.v_proj",
74
+ "language_model.layers.9.self_attn.v_proj",
75
+ "language_model.layers.16.self_attn.q_proj",
76
+ "language_model.layers.19.self_attn.q_proj",
77
+ "language_model.layers.24.self_attn.q_proj",
78
+ "language_model.layers.11.self_attn.q_proj",
79
+ "40.self_attn.k_proj",
80
+ "46.self_attn.q_proj",
81
+ "language_model.layers.7.self_attn.q_proj",
82
+ "36.self_attn.k_proj",
83
+ "language_model.layers.12.self_attn.q_proj",
84
+ "language_model.layers.21.self_attn.k_proj",
85
+ "41.self_attn.k_proj",
86
+ "40.self_attn.q_proj",
87
+ "32.self_attn.k_proj",
88
+ "language_model.layers.18.self_attn.k_proj",
89
+ "29.self_attn.v_proj",
90
+ "44.self_attn.k_proj",
91
+ "42.self_attn.k_proj",
92
+ "language_model.layers.25.self_attn.q_proj",
93
+ "33.self_attn.v_proj",
94
+ "28.self_attn.k_proj",
95
+ "language_model.layers.24.self_attn.v_proj",
96
+ "language_model.layers.5.self_attn.q_proj",
97
+ "language_model.layers.9.self_attn.k_proj",
98
+ "44.self_attn.v_proj",
99
+ "42.self_attn.v_proj",
100
+ "language_model.layers.11.self_attn.k_proj",
101
+ "32.self_attn.q_proj",
102
+ "language_model.layers.20.self_attn.k_proj",
103
+ "language_model.layers.26.self_attn.q_proj",
104
+ "language_model.layers.14.self_attn.k_proj",
105
+ "30.self_attn.q_proj",
106
+ "27.self_attn.k_proj",
107
+ "40.self_attn.v_proj",
108
+ "37.self_attn.v_proj",
109
+ "31.self_attn.k_proj",
110
+ "language_model.layers.2.self_attn.q_proj",
111
+ "language_model.layers.24.self_attn.k_proj",
112
+ "46.self_attn.k_proj",
113
+ "language_model.layers.12.self_attn.v_proj",
114
+ "41.self_attn.v_proj",
115
+ "language_model.layers.3.self_attn.v_proj",
116
+ "46.self_attn.v_proj",
117
+ "41.self_attn.q_proj",
118
+ "44.self_attn.q_proj",
119
+ "42.self_attn.q_proj",
120
+ "language_model.layers.16.self_attn.v_proj",
121
+ "language_model.layers.4.self_attn.q_proj",
122
+ "language_model.layers.19.self_attn.v_proj",
123
+ "up_proj",
124
+ "32.self_attn.v_proj",
125
+ "35.self_attn.k_proj",
126
+ "language_model.layers.20.self_attn.q_proj",
127
+ "38.self_attn.v_proj",
128
+ "language_model.layers.14.self_attn.q_proj",
129
+ "language_model.layers.20.self_attn.v_proj",
130
+ "43.self_attn.k_proj",
131
+ "language_model.layers.12.self_attn.k_proj",
132
+ "30.self_attn.k_proj",
133
+ "language_model.layers.1.self_attn.q_proj",
134
+ "34.self_attn.k_proj",
135
+ "39.self_attn.q_proj",
136
+ "37.self_attn.q_proj",
137
+ "language_model.layers.18.self_attn.v_proj",
138
+ "language_model.layers.26.self_attn.v_proj",
139
+ "language_model.layers.6.self_attn.q_proj",
140
+ "language_model.layers.21.self_attn.v_proj",
141
+ "language_model.layers.8.self_attn.k_proj",
142
+ "38.self_attn.q_proj",
143
+ "down_proj",
144
+ "language_model.layers.3.self_attn.q_proj",
145
+ "47.self_attn.q_proj",
146
+ "language_model.layers.22.self_attn.v_proj",
147
+ "33.self_attn.k_proj",
148
+ "36.self_attn.q_proj",
149
+ "language_model.layers.15.self_attn.q_proj",
150
+ "language_model.layers.0.self_attn.v_proj",
151
+ "language_model.layers.3.self_attn.k_proj",
152
+ "37.self_attn.k_proj",
153
+ "39.self_attn.v_proj",
154
+ "43.self_attn.v_proj",
155
+ "language_model.layers.2.self_attn.k_proj",
156
+ "language_model.layers.6.self_attn.v_proj",
157
+ "language_model.layers.13.self_attn.k_proj",
158
+ "36.self_attn.v_proj",
159
+ "language_model.layers.11.self_attn.v_proj",
160
+ "language_model.layers.17.self_attn.q_proj",
161
+ "language_model.layers.5.self_attn.v_proj",
162
+ "language_model.layers.10.self_attn.q_proj",
163
+ "38.self_attn.k_proj",
164
+ "35.self_attn.v_proj",
165
+ "language_model.layers.13.self_attn.q_proj",
166
+ "language_model.layers.5.self_attn.k_proj",
167
+ "language_model.layers.7.self_attn.v_proj",
168
+ "45.self_attn.q_proj",
169
+ "34.self_attn.q_proj",
170
+ "35.self_attn.q_proj",
171
+ "language_model.layers.21.self_attn.q_proj",
172
+ "language_model.layers.22.self_attn.q_proj",
173
+ "language_model.layers.1.self_attn.v_proj",
174
+ "language_model.layers.17.self_attn.v_proj",
175
+ "language_model.layers.8.self_attn.v_proj",
176
+ "39.self_attn.k_proj",
177
+ "language_model.layers.6.self_attn.k_proj",
178
+ "language_model.layers.23.self_attn.v_proj",
179
+ "47.self_attn.k_proj",
180
+ "language_model.layers.25.self_attn.k_proj",
181
+ "language_model.layers.15.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_direct/checkpoint-128/adapter_model.safetensors ADDED
@@ -0,0 +1,3 @@
 
 
 
 
1
+ version https://git-lfs.github.com/spec/v1
2
+ oid sha256:323d1c6b74ee4dd756976a72fdfafe86e47080fb34411ddae4d5a6a60fb8fac4
3
+ size 523864568
rl_grpo/control_4x_direct/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/rl3_direct_b/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
+ "28.self_attn.v_proj",
35
+ "language_model.layers.1.self_attn.k_proj",
36
+ "language_model.layers.22.self_attn.k_proj",
37
+ "language_model.layers.25.self_attn.v_proj",
38
+ "45.self_attn.k_proj",
39
+ "27.self_attn.v_proj",
40
+ "language_model.layers.9.self_attn.q_proj",
41
+ "language_model.layers.8.self_attn.q_proj",
42
+ "language_model.layers.23.self_attn.q_proj",
43
+ "31.self_attn.q_proj",
44
+ "language_model.layers.0.self_attn.k_proj",
45
+ "language_model.layers.23.self_attn.k_proj",
46
+ "27.self_attn.q_proj",
47
+ "34.self_attn.v_proj",
48
+ "language_model.layers.10.self_attn.v_proj",
49
+ "language_model.layers.10.self_attn.k_proj",
50
+ "language_model.layers.16.self_attn.k_proj",
51
+ "29.self_attn.q_proj",
52
+ "language_model.layers.19.self_attn.k_proj",
53
+ "28.self_attn.q_proj",
54
+ "47.self_attn.v_proj",
55
+ "43.self_attn.q_proj",
56
+ "language_model.layers.14.self_attn.v_proj",
57
+ "31.self_attn.v_proj",
58
+ "language_model.layers.4.self_attn.v_proj",
59
+ "language_model.layers.4.self_attn.k_proj",
60
+ "language_model.layers.17.self_attn.k_proj",
61
+ "33.self_attn.q_proj",
62
+ "30.self_attn.v_proj",
63
+ "gate_proj",
64
+ "language_model.layers.7.self_attn.k_proj",
65
+ "language_model.layers.18.self_attn.q_proj",
66
+ "language_model.layers.2.self_attn.v_proj",
67
+ "45.self_attn.v_proj",
68
+ "language_model.layers.26.self_attn.k_proj",
69
+ "29.self_attn.k_proj",
70
+ "o_proj",
71
+ "language_model.layers.0.self_attn.q_proj",
72
+ "language_model.layers.15.self_attn.v_proj",
73
+ "language_model.layers.13.self_attn.v_proj",
74
+ "language_model.layers.9.self_attn.v_proj",
75
+ "language_model.layers.16.self_attn.q_proj",
76
+ "language_model.layers.19.self_attn.q_proj",
77
+ "language_model.layers.24.self_attn.q_proj",
78
+ "language_model.layers.11.self_attn.q_proj",
79
+ "40.self_attn.k_proj",
80
+ "46.self_attn.q_proj",
81
+ "language_model.layers.7.self_attn.q_proj",
82
+ "36.self_attn.k_proj",
83
+ "language_model.layers.12.self_attn.q_proj",
84
+ "language_model.layers.21.self_attn.k_proj",
85
+ "41.self_attn.k_proj",
86
+ "40.self_attn.q_proj",
87
+ "32.self_attn.k_proj",
88
+ "language_model.layers.18.self_attn.k_proj",
89
+ "29.self_attn.v_proj",
90
+ "44.self_attn.k_proj",
91
+ "42.self_attn.k_proj",
92
+ "language_model.layers.25.self_attn.q_proj",
93
+ "33.self_attn.v_proj",
94
+ "28.self_attn.k_proj",
95
+ "language_model.layers.24.self_attn.v_proj",
96
+ "language_model.layers.5.self_attn.q_proj",
97
+ "language_model.layers.9.self_attn.k_proj",
98
+ "44.self_attn.v_proj",
99
+ "42.self_attn.v_proj",
100
+ "language_model.layers.11.self_attn.k_proj",
101
+ "32.self_attn.q_proj",
102
+ "language_model.layers.20.self_attn.k_proj",
103
+ "language_model.layers.26.self_attn.q_proj",
104
+ "language_model.layers.14.self_attn.k_proj",
105
+ "30.self_attn.q_proj",
106
+ "27.self_attn.k_proj",
107
+ "40.self_attn.v_proj",
108
+ "37.self_attn.v_proj",
109
+ "31.self_attn.k_proj",
