Hugging Face's logo

aaroncool9
/
NanoJev

PyTorch
English
Chinese
nanojev
decision-model
probability-distributions
parallel-inference
navigation
Model card Files
xet
 Community
aaroncool9 C-Tianyu commited on
Commit
bf577dc
·
0 Parent(s):

Duplicate from C-Tianyu/NanoJev

Browse files

Co-authored-by: Tianyu Chen <C-Tianyu@users.noreply.huggingface.co>

This view is limited to 50 files because it contains too many changes.   See raw diff
Files changed (50) hide show
  1. .gitattributes +43 -0
  2. GAMES_MODEL_MANIFEST.json +842 -0
  3. MODEL_MANIFEST.json +383 -0
  4. README.md +224 -0
  5. backbone_config/config.json +63 -0
  6. best.safetensors +3 -0
  7. config.json +66 -0
  8. stage1/backbone_config/config.json +63 -0
  9. stage1/best.safetensors +3 -0
  10. stage1/config.json +66 -0
  11. stage1/tokenizer/chat_template.jinja +89 -0
  12. stage1/tokenizer/tokenizer.json +3 -0
  13. stage1/tokenizer/tokenizer_config.json +30 -0
  14. tokenizer/chat_template.jinja +89 -0
  15. tokenizer/tokenizer.json +3 -0
  16. tokenizer/tokenizer_config.json +30 -0
  17. variants/events_brier_seed17/backbone_config/config.json +63 -0
  18. variants/events_brier_seed17/best.safetensors +3 -0
  19. variants/events_brier_seed17/config.json +69 -0
  20. variants/events_brier_seed17/summary.json +136 -0
  21. variants/events_brier_seed17/target_audit.json +0 -0
  22. variants/events_brier_seed17/tokenizer/chat_template.jinja +89 -0
  23. variants/events_brier_seed17/tokenizer/tokenizer.json +3 -0
  24. variants/events_brier_seed17/tokenizer/tokenizer_config.json +30 -0
  25. variants/events_brier_seed17/train_log.json +964 -0
  26. variants/events_ce_seed17/backbone_config/config.json +63 -0
  27. variants/events_ce_seed17/best.safetensors +3 -0
  28. variants/events_ce_seed17/config.json +69 -0
  29. variants/events_ce_seed17/summary.json +136 -0
  30. variants/events_ce_seed17/target_audit.json +0 -0
  31. variants/events_ce_seed17/tokenizer/chat_template.jinja +89 -0
  32. variants/events_ce_seed17/tokenizer/tokenizer.json +3 -0
  33. variants/events_ce_seed17/tokenizer/tokenizer_config.json +30 -0
  34. variants/events_ce_seed17/train_log.json +964 -0
  35. variants/events_paired_seed17/backbone_config/config.json +63 -0
  36. variants/events_paired_seed17/best.safetensors +3 -0
  37. variants/events_paired_seed17/config.json +69 -0
  38. variants/events_paired_seed17/summary.json +136 -0
  39. variants/events_paired_seed17/target_audit.json +0 -0
  40. variants/events_paired_seed17/tokenizer/chat_template.jinja +89 -0
  41. variants/events_paired_seed17/tokenizer/tokenizer.json +3 -0
  42. variants/events_paired_seed17/tokenizer/tokenizer_config.json +30 -0
  43. variants/events_paired_seed17/train_log.json +964 -0
  44. variants/games_api_seed17/backbone_config/config.json +63 -0