110
+ "language_model.layers.2.self_attn.q_proj",
111
+ "language_model.layers.24.self_attn.k_proj",
112
+ "46.self_attn.k_proj",
113
+ "language_model.layers.12.self_attn.v_proj",
114
+ "41.self_attn.v_proj",
115
+ "language_model.layers.3.self_attn.v_proj",
116
+ "46.self_attn.v_proj",
117
+ "41.self_attn.q_proj",
118
+ "44.self_attn.q_proj",
119
+ "42.self_attn.q_proj",
120
+ "language_model.layers.16.self_attn.v_proj",
121
+ "language_model.layers.4.self_attn.q_proj",
122
+ "language_model.layers.19.self_attn.v_proj",
123
+ "up_proj",
124
+ "32.self_attn.v_proj",
125
+ "35.self_attn.k_proj",
126
+ "language_model.layers.20.self_attn.q_proj",
127
+ "38.self_attn.v_proj",
128
+ "language_model.layers.14.self_attn.q_proj",
129
+ "language_model.layers.20.self_attn.v_proj",
130
+ "43.self_attn.k_proj",
131
+ "language_model.layers.12.self_attn.k_proj",
132
+ "30.self_attn.k_proj",
133
+ "language_model.layers.1.self_attn.q_proj",
134
+ "34.self_attn.k_proj",
135
+ "39.self_attn.q_proj",
136
+ "37.self_attn.q_proj",
137
+ "language_model.layers.18.self_attn.v_proj",
138
+ "language_model.layers.26.self_attn.v_proj",
139
+ "language_model.layers.6.self_attn.q_proj",
140
+ "language_model.layers.21.self_attn.v_proj",
141
+ "language_model.layers.8.self_attn.k_proj",
142
+ "38.self_attn.q_proj",
143
+ "down_proj",
144
+ "language_model.layers.3.self_attn.q_proj",
145
+ "47.self_attn.q_proj",
146
+ "language_model.layers.22.self_attn.v_proj",
147
+ "33.self_attn.k_proj",
148
+ "36.self_attn.q_proj",
149
+ "language_model.layers.15.self_attn.q_proj",
150
+ "language_model.layers.0.self_attn.v_proj",
151
+ "language_model.layers.3.self_attn.k_proj",
152
+ "37.self_attn.k_proj",
153
+ "39.self_attn.v_proj",
154
+ "43.self_attn.v_proj",
155
+ "language_model.layers.2.self_attn.k_proj",
156
+ "language_model.layers.6.self_attn.v_proj",
157
+ "language_model.layers.13.self_attn.k_proj",
158
+ "36.self_attn.v_proj",
159
+ "language_model.layers.11.self_attn.v_proj",
160
+ "language_model.layers.17.self_attn.q_proj",
161
+ "language_model.layers.5.self_attn.v_proj",
162
+ "language_model.layers.10.self_attn.q_proj",
163
+ "38.self_attn.k_proj",
164
+ "35.self_attn.v_proj",
165
+ "language_model.layers.13.self_attn.q_proj",
166
+ "language_model.layers.5.self_attn.k_proj",
167
+ "language_model.layers.7.self_attn.v_proj",
168
+ "45.self_attn.q_proj",
169
+ "34.self_attn.q_proj",
170
+ "35.self_attn.q_proj",
171
+ "language_model.layers.21.self_attn.q_proj",
172
+ "language_model.layers.22.self_attn.q_proj",
173
+ "language_model.layers.1.self_attn.v_proj",
174
+ "language_model.layers.17.self_attn.v_proj",
175
+ "language_model.layers.8.self_attn.v_proj",
176
+ "39.self_attn.k_proj",
177
+ "language_model.layers.6.self_attn.k_proj",
178
+ "language_model.layers.23.self_attn.v_proj",
179
+ "47.self_attn.k_proj",
180
+ "language_model.layers.25.self_attn.k_proj",
181
+ "language_model.layers.15.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_direct/checkpoint-16/adapter_model.safetensors ADDED
@@ -0,0 +1,3 @@
 
 
 
 
1
+ version https://git-lfs.github.com/spec/v1
2
+ oid sha256:acacece8e809eac18be89955b47cb429dc840011236db0f5cec33ac8f8f8cf1c
3
+ size 523864568
rl_grpo/control_4x_direct/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/rl3_direct_b/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
+ "28.self_attn.v_proj",
35
+ "language_model.layers.1.self_attn.k_proj",
36
+ "language_model.layers.22.self_attn.k_proj",
37
+ "language_model.layers.25.self_attn.v_proj",
38
+ "45.self_attn.k_proj",
39
+ "27.self_attn.v_proj",
40
+ "language_model.layers.9.self_attn.q_proj",
41
+ "language_model.layers.8.self_attn.q_proj",
42
+ "language_model.layers.23.self_attn.q_proj",
43
+ "31.self_attn.q_proj",
44
+ "language_model.layers.0.self_attn.k_proj",
45
+ "language_model.layers.23.self_attn.k_proj",
46
+ "27.self_attn.q_proj",
47
+ "34.self_attn.v_proj",
48
+ "language_model.layers.10.self_attn.v_proj",
49
+ "language_model.layers.10.self_attn.k_proj",
50
+ "language_model.layers.16.self_attn.k_proj",
51
+ "29.self_attn.q_proj",
52
+ "language_model.layers.19.self_attn.k_proj",
53
+ "28.self_attn.q_proj",
54
+ "47.self_attn.v_proj",
55
+ "43.self_attn.q_proj",
56
+ "language_model.layers.14.self_attn.v_proj",
57
+ "31.self_attn.v_proj",
58
+ "language_model.layers.4.self_attn.v_proj",
59
+ "language_model.layers.4.self_attn.k_proj",
60
+ "language_model.layers.17.self_attn.k_proj",
61
+ "33.self_attn.q_proj",
62
+ "30.self_attn.v_proj",
63
+ "gate_proj",
64
+ "language_model.layers.7.self_attn.k_proj",
65
+ "language_model.layers.18.self_attn.q_proj",
66
+ "language_model.layers.2.self_attn.v_proj",
67
+ "45.self_attn.v_proj",
68
+ "language_model.layers.26.self_attn.k_proj",
69
+ "29.self_attn.k_proj",
70
+ "o_proj",
71
+ "language_model.layers.0.self_attn.q_proj",
72
+ "language_model.layers.15.self_attn.v_proj",
73
+ "language_model.layers.13.self_attn.v_proj",
74
+ "language_model.layers.9.self_attn.v_proj",
75
+ "language_model.layers.16.self_attn.q_proj",
76
+ "language_model.layers.19.self_attn.q_proj",
77
+ "language_model.layers.24.self_attn.q_proj",
78
+ "language_model.layers.11.self_attn.q_proj",
79
+ "40.self_attn.k_proj",
80
+ "46.self_attn.q_proj",
81
+ "language_model.layers.7.self_attn.q_proj",
82
+ "36.self_attn.k_proj",
83
+ "language_model.layers.12.self_attn.q_proj",
84
+ "language_model.layers.21.self_attn.k_proj",
85
+ "41.self_attn.k_proj",
86
+ "40.self_attn.q_proj",
87
+ "32.self_attn.k_proj",
88
+ "language_model.layers.18.self_attn.k_proj",
89
+ "29.self_attn.v_proj",
90
+ "44.self_attn.k_proj",
91
+ "42.self_attn.k_proj",
92
+ "language_model.layers.25.self_attn.q_proj",
93
+ "33.self_attn.v_proj",
94
+ "28.self_attn.k_proj",
95
+ "language_model.layers.24.self_attn.v_proj",
96
+ "language_model.layers.5.self_attn.q_proj",
97
+ "language_model.layers.9.self_attn.k_proj",
98
+ "44.self_attn.v_proj",
99
+ "42.self_attn.v_proj",
100
+ "language_model.layers.11.self_attn.k_proj",
101
+ "32.self_attn.q_proj",
102
+ "language_model.layers.20.self_attn.k_proj",
103
+ "language_model.layers.26.self_attn.q_proj",
104
+ "language_model.layers.14.self_attn.k_proj",
105
+ "30.self_attn.q_proj",
106
+ "27.self_attn.k_proj",
107
+ "40.self_attn.v_proj",
108
+ "37.self_attn.v_proj",
109
+ "31.self_attn.k_proj",
110
+ "language_model.layers.2.self_attn.q_proj",
111
+ "language_model.layers.24.self_attn.k_proj",
112
+ "46.self_attn.k_proj",
113
+ "language_model.layers.12.self_attn.v_proj",
114
+ "41.self_attn.v_proj",
115
+ "language_model.layers.3.self_attn.v_proj",
116
+ "46.self_attn.v_proj",
117
+ "41.self_attn.q_proj",
118
+ "44.self_attn.q_proj",
119
+ "42.self_attn.q_proj",
120
+ "language_model.layers.16.self_attn.v_proj",
121
+ "language_model.layers.4.self_attn.q_proj",
122
+ "language_model.layers.19.self_attn.v_proj",
123
+ "up_proj",
124
+ "32.self_attn.v_proj",
125
+ "35.self_attn.k_proj",
126
+ "language_model.layers.20.self_attn.q_proj",
127
+ "38.self_attn.v_proj",
128
+ "language_model.layers.14.self_attn.q_proj",
129
+ "language_model.layers.20.self_attn.v_proj",
130
+ "43.self_attn.k_proj",
131
+ "language_model.layers.12.self_attn.k_proj",
132
+ "30.self_attn.k_proj",
133
+ "language_model.layers.1.self_attn.q_proj",
134
+ "34.self_attn.k_proj",
135
+ "39.self_attn.q_proj",
136
+ "37.self_attn.q_proj",
137
+ "language_model.layers.18.self_attn.v_proj",
138
+ "language_model.layers.26.self_attn.v_proj",
139
+ "language_model.layers.6.self_attn.q_proj",
140
+ "language_model.layers.21.self_attn.v_proj",
141
+ "language_model.layers.8.self_attn.k_proj",
142
+ "38.self_attn.q_proj",
143
+ "down_proj",
144
+ "language_model.layers.3.self_attn.q_proj",