  45. variants/games_api_seed17/best.safetensors +3 -0
  46. variants/games_api_seed17/config.json +74 -0
  47. variants/games_api_seed17/summary.json +136 -0
  48. variants/games_api_seed17/target_audit.json +0 -0
  49. variants/games_api_seed17/tokenizer/chat_template.jinja +89 -0
  50. variants/games_api_seed17/tokenizer/tokenizer.json +3 -0
.gitattributes ADDED
@@ -0,0 +1,43 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ *.7z filter=lfs diff=lfs merge=lfs -text
2
+ *.arrow filter=lfs diff=lfs merge=lfs -text
3
+ *.bin filter=lfs diff=lfs merge=lfs -text
4
+ *.bz2 filter=lfs diff=lfs merge=lfs -text
5
+ *.ckpt filter=lfs diff=lfs merge=lfs -text
6
+ *.ftz filter=lfs diff=lfs merge=lfs -text
7
+ *.gz filter=lfs diff=lfs merge=lfs -text
8
+ *.h5 filter=lfs diff=lfs merge=lfs -text
9
+ *.joblib filter=lfs diff=lfs merge=lfs -text
10
+ *.lfs.* filter=lfs diff=lfs merge=lfs -text
11
+ *.mlmodel filter=lfs diff=lfs merge=lfs -text
12
+ *.model filter=lfs diff=lfs merge=lfs -text
13
+ *.msgpack filter=lfs diff=lfs merge=lfs -text
14
+ *.npy filter=lfs diff=lfs merge=lfs -text
15
+ *.npz filter=lfs diff=lfs merge=lfs -text
16
+ *.onnx filter=lfs diff=lfs merge=lfs -text
17
+ *.ot filter=lfs diff=lfs merge=lfs -text
18
+ *.parquet filter=lfs diff=lfs merge=lfs -text
19
+ *.pb filter=lfs diff=lfs merge=lfs -text
20
+ *.pickle filter=lfs diff=lfs merge=lfs -text
21
+ *.pkl filter=lfs diff=lfs merge=lfs -text
22
+ *.pt filter=lfs diff=lfs merge=lfs -text
23
+ *.pth filter=lfs diff=lfs merge=lfs -text
24
+ *.rar filter=lfs diff=lfs merge=lfs -text
25
+ *.safetensors filter=lfs diff=lfs merge=lfs -text
26
+ saved_model/**/* filter=lfs diff=lfs merge=lfs -text
27
+ *.tar.* filter=lfs diff=lfs merge=lfs -text
28
+ *.tar filter=lfs diff=lfs merge=lfs -text
29
+ *.tflite filter=lfs diff=lfs merge=lfs -text
30
+ *.tgz filter=lfs diff=lfs merge=lfs -text
31
+ *.wasm filter=lfs diff=lfs merge=lfs -text
32
+ *.xz filter=lfs diff=lfs merge=lfs -text
33
+ *.zip filter=lfs diff=lfs merge=lfs -text
34
+ *.zst filter=lfs diff=lfs merge=lfs -text
35
+ *tfevents* filter=lfs diff=lfs merge=lfs -text
36
+ stage1/tokenizer/tokenizer.json filter=lfs diff=lfs merge=lfs -text
37
+ tokenizer/tokenizer.json filter=lfs diff=lfs merge=lfs -text
38
+ variants/events_brier_seed17/tokenizer/tokenizer.json filter=lfs diff=lfs merge=lfs -text
39
+ variants/events_ce_seed17/tokenizer/tokenizer.json filter=lfs diff=lfs merge=lfs -text
40
+ variants/events_paired_seed17/tokenizer/tokenizer.json filter=lfs diff=lfs merge=lfs -text
41
+ variants/games_api_seed17/tokenizer/tokenizer.json filter=lfs diff=lfs merge=lfs -text
42
+ variants/games_gold_seed17/tokenizer/tokenizer.json filter=lfs diff=lfs merge=lfs -text
43
+ variants/local_atomic_seed17/tokenizer/tokenizer.json filter=lfs diff=lfs merge=lfs -text
GAMES_MODEL_MANIFEST.json ADDED
@@ -0,0 +1,842 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ {
2
+ "schema": "nanojev-games-model-release-v1",
3
+ "model_repository": "C-Tianyu/NanoJev",
4
+ "dataset_repository": "C-Tianyu/NanoJev-Data",
5
+ "base_model": "Qwen/Qwen3-0.6B",
6
+ "base_model_revision": "c1899de289a04d12100db370d81485cdf75e47ca",
7
+ "weight_staging": "hard links to existing frozen weights; configuration and audit files are byte-for-byte copies; no symlinks",
8
+ "variants": [
9
+ {
10
+ "name": "local_atomic_seed17",
11
+ "path": "variants/local_atomic_seed17",
12
+ "source_checkpoint": "/home/rwang/nanojev_dev_v4_20260917/runs/local_atomic_seed17",
13
+ "weight_sha256": "2b06e4423861f47da7ca53ad9c23a26aebdbfe79de3a238b36c2954ae9c6abf0",
14
+ "best_step": 300,
15
+ "best_dev_target_ce": 0.48294000757494554,
16
+ "training_seconds": 242.69976669549942,
17
+ "objective": "gold_distribution",
18
+ "loss": "ce",
19
+ "data_source_directory": "data/local_maze_v1",
20
+ "data_files": [
21
+ {
22
+ "source_path": "data/local_maze_v1/train.jsonl",
23
+ "sha256": "502f2840f3e15ffac76554f35ef6c2e41c6e25f58f77b5ee3a2f1196cba0c41a",
24
+ "bytes": 754633
25
+ },
26
+ {
27
+ "source_path": "data/local_maze_v1/dev.jsonl",
28
+ "sha256": "bd7f9d4402112770f571210677e329e8673a38caa6bb798b744ee47bc793b9a5",
29
+ "bytes": 262359
30
+ },
31
+ {
32
+ "source_path": "data/local_maze_v1/calibration.jsonl",
33
+ "sha256": "94a233e71eaf3460c939bb0b4b147a0a951931001097070ea7ac8b4b78f39ace",
34
+ "bytes": 252234
35
+ },
36
+ {
37
+ "source_path": "data/local_maze_v1/test.jsonl",
38
+ "sha256": "be7783c395715f9fa2aae7c793c4e5b7c4b1dfe3fcb0880544b577e85b8f3b10",
39
+ "bytes": 232377
40
+ },
41
+ {
42
+ "source_path": "data/local_maze_v1/ood.jsonl",
43
+ "sha256": "e6772998a189880c780b6408c0e348c77324a28204e390767dec54f475d61f4e",
44
+ "bytes": 195310
45
+ }
46
+ ],
47
+ "files": [
48
+ {
49
+ "path": "variants/local_atomic_seed17/best.safetensors",
50
+ "bytes": 2385039280,
51
+ "sha256": "2b06e4423861f47da7ca53ad9c23a26aebdbfe79de3a238b36c2954ae9c6abf0"
52
+ },
53
+ {
54
+ "path": "variants/local_atomic_seed17/config.json",
55
+ "bytes": 2498,
56
+ "sha256": "3a6b014a77eaa3a79155a9c0c84ce2c005b5a5c66fdf88445067657ac9def176"
57
+ },
58
+ {
59
+ "path": "variants/local_atomic_seed17/train_log.json",
60
+ "bytes": 85584,
61
+ "sha256": "5e2a39713c30d714dddd8b7d989aecf1a8105fee24e74b42ef8943a1fd4dff09"
62
+ },
63
+ {
64
+ "path": "variants/local_atomic_seed17/summary.json",
65
+ "bytes": 4341,
66
+ "sha256": "3da659f939b0cc8d44614d9258f8660482db548a3f62e6f5c8604f15a9770980"
67
+ },
68
+ {
69
+ "path": "variants/local_atomic_seed17/target_audit.json",
70
+ "bytes": 701867,
71
+ "sha256": "ae1cde9eeedf6e84266205fba29b32bcb58489ca7da4f17704d74d2964d6cc79"
72