145
+ "47.self_attn.q_proj",
146
+ "language_model.layers.22.self_attn.v_proj",
147
+ "33.self_attn.k_proj",
148
+ "36.self_attn.q_proj",
149
+ "language_model.layers.15.self_attn.q_proj",
150
+ "language_model.layers.0.self_attn.v_proj",
151
+ "language_model.layers.3.self_attn.k_proj",
152
+ "37.self_attn.k_proj",
153
+ "39.self_attn.v_proj",
154
+ "43.self_attn.v_proj",
155
+ "language_model.layers.2.self_attn.k_proj",
156
+ "language_model.layers.6.self_attn.v_proj",
157
+ "language_model.layers.13.self_attn.k_proj",
158
+ "36.self_attn.v_proj",
159
+ "language_model.layers.11.self_attn.v_proj",
160
+ "language_model.layers.17.self_attn.q_proj",
161
+ "language_model.layers.5.self_attn.v_proj",
162
+ "language_model.layers.10.self_attn.q_proj",
163
+ "38.self_attn.k_proj",
164
+ "35.self_attn.v_proj",
165
+ "language_model.layers.13.self_attn.q_proj",
166
+ "language_model.layers.5.self_attn.k_proj",
167
+ "language_model.layers.7.self_attn.v_proj",
168
+ "45.self_attn.q_proj",
169
+ "34.self_attn.q_proj",
170
+ "35.self_attn.q_proj",
171
+ "language_model.layers.21.self_attn.q_proj",
172
+ "language_model.layers.22.self_attn.q_proj",
173
+ "language_model.layers.1.self_attn.v_proj",
174
+ "language_model.layers.17.self_attn.v_proj",
175
+ "language_model.layers.8.self_attn.v_proj",
176
+ "39.self_attn.k_proj",
177
+ "language_model.layers.6.self_attn.k_proj",
178
+ "language_model.layers.23.self_attn.v_proj",
179
+ "47.self_attn.k_proj",
180
+ "language_model.layers.25.self_attn.k_proj",
181
+ "language_model.layers.15.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_direct/checkpoint-256/adapter_model.safetensors ADDED
@@ -0,0 +1,3 @@
 
 
 
 
1
+ version https://git-lfs.github.com/spec/v1
2
+ oid sha256:f7067b4606a8c8ffce0b8c4eb6e0a8bca6f6d42739bfc831e95df5a24f00c921
3
+ size 523864568
rl_grpo/control_4x_direct/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": 0.025,
19
+ "completions/max_length": 72.7,
20
+ "completions/max_terminated_length": 32.775,
21
+ "completions/mean_length": 27.725,
22
+ "completions/mean_terminated_length": 21.900297832489013,
23
+ "completions/min_length": 18.25,
24
+ "completions/min_terminated_length": 18.25,
25
+ "entropy": 0.20645992923527956,
26
+ "epoch": 0.0048828125,
27
+ "frac_reward_zero_std": 0.05,
28
+ "grad_norm": 0.09425471723079681,
29
+ "learning_rate": 9.648437500000002e-06,
30
+ "loss": -0.009961342811584473,
31
+ "num_tokens": 224184.0,
32
+ "reward": 0.434375,
33
+ "reward_std": 0.3751695781946182,
34
+ "rewards/reward_func/mean": 0.434375,
35
+ "rewards/reward_func/std": 0.3751695781946182,
36
+ "sampling/importance_sampling_ratio/max": 1.311442382633686,
37
+ "sampling/importance_sampling_ratio/mean": 0.9916193425655365,
38
+ "sampling/importance_sampling_ratio/min": 0.7700684994459153,
39
+ "sampling/sampling_logp_difference/max": 0.26726830303668975,
40
+ "sampling/sampling_logp_difference/mean": 0.013832275808090344,
41
+ "step": 10,
42
+ "step_time": 13.621745687513613
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": 0.028125,
51
+ "completions/max_length": 79.625,
52
+ "completions/max_terminated_length": 39.35,
53
+ "completions/mean_length": 30.3,
54
+ "completions/mean_terminated_length": 24.513244247436525,
55
+ "completions/min_length": 21.0,
56
+ "completions/min_terminated_length": 21.0,
57
+ "entropy": 0.17007166063413023,
58
+ "epoch": 0.009765625,
59
+ "frac_reward_zero_std": 0.075,
60
+ "grad_norm": 0.07826346904039383,
61
+ "learning_rate": 9.2578125e-06,
62
+ "loss": -0.003023067116737366,
63
+ "num_tokens": 462728.0,
64
+ "reward": 0.3953125,
65
+ "reward_std": 0.3438856557011604,
66
+ "rewards/reward_func/mean": 0.3953125,
67
+ "rewards/reward_func/std": 0.3438856557011604,
68
+ "sampling/importance_sampling_ratio/max": 1.552274227142334,
69
+ "sampling/importance_sampling_ratio/mean": 0.9810440734028816,
70
+ "sampling/importance_sampling_ratio/min": 0.3766255252063274,
71
+ "sampling/sampling_logp_difference/max": 0.9436389967799187,
72
+ "sampling/sampling_logp_difference/mean": 0.02956465636380017,
73
+ "step": 20,
74
+ "step_time": 14.440491062961518
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": 0.00625,
83
+ "completions/max_length": 44.225,
84
+ "completions/max_terminated_length": 32.85,
85
+ "completions/mean_length": 24.21875,
86
+ "completions/mean_terminated_length": 22.787053632736207,
87
+ "completions/min_length": 20.6,
88
+ "completions/min_terminated_length": 20.6,
89
+ "entropy": 0.10943786213174463,
90
+ "epoch": 0.0146484375,
91
+ "frac_reward_zero_std": 0.325,
92
+ "grad_norm": 0.15092740952968597,
93
+ "learning_rate": 8.867187500000001e-06,
94
+ "loss": -0.006308811157941819,
95
+ "num_tokens": 690214.0,
96
+ "reward": 0.6296875,
97
+ "reward_std": 0.2346084039658308,
98
+ "rewards/reward_func/mean": 0.6296875,
99
+ "rewards/reward_func/std": 0.2346084039658308,
100
+ "sampling/importance_sampling_ratio/max": 1.5436350837349893,
101
+ "sampling/importance_sampling_ratio/mean": 0.961260037124157,
102
+ "sampling/importance_sampling_ratio/min": 0.3657877079676837,
103
+ "sampling/sampling_logp_difference/max": 1.4533971292898058,
104
+ "sampling/sampling_logp_difference/mean": 0.031207872147206217,
105
+ "step": 30,
106
+ "step_time": 11.861421128408983
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": 0.0,
115
+ "completions/max_length": 27.85,
116
+ "completions/max_terminated_length": 27.85,
117
+ "completions/mean_length": 21.678125,
118
+ "completions/mean_terminated_length": 21.678125,
119
+ "completions/min_length": 20.4,
120
+ "completions/min_terminated_length": 20.4,
121
+ "entropy": 0.07125286967493594,
122
+ "epoch": 0.01953125,
123
+ "frac_reward_zero_std": 0.575,
124
+ "grad_norm": 0.24056176841259003,
125
+ "learning_rate": 8.4765625e-06,
126
+ "loss": -0.0008091941475868225,
127
+ "num_tokens": 912823.0,
128
+ "reward": 0.646875,
129
+ "reward_std": 0.13791176080703735,
130
+ "rewards/reward_func/mean": 0.646875,
131
+ "rewards/reward_func/std": 0.13791176080703735,
132
+ "sampling/importance_sampling_ratio/max": 1.2505815471755342,
133
+ "sampling/importance_sampling_ratio/mean": 0.8135982143459841,
134
+ "sampling/importance_sampling_ratio/min": 0.3564261724299286,
135
+ "sampling/sampling_logp_difference/max": 1.6297439455986023,
136
+ "sampling/sampling_logp_difference/mean": 0.04854685590835288,
137
+ "step": 40,
138
+ "step_time": 10.765326025546528
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": 0.003125,
147
+ "completions/max_length": 34.625,
148
+ "completions/max_terminated_length": 28.9,
149
+ "completions/mean_length": 22.159375,
150
+ "completions/mean_terminated_length": 21.43616075515747,
151
+ "completions/min_length": 20.025,
152
+ "completions/min_terminated_length": 20.025,
153
+ "entropy": 0.05924436629284173,
154
+ "epoch": 0.0244140625,
155
+ "frac_reward_zero_std": 0.675,
156
+ "grad_norm": 0.0,
157
+ "learning_rate": 8.085937500000001e-06,
158
+ "loss": -0.007848377525806426,
159
+ "num_tokens": 1136274.0,
160
+ "reward": 0.63125,
161
+ "reward_std": 0.11495114155113698,
162
+ "rewards/reward_func/mean": 0.63125,
163
+ "rewards/reward_func/std": 0.11495114155113698,
164
+ "sampling/importance_sampling_ratio/max": 1.0875981040298939,
165
+ "sampling/importance_sampling_ratio/mean": 0.8007158821448683,
166
+ "sampling/importance_sampling_ratio/min": 0.330442966746341,
167
+ "sampling/sampling_logp_difference/max": 1.781578005850315,
168
+ "sampling/sampling_logp_difference/mean": 0.04431041326606646,
169
+ "step": 50,
170
+ "step_time": 11.050053541082889
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": 0.003125,
179
+ "completions/max_length": 28.525,
180
+ "completions/max_terminated_length": 22.525,