+ },
73
+ {
74
+ "path": "variants/local_atomic_seed17/tokenizer/chat_template.jinja",
75
+ "bytes": 4168,
76
+ "sha256": "a55ee1b1660128b7098723e0abcd92caa0788061051c62d51cbe87d9cf1974d8"
77
+ },
78
+ {
79
+ "path": "variants/local_atomic_seed17/tokenizer/tokenizer.json",
80
+ "bytes": 11422650,
81
+ "sha256": "be75606093db2094d7cd20f3c2f385c212750648bd6ea4fb2bf507a6a4c55506"
82
+ },
83
+ {
84
+ "path": "variants/local_atomic_seed17/tokenizer/tokenizer_config.json",
85
+ "bytes": 692,
86
+ "sha256": "1cc816812993bff176eb4f7495433b736f06fba9b6e7b05cac7b4a1780650c95"
87
+ },
88
+ {
89
+ "path": "variants/local_atomic_seed17/backbone_config/config.json",
90
+ "bytes": 1417,
91
+ "sha256": "30c011854471509747858ac07ffb8a4dab2bc6c04035b184729b69500c557c54"
92
+ }
93
+ ],
94
+ "initial_checkpoint_sha256": "fff62d1412685c1714eaa386acb603f9690371fb3cc8ad03dc41319302597c28",
95
+ "minimum_inference_patterns": [
96
+ "best.safetensors",
97
+ "config.json",
98
+ "tokenizer/*",
99
+ "backbone_config/*"
100
+ ]
101
+ },
102
+ {
103
+ "name": "games_gold_seed17",
104
+ "path": "variants/games_gold_seed17",
105
+ "source_checkpoint": "/home/rwang/nanojev_dev_v4_20260917/runs/games_gold_seed17",
106
+ "weight_sha256": "5a395c56fad2a9bd4fb664ef78e1e06c19102886d97dd56dfe2676e5896c7e3a",
107
+ "best_step": 100,
108
+ "best_dev_target_ce": 0.6079203053015539,
109
+ "training_seconds": 186.51977808214724,
110
+ "objective": "gold_distribution",
111
+ "loss": "ce",
112
+ "data_source_directory": "data/scaled_games_v4b/policy",
113
+ "data_files": [
114
+ {
115
+ "source_path": "data/scaled_games_v4b/policy/train.jsonl",
116
+ "sha256": "af6065315d383bce2f137b3095cc9912aafe1edf0a3ee7c119c9b7af4a0ac399",
117
+ "bytes": 1776280
118
+ },
119
+ {
120
+ "source_path": "data/scaled_games_v4b/policy/dev.jsonl",
121
+ "sha256": "5e34f86e60dffce37d9454c4ffb978c987ba8467de35c5f79dab99bdb24f3d6b",
122
+ "bytes": 601279
123
+ },
124
+ {
125
+ "source_path": "data/scaled_games_v4b/policy/calibration.jsonl",
126
+ "sha256": "c2b279ecd835068c20bc328c34bca970feee1e8afab67b1f7dea0c201f0d824f",
127
+ "bytes": 584303
128
+ },
129
+ {
130
+ "source_path": "data/scaled_games_v4b/policy/test.jsonl",
131
+ "sha256": "54e573515e5d1e2c55308763a857d9bb46f2c02b2cbd23e50a4544587a69501c",
132
+ "bytes": 569530
133
+ },
134
+ {
135
+ "source_path": "data/scaled_games_v4b/policy/ood.jsonl",
136
+ "sha256": "ac4ea92590a95f40cd2134ad64f450c369601745b36d91ebf5692d9f39ea3135",
137
+ "bytes": 375280
138
+ }
139
+ ],
140
+ "files": [
141
+ {
142
+ "path": "variants/games_gold_seed17/best.safetensors",
143
+ "bytes": 2385039280,
144
+ "sha256": "5a395c56fad2a9bd4fb664ef78e1e06c19102886d97dd56dfe2676e5896c7e3a"
145
+ },
146
+ {
147
+ "path": "variants/games_gold_seed17/config.json",
148
+ "bytes": 2562,
149