181
+ "completions/mean_length": 21.609375,
182
+ "completions/mean_terminated_length": 20.859375,
183
+ "completions/min_length": 20.275,
184
+ "completions/min_terminated_length": 20.275,
185
+ "entropy": 0.052522658626548945,
186
+ "epoch": 0.029296875,
187
+ "frac_reward_zero_std": 0.575,
188
+ "grad_norm": 0.16273297369480133,
189
+ "learning_rate": 7.6953125e-06,
190
+ "loss": -0.006116054579615593,
191
+ "num_tokens": 1362101.0,
192
+ "reward": 0.6859375,
193
+ "reward_std": 0.12590925730764865,
194
+ "rewards/reward_func/mean": 0.6859375,
195
+ "rewards/reward_func/std": 0.12590925730764865,
196
+ "sampling/importance_sampling_ratio/max": 1.1167198088951409,
197
+ "sampling/importance_sampling_ratio/mean": 0.8936128420755267,
198
+ "sampling/importance_sampling_ratio/min": 0.46313620423464813,
199
+ "sampling/sampling_logp_difference/max": 1.4776780301705004,
200
+ "sampling/sampling_logp_difference/mean": 0.03263402148731984,
201
+ "step": 60,
202
+ "step_time": 10.778436978626996
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": 0.0,
211
+ "completions/max_length": 22.05,
212
+ "completions/max_terminated_length": 22.05,
213
+ "completions/mean_length": 20.4125,
214
+ "completions/mean_terminated_length": 20.4125,
215
+ "completions/min_length": 19.95,
216
+ "completions/min_terminated_length": 19.95,
217
+ "entropy": 0.05537974808830768,
218
+ "epoch": 0.0341796875,
219
+ "frac_reward_zero_std": 0.675,
220
+ "grad_norm": 0.0,
221
+ "learning_rate": 7.3046875000000005e-06,
222
+ "loss": -0.00045838202349841596,
223
+ "num_tokens": 1586225.0,
224
+ "reward": 0.8203125,
225
+ "reward_std": 0.11835320964455605,
226
+ "rewards/reward_func/mean": 0.8203125,
227
+ "rewards/reward_func/std": 0.11835320964455605,
228
+ "sampling/importance_sampling_ratio/max": 1.069625974689552,
229
+ "sampling/importance_sampling_ratio/mean": 0.8583247686372488,
230
+ "sampling/importance_sampling_ratio/min": 0.5171208131927415,
231
+ "sampling/sampling_logp_difference/max": 0.9130201599095017,
232
+ "sampling/sampling_logp_difference/mean": 0.03456720225949539,
233
+ "step": 70,
234
+ "step_time": 10.363929512281903
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": 0.0,
243
+ "completions/max_length": 22.4,
244
+ "completions/max_terminated_length": 22.4,
245
+ "completions/mean_length": 20.3375,
246
+ "completions/mean_terminated_length": 20.3375,
247
+ "completions/min_length": 19.85,
248
+ "completions/min_terminated_length": 19.85,
249
+ "entropy": 0.05415882607921958,
250
+ "epoch": 0.0390625,
251
+ "frac_reward_zero_std": 0.75,
252
+ "grad_norm": 0.0,
253
+ "learning_rate": 6.9140625e-06,
254
+ "loss": -0.006630612909793854,
255
+ "num_tokens": 1800437.0,
256
+ "reward": 0.7515625,
257
+ "reward_std": 0.07652263343334198,
258
+ "rewards/reward_func/mean": 0.7515625,
259
+ "rewards/reward_func/std": 0.07652263343334198,
260
+ "sampling/importance_sampling_ratio/max": 1.1442169517162255,
261
+ "sampling/importance_sampling_ratio/mean": 0.8566872804425657,
262
+ "sampling/importance_sampling_ratio/min": 0.45216031289601233,
263
+ "sampling/sampling_logp_difference/max": 1.3490417491644622,
264
+ "sampling/sampling_logp_difference/mean": 0.031430702609941366,
265
+ "step": 80,
266
+ "step_time": 10.047211606008933
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": 0.0,
275
+ "completions/max_length": 27.625,
276
+ "completions/max_terminated_length": 27.625,
277
+ "completions/mean_length": 21.778125,
278
+ "completions/mean_terminated_length": 21.778125,
279
+ "completions/min_length": 20.775,
280
+ "completions/min_terminated_length": 20.775,
281
+ "entropy": 0.05513966418802738,
282
+ "epoch": 0.0439453125,
283
+ "frac_reward_zero_std": 0.8,
284
+ "grad_norm": 0.009913666173815727,
285
+ "learning_rate": 6.5234375e-06,
286
+ "loss": -0.0063079454004764555,
287
+ "num_tokens": 2036358.0,
288
+ "reward": 0.7828125,
289
+ "reward_std": 0.04545377567410469,
290
+ "rewards/reward_func/mean": 0.7828125,
291
+ "rewards/reward_func/std": 0.04545377567410469,
292
+ "sampling/importance_sampling_ratio/max": 1.0621035307645799,
293
+ "sampling/importance_sampling_ratio/mean": 0.8827658608555794,
294
+ "sampling/importance_sampling_ratio/min": 0.5341659176192479,
295
+ "sampling/sampling_logp_difference/max": 1.0174065412953497,
296
+ "sampling/sampling_logp_difference/mean": 0.02145701235276647,
297
+ "step": 90,
298
+ "step_time": 11.223674135352486
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": 0.003125,
307
+ "completions/max_length": 27.975,
308
+ "completions/max_terminated_length": 22.25,
309
+ "completions/mean_length": 21.653125,
310
+ "completions/mean_terminated_length": 20.932142877578734,
311
+ "completions/min_length": 20.625,
312
+ "completions/min_terminated_length": 20.625,
313
+ "entropy": 0.042366575170308354,
314
+ "epoch": 0.048828125,
315
+ "frac_reward_zero_std": 0.8,
316
+ "grad_norm": 0.004494183696806431,
317
+ "learning_rate": 6.132812500000001e-06,
318
+ "loss": -0.0017893832176923753,
319
+ "num_tokens": 2264295.0,
320
+ "reward": 0.784375,
321
+ "reward_std": 0.0588643092662096,
322
+ "rewards/reward_func/mean": 0.784375,
323
+ "rewards/reward_func/std": 0.0588643092662096,
324
+ "sampling/importance_sampling_ratio/max": 1.1741500362753867,
325
+ "sampling/importance_sampling_ratio/mean": 0.9192762583494186,
326
+ "sampling/importance_sampling_ratio/min": 0.5411814996172325,
327
+ "sampling/sampling_logp_difference/max": 0.9970868226140738,
328
+ "sampling/sampling_logp_difference/mean": 0.024688148850691503,
329
+ "step": 100,
330
+ "step_time": 10.722826294903644
331
+ },
332
+ {
333
+ "clip_ratio/high_max": 0.0009999999776482583,
334
+ "clip_ratio/high_mean": 0.0003749999916180968,
335
+ "clip_ratio/low_mean": 0.0,
336
+ "clip_ratio/low_min": 0.0,
337
+ "clip_ratio/region_mean": 0.0003749999916180968,
338
+ "completions/clipped_ratio": 0.0,
339
+ "completions/max_length": 23.0,
340
+ "completions/max_terminated_length": 23.0,
341
+ "completions/mean_length": 21.11875,
342
+ "completions/mean_terminated_length": 21.11875,
343
+ "completions/min_length": 20.25,
344
+ "completions/min_terminated_length": 20.25,
345
+ "entropy": 0.044411231228150425,
346
+ "epoch": 0.0537109375,
347
+ "frac_reward_zero_std": 0.675,
348
+ "grad_norm": 0.13850340247154236,
349
+ "learning_rate": 5.742187500000001e-06,
350
+ "loss": 0.004874607920646668,
351
+ "num_tokens": 2490965.0,
352
+ "reward": 0.7359375,
353
+ "reward_std": 0.10600103959441184,
354
+ "rewards/reward_func/mean": 0.7359375,
355
+ "rewards/reward_func/std": 0.10600103959441184,
356
+ "sampling/importance_sampling_ratio/max": 1.064456513787445,
357
+ "sampling/importance_sampling_ratio/mean": 0.7856207518299925,
358
+ "sampling/importance_sampling_ratio/min": 0.46712204770547033,
359
+ "sampling/sampling_logp_difference/max": 1.712026689387858,
360
+ "sampling/sampling_logp_difference/mean": 0.054023325006710365,
361
+ "step": 110,
362
+ "step_time": 10.362062294571661
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": 0.0,
371
+ "completions/max_length": 23.475,
372
+ "completions/max_terminated_length": 23.475,
373
+ "completions/mean_length": 19.696875,
374
+ "completions/mean_terminated_length": 19.696875,
375
+ "completions/min_length": 19.05,
376
+ "completions/min_terminated_length": 19.05,
377
+ "entropy": 0.05922343819402158,
378
+ "epoch": 0.05859375,
379
+ "frac_reward_zero_std": 0.825,
380
+ "grad_norm": 0.0,
381
+ "learning_rate": 5.3515625000000005e-06,
382
+ "loss": 0.00022218506783246993,
383
+ "num_tokens": 2690788.0,
384
+ "reward": 0.778125,
385
+ "reward_std": 0.06519497111439705,
386
+ "rewards/reward_func/mean": 0.778125,
387
+ "rewards/reward_func/std": 0.06519497111439705,
388
+ "sampling/importance_sampling_ratio/max": 1.1133794834197033,
389
+ "sampling/importance_sampling_ratio/mean": 0.9058507873967756,
390
+ "sampling/importance_sampling_ratio/min": 0.5693986285477877,
391
+ "sampling/sampling_logp_difference/max": 0.7275112386792898,
392
+ "sampling/sampling_logp_difference/mean": 0.02245992979442235,
393
+ "step": 120,
394
+ "step_time": 9.425406742095948
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": 0.0,
403
+ "completions/max_length": 21.25,
404
+ "completions/max_terminated_length": 21.25,
405
+ "completions/mean_length": 19.884375,
406
+ "completions/mean_terminated_length": 19.884375,
407
+ "completions/min_length": 19.525,
408
+ "completions/min_terminated_length": 19.525,
409
+ "entropy": 0.03547353339381516,
410
+ "epoch": 0.0634765625,
411
+ "frac_reward_zero_std": 0.925,
412
+ "grad_norm": 0.0,
413
+ "learning_rate": 4.9609375e-06,
414
+ "loss": -0.002609885111451149,
415
+ "num_tokens": 2902247.0,
416
+ "reward": 0.74375,
417
+ "reward_std": 0.019727617129683496,
418
+ "rewards/reward_func/mean": 0.74375,
419
+ "rewards/reward_func/std": 0.019727617129683496,
420
+ "sampling/importance_sampling_ratio/max": 1.114166632387787,
421
+ "sampling/importance_sampling_ratio/mean": 0.9661277275765314,
422
+ "sampling/importance_sampling_ratio/min": 0.7674026334891095,
423
+ "sampling/sampling_logp_difference/max": 0.5253889178857207,
424
+ "sampling/sampling_logp_difference/mean": 0.014963770058238878,
425
+ "step": 130,
426
+ "step_time": 9.826964975590817
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": 0.0,
435
+ "completions/max_length": 26.825,
436
+ "completions/max_terminated_length": 26.825,
437
+ "completions/mean_length": 20.95625,
438
+ "completions/mean_terminated_length": 20.95625,
439
+ "completions/min_length": 19.9,
440
+ "completions/min_terminated_length": 19.9,
441
+ "entropy": 0.03855611260514706,
442
+ "epoch": 0.068359375,
443
+ "frac_reward_zero_std": 0.9,
444
+ "grad_norm": 0.0297012310475111,
445
+ "learning_rate": 4.5703125e-06,
446
+ "loss": 0.00613410472869873,
447
+ "num_tokens": 3123345.0,
448
+ "reward": 0.809375,
449
+ "reward_std": 0.02346404492855072,
450
+ "rewards/reward_func/mean": 0.809375,
451
+ "rewards/reward_func/std": 0.02346404492855072,
452
+ "sampling/importance_sampling_ratio/max": 1.1988712593913078,
453
+ "sampling/importance_sampling_ratio/mean": 1.0076340302824973,
454
+ "sampling/importance_sampling_ratio/min": 0.7764216745970771,
455
+ "sampling/sampling_logp_difference/max": 0.5479300385806709,
456
+ "sampling/sampling_logp_difference/mean": 0.00877691409914405,
457
+ "step": 140,
458
+ "step_time": 10.500897655263543
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": 0.0,
467
+ "completions/max_length": 22.475,
468
+ "completions/max_terminated_length": 22.475,
469
+ "completions/mean_length": 20.99375,
470
+ "completions/mean_terminated_length": 20.99375,
471
+ "completions/min_length": 20.525,
472
+ "completions/min_terminated_length": 20.525,
473
+ "entropy": 0.036207933886907996,
474
+ "epoch": 0.0732421875,
475
+ "frac_reward_zero_std": 0.725,
476
+ "grad_norm": 0.0,
477
+ "learning_rate": 4.1796875e-06,
478
+ "loss": 0.0010990417562425137,
479
+ "num_tokens": 3348615.0,
480
+ "reward": 0.7890625,
481
+ "reward_std": 0.06975407116115093,
482
+ "rewards/reward_func/mean": 0.7890625,
483
+ "rewards/reward_func/std": 0.06975407116115093,
484
+ "sampling/importance_sampling_ratio/max": 1.0597496312810109,
485
+ "sampling/importance_sampling_ratio/mean": 0.9259018765587825,
486
+ "sampling/importance_sampling_ratio/min": 0.6415333994431421,
487
+ "sampling/sampling_logp_difference/max": 1.0824034682358614,
488
+ "sampling/sampling_logp_difference/mean": 0.01923654735146556,
489
+ "step": 150,
490
+ "step_time": 10.479148641903885
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": 0.0,
499
+ "completions/max_length": 22.75,
500
+ "completions/max_terminated_length": 22.75,
501
+ "completions/mean_length": 20.9375,
502
+ "completions/mean_terminated_length": 20.9375,
503
+ "completions/min_length": 20.45,
504
+ "completions/min_terminated_length": 20.45,
505
+ "entropy": 0.038048108806833625,
506
+ "epoch": 0.078125,
507
+ "frac_reward_zero_std": 0.775,
508
+ "grad_norm": 0.5779051184654236,
509
+ "learning_rate": 3.7890625e-06,
510
+ "loss": -0.0014641446061432362,
511
+ "num_tokens": 3571307.0,
512
+ "reward": 0.765625,
513
+ "reward_std": 0.067339675873518,
514
+ "rewards/reward_func/mean": 0.765625,
515
+ "rewards/reward_func/std": 0.067339675873518,
516
+ "sampling/importance_sampling_ratio/max": 1.0659763468895107,
517
+ "sampling/importance_sampling_ratio/mean": 0.9048170571681112,
518
+ "sampling/importance_sampling_ratio/min": 0.7217138898442499,
519
+ "sampling/sampling_logp_difference/max": 0.6475832529831678,
520
+ "sampling/sampling_logp_difference/mean": 0.016887320946261754,
521
+ "step": 160,
522
+ "step_time": 10.295554198813624
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": 0.0,
531
+ "completions/max_length": 22.5,
532
+ "completions/max_terminated_length": 22.5,
533
+ "completions/mean_length": 20.9875,
534
+ "completions/mean_terminated_length": 20.9875,
535
+ "completions/min_length": 20.5,
536
+ "completions/min_terminated_length": 20.5,
537
+ "entropy": 0.035747832129709424,
538
+ "epoch": 0.0830078125,
539
+ "frac_reward_zero_std": 0.725,
540
+ "grad_norm": 0.0,
541
+ "learning_rate": 3.3984375000000004e-06,
542
+ "loss": 0.0010564704425632954,
543
+ "num_tokens": 3798271.0,
544
+ "reward": 0.8375,
545
+ "reward_std": 0.09189608357846737,
546
+ "rewards/reward_func/mean": 0.8375,
547
+ "rewards/reward_func/std": 0.09189608357846737,
548
+ "sampling/importance_sampling_ratio/max": 1.1843287393450737,
549
+ "sampling/importance_sampling_ratio/mean": 1.001787456870079,
550
+ "sampling/importance_sampling_ratio/min": 0.7098866928688949,
551
+ "sampling/sampling_logp_difference/max": 0.8293700889218598,
552
+ "sampling/sampling_logp_difference/mean": 0.01597311543227988,
553
+ "step": 170,
554
+ "step_time": 10.606972057255916
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": 0.0,
563
+ "completions/max_length": 22.55,
564
+ "completions/max_terminated_length": 22.55,
565
+ "completions/mean_length": 21.1125,
566
+ "completions/mean_terminated_length": 21.1125,
567
+ "completions/min_length": 20.7,
568
+ "completions/min_terminated_length": 20.7,
569
+ "entropy": 0.033882539323531094,
570
+ "epoch": 0.087890625,
571
+ "frac_reward_zero_std": 0.8,
572
+ "grad_norm": 0.013551953248679638,
573
+ "learning_rate": 3.0078125000000003e-06,
574
+ "loss": -0.003285379707813263,
575
+ "num_tokens": 4022147.0,
576
+ "reward": 0.853125,
577
+ "reward_std": 0.054467161372303965,
578
+ "rewards/reward_func/mean": 0.853125,
579
+ "rewards/reward_func/std": 0.054467161372303965,
580
+ "sampling/importance_sampling_ratio/max": 1.1137135684490205,
581
+ "sampling/importance_sampling_ratio/mean": 0.9377105679363013,
582
+ "sampling/importance_sampling_ratio/min": 0.6365731901722029,
583
+ "sampling/sampling_logp_difference/max": 0.9814381817355752,
584
+ "sampling/sampling_logp_difference/mean": 0.016642241284716874,
585
+ "step": 180,
586
+ "step_time": 10.302958910865708
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": 0.0,
595
+ "completions/max_length": 22.075,
596
+ "completions/max_terminated_length": 22.075,
597
+ "completions/mean_length": 20.746875,
598
+ "completions/mean_terminated_length": 20.746875,
599
+ "completions/min_length": 20.35,
600
+ "completions/min_terminated_length": 20.35,
601
+ "entropy": 0.03087269093375653,
602
+ "epoch": 0.0927734375,
603
+ "frac_reward_zero_std": 0.85,
604