+ "sha256": "965c5905315315ff591d6c2eb581146b749104e65afdd463e3a3118f70baaaff"
150
+ },
151
+ {
152
+ "path": "variants/games_gold_seed17/train_log.json",
153
+ "bytes": 42887,
154
+ "sha256": "a1c2de04bf59f1da31ab4255be69a5b5960fde183b8a459d2cd8b1f7c618ae28"
155
+ },
156
+ {
157
+ "path": "variants/games_gold_seed17/summary.json",
158
+ "bytes": 4530,
159
+ "sha256": "f7de836505d40a4f779258155455fc1d3d8194e468cae269bcb942c13441ff5c"
160
+ },
161
+ {
162
+ "path": "variants/games_gold_seed17/target_audit.json",
163
+ "bytes": 1831844,
164
+ "sha256": "62e50d044875e9914b7edce5d1321c873eb923a31fefaf0a32d8a13e0e755b1b"
165
+ },
166
+ {
167
+ "path": "variants/games_gold_seed17/tokenizer/chat_template.jinja",
168
+ "bytes": 4168,
169
+ "sha256": "a55ee1b1660128b7098723e0abcd92caa0788061051c62d51cbe87d9cf1974d8"
170
+ },
171
+ {
172
+ "path": "variants/games_gold_seed17/tokenizer/tokenizer.json",
173
+ "bytes": 11422650,
174
+ "sha256": "be75606093db2094d7cd20f3c2f385c212750648bd6ea4fb2bf507a6a4c55506"
175
+ },
176
+ {
177
+ "path": "variants/games_gold_seed17/tokenizer/tokenizer_config.json",
178
+ "bytes": 692,
179
+ "sha256": "1cc816812993bff176eb4f7495433b736f06fba9b6e7b05cac7b4a1780650c95"
180
+ },
181
+ {
182
+ "path": "variants/games_gold_seed17/backbone_config/config.json",
183
+ "bytes": 1417,
184
+ "sha256": "30c011854471509747858ac07ffb8a4dab2bc6c04035b184729b69500c557c54"
185
+ }
186
+ ],
187
+ "initial_checkpoint_sha256": "fff62d1412685c1714eaa386acb603f9690371fb3cc8ad03dc41319302597c28",
188
+ "minimum_inference_patterns": [
189
+ "best.safetensors",
190
+ "config.json",
191
+ "tokenizer/*",
192
+ "backbone_config/*"
193
+ ]
194
+ },
195
+ {
196
+ "name": "games_api_seed17",
197
+ "path": "variants/games_api_seed17",
198
+ "source_checkpoint": "/home/rwang/nanojev_dev_v4_20260917/runs/games_api_seed17",
199
+ "weight_sha256": "45829fc896391847220b67dcf6fc0702427d2f184be384649bcf9343c1552477",
200
+ "best_step": 150,
201
+ "best_dev_target_ce": 0.7130707709730839,
202
+ "training_seconds": 182.26331992074847,
203
+ "objective": "teacher",
204
+ "loss": "ce",
205
+ "data_source_directory": "data/scaled_games_labeled_v4",
206
+ "data_files": [
207
+ {
208
+ "source_path": "data/scaled_games_labeled_v4/train.jsonl",
209
+ "sha256": "e21d05f1c34253687f4e7b78974503e7c05a7e145e4cbfc306246938d9f6317c",
210
+ "bytes": 1542530
211
+ },
212
+ {
213
+ "source_path": "data/scaled_games_labeled_v4/dev.jsonl",
214
+ "sha256": "fd7780fd6bf94263f5653cea8b3575925447981602dba61d0c7ff96b8433ea88",
215
+ "bytes": 524030
216
+ },
217
+ {
218
+ "source_path": "data/scaled_games_labeled_v4/calibration.jsonl",
219
+ "sha256": "9632b050ee1069c4393e275fc878c2681f9c36a501c4c6584df19b3fb88d76fd",
220
+ "bytes": 507336
221
+ },
222
+ {
223
+ "source_path": "data/scaled_games_labeled_v4/test.jsonl",
224
+ "sha256": "82b62b39f56bcd703e6e7c911aafb20203ff7b48ebf4f4f4c3444694e4823fa0",
225
+ "bytes": 494657
226
+ },
227
+ {
228
+ "source_path": "data/scaled_games_labeled_v4/ood.jsonl",
229
+ "sha256": "568c11c100126b3848fb29c9cd3d4b8dc22be42fe2df43a3bb3cf8e7b0128ae1",
230
+ "bytes": 348817
231
+ }
232
+ ],
233
+ "files": [
234
+ {
235
+ "path": "variants/games_api_seed17/best.safetensors",
236
+ "bytes": 2385039280,
237
+ "sha256": "45829fc896391847220b67dcf6fc0702427d2f184be384649bcf9343c1552477"
238
+ },
239
+ {
240
+ "path": "variants/games_api_seed17/config.json",
241
+ "bytes": 2693,
242
+ "sha256": "509c449f3088e167f95ea284370763489ca182e787d7765efbcffc377b767931"
243
+ },
244
+ {
245
+ "path": "variants/games_api_seed17/train_log.json",
246
+ "bytes": 42439,
247
+ "sha256": "aca1693ee0a2c1c1c574416329e28a089adaa93014a071b09ac4350b62e12050"
248
+ },
249
+ {
250
+ "path": "variants/games_api_seed17/summary.json",
251
+ "bytes": 3996,
252
+ "sha256": "304fde47ce4cef20bd06ed088c55f4a8bc2c8da3f68e3f5a9520a5bdbf14dbdb"
253
+ },
254
+ {
255
+ "path": "variants/games_api_seed17/target_audit.json",
256
+ "bytes": 2012764,
257
+ "sha256": "28a1666127b407bbfdaf8bcf99ca386c66a814a37ef78221bda12cdce9f8ab38"
258
+ },
259
+ {
260
+ "path": "variants/games_api_seed17/tokenizer/chat_template.jinja",
261
+ "bytes": 4168,
262
+ "sha256": "a55ee1b1660128b7098723e0abcd92caa0788061051c62d51cbe87d9cf1974d8"
263
+ },
264
+ {
265
+ "path": "variants/games_api_seed17/tokenizer/tokenizer.json",
266
+ "bytes": 11422650,
267
+ "sha256": "be75606093db2094d7cd20f3c2f385c212750648bd6ea4fb2bf507a6a4c55506"
268
+ },
269
+ {
270
+ "path": "variants/games_api_seed17/tokenizer/tokenizer_config.json",
271
+ "bytes": 692,
272
+ "sha256": "1cc816812993bff176eb4f7495433b736f06fba9b6e7b05cac7b4a1780650c95"
273
+ },
274
+ {
275
+ "path": "variants/games_api_seed17/backbone_config/config.json",
276
+ "bytes": 1417,
277
+ "sha256": "30c011854471509747858ac07ffb8a4dab2bc6c04035b184729b69500c557c54"
278
+ }
279
+ ],
280
+ "initial_checkpoint_sha256": "fff62d1412685c1714eaa386acb603f9690371fb3cc8ad03dc41319302597c28",
281
+ "minimum_inference_patterns": [
282
+ "best.safetensors",
283
+ "config.json",
284
+ "tokenizer/*",
285
+ "backbone_config/*"
286
+ ]
287
+ },
288
+ {
289
+ "name": "events_ce_seed17",
290
+ "path": "variants/events_ce_seed17",
291
+ "source_checkpoint": "/home/rwang/nanojev_dev_v4_20260917/runs/events_ce_seed17",
292
+ "weight_sha256": "582c01139e29807b5f2690335271cf5647583cbfc3dd963d5967792bb3963992",
293
+ "best_step": 100,
294
+ "best_dev_target_ce": 0.47277030621569244,
295
+ "training_seconds": 96.74807532411069,
296
+ "objective": "observed_outcome",
297
+ "loss": "ce",
298
+ "data_source_directory": "data/scaled_games_v4/events",
299
+ "data_files": [
300
+ {
301