+ "grad_norm": 0.0333615280687809,
605
+ "learning_rate": 2.6171875e-06,
606
+ "loss": -0.0023753052577376367,
607
+ "num_tokens": 4251578.0,
608
+ "reward": 0.8359375,
609
+ "reward_std": 0.046608568355441096,
610
+ "rewards/reward_func/mean": 0.8359375,
611
+ "rewards/reward_func/std": 0.046608568355441096,
612
+ "sampling/importance_sampling_ratio/max": 1.0908331140875815,
613
+ "sampling/importance_sampling_ratio/mean": 0.9769783303141594,
614
+ "sampling/importance_sampling_ratio/min": 0.7424357914627763,
615
+ "sampling/sampling_logp_difference/max": 0.6119592327158898,
616
+ "sampling/sampling_logp_difference/mean": 0.009374293391010724,
617
+ "step": 190,
618
+ "step_time": 10.53276472822763
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": 0.0,
627
+ "completions/max_length": 22.2,
628
+ "completions/max_terminated_length": 22.2,
629
+ "completions/mean_length": 20.8125,
630
+ "completions/mean_terminated_length": 20.8125,
631
+ "completions/min_length": 20.375,
632
+ "completions/min_terminated_length": 20.375,
633
+ "entropy": 0.03330791248008609,
634
+ "epoch": 0.09765625,
635
+ "frac_reward_zero_std": 0.75,
636
+ "grad_norm": 0.16599611937999725,
637
+ "learning_rate": 2.2265625000000003e-06,
638
+ "loss": -0.008330522477626801,
639
+ "num_tokens": 4482422.0,
640
+ "reward": 0.753125,
641
+ "reward_std": 0.07222737111151219,
642
+ "rewards/reward_func/mean": 0.753125,
643
+ "rewards/reward_func/std": 0.07222737111151219,
644
+ "sampling/importance_sampling_ratio/max": 1.1444828119594603,
645
+ "sampling/importance_sampling_ratio/mean": 0.9122603974701405,
646
+ "sampling/importance_sampling_ratio/min": 0.6283474483560895,
647
+ "sampling/sampling_logp_difference/max": 1.158764861524105,
648
+ "sampling/sampling_logp_difference/mean": 0.035384934321336915,
649
+ "step": 200,
650
+ "step_time": 10.551428495021536
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": 0.0,
659
+ "completions/max_length": 21.4,
660
+ "completions/max_terminated_length": 21.4,
661
+ "completions/mean_length": 20.178125,
662
+ "completions/mean_terminated_length": 20.178125,
663
+ "completions/min_length": 19.85,
664
+ "completions/min_terminated_length": 19.85,
665
+ "entropy": 0.033770469832234085,
666
+ "epoch": 0.1025390625,
667
+ "frac_reward_zero_std": 0.65,
668
+ "grad_norm": 0.006895879749208689,
669
+ "learning_rate": 1.8359375000000002e-06,
670
+ "loss": -0.0038197148591279984,
671
+ "num_tokens": 4705239.0,
672
+ "reward": 0.6484375,
673
+ "reward_std": 0.10259594917297363,
674
+ "rewards/reward_func/mean": 0.6484375,
675
+ "rewards/reward_func/std": 0.10259594917297363,
676
+ "sampling/importance_sampling_ratio/max": 1.1333670931693631,
677
+ "sampling/importance_sampling_ratio/mean": 0.9318572779826354,
678
+ "sampling/importance_sampling_ratio/min": 0.6148406468622852,
679
+ "sampling/sampling_logp_difference/max": 1.1026633579982446,
680
+ "sampling/sampling_logp_difference/mean": 0.02985095392796211,
681
+ "step": 210,
682
+ "step_time": 10.065247177379206
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": 0.0,
691
+ "completions/max_length": 26.25,
692
+ "completions/max_terminated_length": 26.25,
693
+ "completions/mean_length": 22.075,
694
+ "completions/mean_terminated_length": 22.075,
695
+ "completions/min_length": 21.2,
696
+ "completions/min_terminated_length": 21.2,
697
+ "entropy": 0.032827228563837706,
698
+ "epoch": 0.107421875,
699
+ "frac_reward_zero_std": 0.775,
700
+ "grad_norm": 0.0,
701
+ "learning_rate": 1.4453125000000002e-06,
702
+ "loss": -0.0012193472124636174,
703
+ "num_tokens": 4944535.0,
704
+ "reward": 0.7921875,
705
+ "reward_std": 0.07121706642210483,
706
+ "rewards/reward_func/mean": 0.7921875,
707
+ "rewards/reward_func/std": 0.07121706642210483,
708
+ "sampling/importance_sampling_ratio/max": 1.0614363917302398,
709
+ "sampling/importance_sampling_ratio/mean": 0.8753228086490707,
710
+ "sampling/importance_sampling_ratio/min": 0.5934614542433337,
711
+ "sampling/sampling_logp_difference/max": 1.567086279625073,
712
+ "sampling/sampling_logp_difference/mean": 0.034673807610670335,
713
+ "step": 220,
714
+ "step_time": 11.150762310205028
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": 0.0,
723
+ "completions/max_length": 22.125,
724
+ "completions/max_terminated_length": 22.125,
725
+ "completions/mean_length": 19.41875,
726
+ "completions/mean_terminated_length": 19.41875,
727
+ "completions/min_length": 18.9,
728
+ "completions/min_terminated_length": 18.9,
729
+ "entropy": 0.03635700640734285,
730
+ "epoch": 0.1123046875,
731
+ "frac_reward_zero_std": 0.9,
732
+ "grad_norm": 0.0,
733
+ "learning_rate": 1.0546875e-06,
734
+ "loss": -0.0027765242382884027,
735
+ "num_tokens": 5146645.0,
736
+ "reward": 0.76875,
737
+ "reward_std": 0.02619796171784401,
738
+ "rewards/reward_func/mean": 0.76875,
739
+ "rewards/reward_func/std": 0.02619796171784401,
740
+ "sampling/importance_sampling_ratio/max": 1.0704911544919014,
741
+ "sampling/importance_sampling_ratio/mean": 0.9547419771552086,
742
+ "sampling/importance_sampling_ratio/min": 0.7945206198841334,
743
+ "sampling/sampling_logp_difference/max": 0.33006448808591815,
744
+ "sampling/sampling_logp_difference/mean": 0.010264440209721216,
745
+ "step": 230,
746
+ "step_time": 9.255960304173641
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": 0.0,
755
+ "completions/max_length": 22.85,
756
+ "completions/max_terminated_length": 22.85,
757
+ "completions/mean_length": 21.678125,
758
+ "completions/mean_terminated_length": 21.678125,
759
+ "completions/min_length": 21.325,
760
+ "completions/min_terminated_length": 21.325,
761
+ "entropy": 0.03218276887200773,
762
+ "epoch": 0.1171875,
763
+ "frac_reward_zero_std": 0.825,
764
+ "grad_norm": 0.5462760329246521,
765
+ "learning_rate": 6.640625e-07,
766
+ "loss": 8.534137159585953e-05,
767
+ "num_tokens": 5384422.0,
768
+ "reward": 0.846875,
769
+ "reward_std": 0.05034401677548885,
770
+ "rewards/reward_func/mean": 0.846875,
771
+ "rewards/reward_func/std": 0.05034401677548885,
772
+ "sampling/importance_sampling_ratio/max": 1.1618405058979988,
773
+ "sampling/importance_sampling_ratio/mean": 0.9789710864424706,
774
+ "sampling/importance_sampling_ratio/min": 0.7448184862951166,
775
+ "sampling/sampling_logp_difference/max": 0.6628234723117202,
776
+ "sampling/sampling_logp_difference/mean": 0.011719785288732965,
777
+ "step": 240,
778
+ "step_time": 10.748754722694866
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": 0.0,
787
+ "completions/max_length": 21.55,
788
+ "completions/max_terminated_length": 21.55,
789
+ "completions/mean_length": 20.109375,
790
+ "completions/mean_terminated_length": 20.109375,
791
+ "completions/min_length": 19.8,
792
+ "completions/min_terminated_length": 19.8,
793
+ "entropy": 0.02970391695853323,
794
+ "epoch": 0.1220703125,
795
+ "frac_reward_zero_std": 0.825,
796
+ "grad_norm": 0.0,
797
+ "learning_rate": 2.7343750000000003e-07,
798
+ "loss": 0.00286461990326643,
799
+ "num_tokens": 5603753.0,
800
+ "reward": 0.890625,
801
+ "reward_std": 0.05418623387813568,
802
+ "rewards/reward_func/mean": 0.890625,
803
+ "rewards/reward_func/std": 0.05418623387813568,
804
+ "sampling/importance_sampling_ratio/max": 1.1075488299131393,
805
+ "sampling/importance_sampling_ratio/mean": 0.9244932755129411,
806
+ "sampling/importance_sampling_ratio/min": 0.660812682658434,
807
+ "sampling/sampling_logp_difference/max": 0.904436647228431,
808
+ "sampling/sampling_logp_difference/mean": 0.022255377695546485,
809
+ "step": 250,
810
+ "step_time": 10.151665401621722
811
+ }
812
+ ],
813
+ "logging_steps": 10,
814
+ "max_steps": 256,
815
+ "num_input_tokens_seen": 5750993,
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": 4,
832
+ "trial_name": null,
833
+ "trial_params": null
834
+ }
rl_grpo/control_4x_direct/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/rl3_direct_b/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
+ "28.self_attn.v_proj",
35
+ "language_model.layers.1.self_attn.k_proj",
36
+ "language_model.layers.22.self_attn.k_proj",
37
+ "language_model.layers.25.self_attn.v_proj",
38
+ "45.self_attn.k_proj",
39
+ "27.self_attn.v_proj",
40
+ "language_model.layers.9.self_attn.q_proj",
41
+ "language_model.layers.8.self_attn.q_proj",
42
+ "language_model.layers.23.self_attn.q_proj",
43
+ "31.self_attn.q_proj",
44
+ "language_model.layers.0.self_attn.k_proj",
45
+ "language_model.layers.23.self_attn.k_proj",
46
+ "27.self_attn.q_proj",
47
+ "34.self_attn.v_proj",
48
+ "language_model.layers.10.self_attn.v_proj",
49
+ "language_model.layers.10.self_attn.k_proj",
50
+ "language_model.layers.16.self_attn.k_proj",
51
+ "29.self_attn.q_proj",
52
+ "language_model.layers.19.self_attn.k_proj",
53
+ "28.self_attn.q_proj",
54
+ "47.self_attn.v_proj",
55
+ "43.self_attn.q_proj",
56
+ "language_model.layers.14.self_attn.v_proj",
57
+ "31.self_attn.v_proj",
58
+ "language_model.layers.4.self_attn.v_proj",
59
+ "language_model.layers.4.self_attn.k_proj",
60
+ "language_model.layers.17.self_attn.k_proj",
61
+ "33.self_attn.q_proj",
62
+ "30.self_attn.v_proj",
63
+ "gate_proj",
64
+ "language_model.layers.7.self_attn.k_proj",
65
+ "language_model.layers.18.self_attn.q_proj",
66
+ "language_model.layers.2.self_attn.v_proj",
67
+ "45.self_attn.v_proj",
68
+ "language_model.layers.26.self_attn.k_proj",
69
+ "29.self_attn.k_proj",
70
+ "o_proj",
71
+ "language_model.layers.0.self_attn.q_proj",
72
+ "language_model.layers.15.self_attn.v_proj",
73
+ "language_model.layers.13.self_attn.v_proj",
74
+ "language_model.layers.9.self_attn.v_proj",
75
+ "language_model.layers.16.self_attn.q_proj",
76
+ "language_model.layers.19.self_attn.q_proj",
77
+ "language_model.layers.24.self_attn.q_proj",
78
+ "language_model.layers.11.self_attn.q_proj",
79
+ "40.self_attn.k_proj",
80
+ "46.self_attn.q_proj",
81
+ "language_model.layers.7.self_attn.q_proj",
82
+ "36.self_attn.k_proj",
83
+ "language_model.layers.12.self_attn.q_proj",
84
+ "language_model.layers.21.self_attn.k_proj",
85
+ "41.self_attn.k_proj",
86
+ "40.self_attn.q_proj",
87
+ "32.self_attn.k_proj",
88
+ "language_model.layers.18.self_attn.k_proj",
89
+ "29.self_attn.v_proj",
90
+ "44.self_attn.k_proj",
91
+ "42.self_attn.k_proj",
92
+ "language_model.layers.25.self_attn.q_proj",
93
+ "33.self_attn.v_proj",
94
+ "28.self_attn.k_proj",
95
+ "language_model.layers.24.self_attn.v_proj",
96
+ "language_model.layers.5.self_attn.q_proj",
97
+ "language_model.layers.9.self_attn.k_proj",
98
+ "44.self_attn.v_proj",
99
+ "42.self_attn.v_proj",
100
+ "language_model.layers.11.self_attn.k_proj",
101
+ "32.self_attn.q_proj",
102
+ "language_model.layers.20.self_attn.k_proj",
103
+ "language_model.layers.26.self_attn.q_proj",
104
+ "language_model.layers.14.self_attn.k_proj",
105
+ "30.self_attn.q_proj",
106
+ "27.self_attn.k_proj",
107
+ "40.self_attn.v_proj",
108
+ "37.self_attn.v_proj",
109
+ "31.self_attn.k_proj",
110
+ "language_model.layers.2.self_attn.q_proj",
111
+ "language_model.layers.24.self_attn.k_proj",
112
+ "46.self_attn.k_proj",
113
+ "language_model.layers.12.self_attn.v_proj",
114
+ "41.self_attn.v_proj",
115
+ "language_model.layers.3.self_attn.v_proj",
116
+ "46.self_attn.v_proj",
117
+ "41.self_attn.q_proj",
118
+ "44.self_attn.q_proj",
119
+ "42.self_attn.q_proj",
120
+ "language_model.layers.16.self_attn.v_proj",
121
+ "language_model.layers.4.self_attn.q_proj",
122
+ "language_model.layers.19.self_attn.v_proj",
123
+ "up_proj",
124
+ "32.self_attn.v_proj",
125
+ "35.self_attn.k_proj",
126
+ "language_model.layers.20.self_attn.q_proj",
127
+ "38.self_attn.v_proj",
128
+ "language_model.layers.14.self_attn.q_proj",
129
+ "language_model.layers.20.self_attn.v_proj",
130
+ "43.self_attn.k_proj",
131
+ "language_model.layers.12.self_attn.k_proj",
132
+ "30.self_attn.k_proj",
133
+ "language_model.layers.1.self_attn.q_proj",
134
+ "34.self_attn.k_proj",
135
+ "39.self_attn.q_proj",
136
+ "37.self_attn.q_proj",
137
+ "language_model.layers.18.self_attn.v_proj",
138
+ "language_model.layers.26.self_attn.v_proj",
139
+ "language_model.layers.6.self_attn.q_proj",
140
+ "language_model.layers.21.self_attn.v_proj",
141
+ "language_model.layers.8.self_attn.k_proj",
142
+ "38.self_attn.q_proj",
143
+ "down_proj",
144
+ "language_model.layers.3.self_attn.q_proj",
145
+ "47.self_attn.q_proj",
146
+ "language_model.layers.22.self_attn.v_proj",
147
+ "33.self_attn.k_proj",
148
+ "36.self_attn.q_proj",
149
+ "language_model.layers.15.self_attn.q_proj",
150
+ "language_model.layers.0.self_attn.v_proj",
151
+ "language_model.layers.3.self_attn.k_proj",
152
+ "37.self_attn.k_proj",
153
+ "39.self_attn.v_proj",
154
+ "43.self_attn.v_proj",
155
+ "language_model.layers.2.self_attn.k_proj",
156
+ "language_model.layers.6.self_attn.v_proj",
157
+ "language_model.layers.13.self_attn.k_proj",
158
+ "36.self_attn.v_proj",
159
+ "language_model.layers.11.self_attn.v_proj",
160
+ "language_model.layers.17.self_attn.q_proj",
161
+ "language_model.layers.5.self_attn.v_proj",
162
+ "language_model.layers.10.self_attn.q_proj",
163
+ "38.self_attn.k_proj",
164
+ "35.self_attn.v_proj",
165
+ "language_model.layers.13.self_attn.q_proj",
166
+ "language_model.layers.5.self_attn.k_proj",
167
+ "language_model.layers.7.self_attn.v_proj",
168
+ "45.self_attn.q_proj",
169
+ "34.self_attn.q_proj",
170
+ "35.self_attn.q_proj",
171
+ "language_model.layers.21.self_attn.q_proj",
172
+ "language_model.layers.22.self_attn.q_proj",
173
+ "language_model.layers.1.self_attn.v_proj",
174
+ "language_model.layers.17.self_attn.v_proj",
175
+ "language_model.layers.8.self_attn.v_proj",
176
+ "39.self_attn.k_proj",
177
+ "language_model.layers.6.self_attn.k_proj",
178
+ "language_model.layers.23.self_attn.v_proj",
179
+ "47.self_attn.k_proj",
180
+ "language_model.layers.25.self_attn.k_proj",
181
+ "language_model.layers.15.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_direct/checkpoint-32/adapter_model.safetensors ADDED
@@ -0,0 +1,3 @@
 
 
 
 
1
+ version https://git-lfs.github.com/spec/v1
2
+ oid sha256:5dc46fd8508e67cd6f23ea2086b2331837ec3f4953ec8c260b227496a346072c
3
+ size 523864568
rl_grpo/control_4x_direct/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/rl3_direct_b/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
+ "28.self_attn.v_proj",
35
+ "language_model.layers.1.self_attn.k_proj",
36
+ "language_model.layers.22.self_attn.k_proj",
37
+ "language_model.layers.25.self_attn.v_proj",
38
+ "45.self_attn.k_proj",
39
+ "27.self_attn.v_proj",
40
+ "language_model.layers.9.self_attn.q_proj",
41
+ "language_model.layers.8.self_attn.q_proj",
42
+ "language_model.layers.23.self_attn.q_proj",
43
+ "31.self_attn.q_proj",
44
+ "language_model.layers.0.self_attn.k_proj",
45
+ "language_model.layers.23.self_attn.k_proj",
46
+ "27.self_attn.q_proj",
47
+ "34.self_attn.v_proj",
48
+ "language_model.layers.10.self_attn.v_proj",
49
+ "language_model.layers.10.self_attn.k_proj",
50
+ "language_model.layers.16.self_attn.k_proj",
51
+ "29.self_attn.q_proj",
52
+ "language_model.layers.19.self_attn.k_proj",
53
+ "28.self_attn.q_proj",
54
+ "47.self_attn.v_proj",
55
+ "43.self_attn.q_proj",
56
+ "language_model.layers.14.self_attn.v_proj",
57
+ "31.self_attn.v_proj",
58
+ "language_model.layers.4.self_attn.v_proj",
59
+ "language_model.layers.4.self_attn.k_proj",
60
+ "language_model.layers.17.self_attn.k_proj",
61
+ "33.self_attn.q_proj",
62
+ "30.self_attn.v_proj",
63