+ "source_path": "data/scaled_games_v4/events/train.jsonl",
302
+ "sha256": "4cebf698e4b71dc37d4d0e985b6577dd565ac370ac82b6f47f375aa6af69f8c9",
303
+ "bytes": 4061794
304
+ },
305
+ {
306
+ "source_path": "data/scaled_games_v4/events/dev.jsonl",
307
+ "sha256": "0c352ee1064a5fb5d2aacad59cbea57afb897886d3d09e5ca108c866a5f893db",
308
+ "bytes": 1403407
309
+ },
310
+ {
311
+ "source_path": "data/scaled_games_v4/events/calibration.jsonl",
312
+ "sha256": "4a8a8a2b81c3c46df6bb4c8690e97374dd0165cbef40566bb7f87c1f425d5562",
313
+ "bytes": 1345231
314
+ },
315
+ {
316
+ "source_path": "data/scaled_games_v4/events/test.jsonl",
317
+ "sha256": "884bec71891668199925809d8708347f56f7bf06e8363e4a436347967b1423e3",
318
+ "bytes": 1299634
319
+ },
320
+ {
321
+ "source_path": "data/scaled_games_v4/events/ood.jsonl",
322
+ "sha256": "3c00d845524634ca3b34559761cdaf0c52b7b7633d5517d1e7ee7f31697cea48",
323
+ "bytes": 1154862
324
+ }
325
+ ],
326
+ "files": [
327
+ {
328
+ "path": "variants/events_ce_seed17/best.safetensors",
329
+ "bytes": 2385039280,
330
+ "sha256": "582c01139e29807b5f2690335271cf5647583cbfc3dd963d5967792bb3963992"
331
+ },
332
+ {
333
+ "path": "variants/events_ce_seed17/config.json",
334
+ "bytes": 2555,
335
+ "sha256": "52f0309ceeefe51d8cadda95b6f0985978b2d523d758194f682932892d82a566"
336
+ },
337
+ {
338
+ "path": "variants/events_ce_seed17/train_log.json",
339
+ "bytes": 28195,
340
+ "sha256": "c5e20da3ea98546ac71a9d2a264b53d52bcfb3387a7cd5a57d65c7cc03c9623e"
341
+ },
342
+ {
343
+ "path": "variants/events_ce_seed17/summary.json",
344
+ "bytes": 3840,
345
+ "sha256": "b5f1a6592da085ea88ca7676fbaf11b39874b097cdc70a3424c992010b27c3a5"
346
+ },
347
+ {
348
+ "path": "variants/events_ce_seed17/target_audit.json",
349
+ "bytes": 1334523,
350
+ "sha256": "7a9b2635cf87607e9f2a47fa1c2e037f1e5b13c18bcbbc5898cff09796222181"
351
+ },
352
+ {
353
+ "path": "variants/events_ce_seed17/tokenizer/chat_template.jinja",
354
+ "bytes": 4168,
355
+ "sha256": "a55ee1b1660128b7098723e0abcd92caa0788061051c62d51cbe87d9cf1974d8"
356
+ },
357
+ {
358
+ "path": "variants/events_ce_seed17/tokenizer/tokenizer.json",
359
+ "bytes": 11422650,
360
+ "sha256": "be75606093db2094d7cd20f3c2f385c212750648bd6ea4fb2bf507a6a4c55506"
361
+ },
362
+ {
363
+ "path": "variants/events_ce_seed17/tokenizer/tokenizer_config.json",
364
+ "bytes": 692,
365
+ "sha256": "1cc816812993bff176eb4f7495433b736f06fba9b6e7b05cac7b4a1780650c95"
366
+ },
367
+ {
368
+ "path": "variants/events_ce_seed17/backbone_config/config.json",
369
+ "bytes": 1417,
370
+ "sha256": "30c011854471509747858ac07ffb8a4dab2bc6c04035b184729b69500c557c54"
371
+ }
372
+ ],
373
+ "initial_checkpoint_sha256": "fff62d1412685c1714eaa386acb603f9690371fb3cc8ad03dc41319302597c28",
374
+ "minimum_inference_patterns": [
375
+ "best.safetensors",
376
+ "config.json",