+ "gate_proj",
64
+ "language_model.layers.7.self_attn.k_proj",
65
+ "language_model.layers.18.self_attn.q_proj",
66
+ "language_model.layers.2.self_attn.v_proj",
67
+ "45.self_attn.v_proj",
68
+ "language_model.layers.26.self_attn.k_proj",
69
+ "29.self_attn.k_proj",
70
+ "o_proj",
71
+ "language_model.layers.0.self_attn.q_proj",
72
+ "language_model.layers.15.self_attn.v_proj",
73
+ "language_model.layers.13.self_attn.v_proj",
74
+ "language_model.layers.9.self_attn.v_proj",
75
+ "language_model.layers.16.self_attn.q_proj",
76
+ "language_model.layers.19.self_attn.q_proj",
77
+ "language_model.layers.24.self_attn.q_proj",
78
+ "language_model.layers.11.self_attn.q_proj",
79
+ "40.self_attn.k_proj",
80
+ "46.self_attn.q_proj",
81
+ "language_model.layers.7.self_attn.q_proj",
82
+ "36.self_attn.k_proj",
83
+ "language_model.layers.12.self_attn.q_proj",
84
+ "language_model.layers.21.self_attn.k_proj",
85
+ "41.self_attn.k_proj",
86
+ "40.self_attn.q_proj",
87
+ "32.self_attn.k_proj",
88
+ "language_model.layers.18.self_attn.k_proj",
89
+ "29.self_attn.v_proj",
90
+ "44.self_attn.k_proj",
91
+ "42.self_attn.k_proj",
92
+ "language_model.layers.25.self_attn.q_proj",
93
+ "33.self_attn.v_proj",
94
+ "28.self_attn.k_proj",
95
+ "language_model.layers.24.self_attn.v_proj",
96
+ "language_model.layers.5.self_attn.q_proj",
97
+ "language_model.layers.9.self_attn.k_proj",
98
+ "44.self_attn.v_proj",
99
+ "42.self_attn.v_proj",
100
+ "language_model.layers.11.self_attn.k_proj",
101
+ "32.self_attn.q_proj",
102
+ "language_model.layers.20.self_attn.k_proj",
103
+ "language_model.layers.26.self_attn.q_proj",
104
+ "language_model.layers.14.self_attn.k_proj",
105
+ "30.self_attn.q_proj",
106
+ "27.self_attn.k_proj",
107
+ "40.self_attn.v_proj",
108
+ "37.self_attn.v_proj",
109
+ "31.self_attn.k_proj",
110
+ "language_model.layers.2.self_attn.q_proj",
111
+ "language_model.layers.24.self_attn.k_proj",
112
+ "46.self_attn.k_proj",
113
+ "language_model.layers.12.self_attn.v_proj",
114
+ "41.self_attn.v_proj",
115
+ "language_model.layers.3.self_attn.v_proj",
116
+ "46.self_attn.v_proj",
117
+ "41.self_attn.q_proj",
118
+ "44.self_attn.q_proj",
119
+ "42.self_attn.q_proj",
120
+ "language_model.layers.16.self_attn.v_proj",
121
+ "language_model.layers.4.self_attn.q_proj",
122
+ "language_model.layers.19.self_attn.v_proj",
123
+ "up_proj",
124
+ "32.self_attn.v_proj",
125
+ "35.self_attn.k_proj",
126
+ "language_model.layers.20.self_attn.q_proj",
127
+ "38.self_attn.v_proj",
128
+ "language_model.layers.14.self_attn.q_proj",
129
+ "language_model.layers.20.self_attn.v_proj",
130
+ "43.self_attn.k_proj",
131
+ "language_model.layers.12.self_attn.k_proj",
132
+ "30.self_attn.k_proj",
133
+ "language_model.layers.1.self_attn.q_proj",
134
+ "34.self_attn.k_proj",
135
+ "39.self_attn.q_proj",
136
+ "37.self_attn.q_proj",
137
+ "language_model.layers.18.self_attn.v_proj",
138
+ "language_model.layers.26.self_attn.v_proj",
139
+ "language_model.layers.6.self_attn.q_proj",
140
+ "language_model.layers.21.self_attn.v_proj",
141
+ "language_model.layers.8.self_attn.k_proj",
142
+ "38.self_attn.q_proj",
143
+ "down_proj",
144
+ "language_model.layers.3.self_attn.q_proj",
145
+ "47.self_attn.q_proj",
146
+ "language_model.layers.22.self_attn.v_proj",
147
+ "33.self_attn.k_proj",
148
+ "36.self_attn.q_proj",
149
+ "language_model.layers.15.self_attn.q_proj",
150
+ "language_model.layers.0.self_attn.v_proj",
151
+ "language_model.layers.3.self_attn.k_proj",
152
+ "37.self_attn.k_proj",
153
+ "39.self_attn.v_proj",
154
+ "43.self_attn.v_proj",
155
+ "language_model.layers.2.self_attn.k_proj",
156
+ "language_model.layers.6.self_attn.v_proj",
157
+ "language_model.layers.13.self_attn.k_proj",
158
+ "36.self_attn.v_proj",
159
+ "language_model.layers.11.self_attn.v_proj",
160
+ "language_model.layers.17.self_attn.q_proj",
161
+ "language_model.layers.5.self_attn.v_proj",
162
+ "language_model.layers.10.self_attn.q_proj",
163
+ "38.self_attn.k_proj",
164
+ "35.self_attn.v_proj",
165
+ "language_model.layers.13.self_attn.q_proj",
166
+ "language_model.layers.5.self_attn.k_proj",
167
+ "language_model.layers.7.self_attn.v_proj",
168
+ "45.self_attn.q_proj",
169
+ "34.self_attn.q_proj",
170
+ "35.self_attn.q_proj",
171
+ "language_model.layers.21.self_attn.q_proj",
172
+ "language_model.layers.22.self_attn.q_proj",
173
+ "language_model.layers.1.self_attn.v_proj",
174
+ "language_model.layers.17.self_attn.v_proj",
175
+ "language_model.layers.8.self_attn.v_proj",
176
+ "39.self_attn.k_proj",
177
+ "language_model.layers.6.self_attn.k_proj",
178
+ "language_model.layers.23.self_attn.v_proj",
179
+ "47.self_attn.k_proj",
180
+ "language_model.layers.25.self_attn.k_proj",
181
+ "language_model.layers.15.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_direct/checkpoint-64/adapter_model.safetensors ADDED
@@ -0,0 +1,3 @@
 
 
 
 
1
+ version https://git-lfs.github.com/spec/v1
2
+ oid sha256:46ca7dbbced1b4602a9d6277e9584cc60e288fa2037a088fbb6c1ee328d6e690
3
+ size 523864568
rl_grpo/control_4x_direct/checkpoint.json ADDED
@@ -0,0 +1,132 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ {
2
+ "experiment": "scimt-train:dispatch-rl-control_4x_direct",
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/rl3_direct_b/data/direct/train.jsonl",
8
+ "dataset_meta": {
9
+ "adhoc": true
10
+ },
11
+ "stage": null,
12
+ "seed": 42,
13
+ "load_checkpoint_path": "/workspace/rl3_direct_b/parent",
14
+ "grpo": {
15
+ "episodes": 8192,
16
+ "group_size": 8,
17
+ "max_prompt_length": 3072,
18
+ "max_completion_length": 256,
19
+ "per_device_batch_size": 4,
20
+ "gradient_accumulation_steps": 8,
21
+ "steps_per_generation": null,
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": 0.45,
28
+ "vllm_max_model_len": 3328,
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_direct",
42
+ "resume_from_checkpoint": null,
43
+ "rollout_log_dir": "/workspace/rl3_direct_b/training/control_4x_direct/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_direct",
62
+ "pointer_file": "/workspace/rl3_direct_b/training/control_4x_direct/train/ckpt_dispatch-rl-control_4x_direct.txt",
63
+ "backend": "hf_grpo",
64
+ "sampler": "/workspace/rl3_direct_b/training/control_4x_direct/train/sampler",
65
+ "state": "/workspace/rl3_direct_b/training/control_4x_direct/train/trainer/checkpoint-256",
66
+ "model": "google/gemma-3-12b-pt",
67
+ "meta": {
68
+ "experiment": "scimt-train:dispatch-rl-control_4x_direct",
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/rl3_direct_b/data/direct/train.jsonl",
74
+ "dataset_meta": {
75
+ "adhoc": true
76
+ },
77
+ "stage": null,
78
+ "seed": 42,
79
+ "load_checkpoint_path": "/workspace/rl3_direct_b/parent",
80
+ "grpo": {
81
+ "episodes": 8192,
82
+ "group_size": 8,
83
+ "max_prompt_length": 3072,
84
+ "max_completion_length": 256,
85
+ "per_device_batch_size": 4,
86
+ "gradient_accumulation_steps": 8,
87
+ "steps_per_generation": null,
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": 0.45,
94
+ "vllm_max_model_len": 3328,
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_direct",
108
+ "resume_from_checkpoint": null,
109
+ "rollout_log_dir": "/workspace/rl3_direct_b/training/control_4x_direct/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_direct",
128
+ "pointer_file": "/workspace/rl3_direct_b/training/control_4x_direct/train/ckpt_dispatch-rl-control_4x_direct.txt"
129
+ },
130
+ "sampler_path": "/workspace/rl3_direct_b/training/control_4x_direct/train/sampler",
131
+ "state_path": "/workspace/rl3_direct_b/training/control_4x_direct/train/trainer/checkpoint-256"
132
+ }
rl_grpo/control_4x_direct/logs/raw_rollouts.rank-0.jsonl ADDED
The diff for this file is too large to render. See raw diff