377
+ "tokenizer/*",
378
+ "backbone_config/*"
379
+ ]
380
+ },
381
+ {
382
+ "name": "events_brier_seed17",
383
+ "path": "variants/events_brier_seed17",
384
+ "source_checkpoint": "/home/rwang/nanojev_dev_v4_20260917/runs/events_brier_seed17",
385
+ "weight_sha256": "0a8d8b14c97743eed4310d461a7ae5cff0400cb87eaa6dc1f8c5dc3c59988651",
386
+ "best_step": 100,
387
+ "best_dev_target_ce": 0.4421807345122086,
388
+ "training_seconds": 94.21064267028123,
389
+ "objective": "observed_outcome",
390
+ "loss": "brier",
391
+ "data_source_directory": "data/scaled_games_v4/events",
392
+ "data_files": [
393
+ {
394
+ "source_path": "data/scaled_games_v4/events/train.jsonl",
395
+ "sha256": "4cebf698e4b71dc37d4d0e985b6577dd565ac370ac82b6f47f375aa6af69f8c9",
396
+ "bytes": 4061794
397
+ },
398
+ {
399
+ "source_path": "data/scaled_games_v4/events/dev.jsonl",
400
+ "sha256": "0c352ee1064a5fb5d2aacad59cbea57afb897886d3d09e5ca108c866a5f893db",
401
+ "bytes": 1403407
402
+ },
403
+ {
404
+ "source_path": "data/scaled_games_v4/events/calibration.jsonl",
405
+ "sha256": "4a8a8a2b81c3c46df6bb4c8690e97374dd0165cbef40566bb7f87c1f425d5562",
406
+ "bytes": 1345231
407
+ },
408
+ {
409
+ "source_path": "data/scaled_games_v4/events/test.jsonl",
410
+ "sha256": "884bec71891668199925809d8708347f56f7bf06e8363e4a436347967b1423e3",
411
+ "bytes": 1299634
412
+ },
413
+ {
414
+ "source_path": "data/scaled_games_v4/events/ood.jsonl",
415
+ "sha256": "3c00d845524634ca3b34559761cdaf0c52b7b7633d5517d1e7ee7f31697cea48",
416
+ "bytes": 1154862
417
+ }
418
+ ],
419
+ "files": [
420
+ {
421
+ "path": "variants/events_brier_seed17/best.safetensors",
422
+ "bytes": 2385039280,
423
+ "sha256": "0a8d8b14c97743eed4310d461a7ae5cff0400cb87eaa6dc1f8c5dc3c59988651"
424
+ },
425
+ {
426
+ "path": "variants/events_brier_seed17/config.json",
427
+ "bytes": 2561,
428
+ "sha256": "bca3a9b31a8f528bc144433f1b2f4f9f27f78f50ceb5957c99949e246c850ac2"
429
+ },
430
+ {
431
+ "path": "variants/events_brier_seed17/train_log.json",
432
+ "bytes": 28219,
433
+ "sha256": "e09fdb39c212589b56af8f5cbae6f51fa1d1f91de73a63636e8661a02f2312f9"
434
+ },
435
+ {
436
+ "path": "variants/events_brier_seed17/summary.json",
437
+ "bytes": 3829,
438
+ "sha256": "daaf0e3739fb244368662944e26b939278a31beac7c1c538ce3e72aa70536def"
439
+ },
440
+ {
441
+ "path": "variants/events_brier_seed17/target_audit.json",
442
+ "bytes": 1334523,
443
+ "sha256": "7a9b2635cf87607e9f2a47fa1c2e037f1e5b13c18bcbbc5898cff09796222181"
444
+ },
445
+ {
446
+ "path": "variants/events_brier_seed17/tokenizer/chat_template.jinja",
447
+ "bytes": 4168,
448
+ "sha256": "a55ee1b1660128b7098723e0abcd92caa0788061051c62d51cbe87d9cf1974d8"
449
+ },
450
+ {
451
+ "path": "variants/events_brier_seed17/tokenizer/tokenizer.json",
452
+ "bytes": 11422650,
453
+ "sha256": "be75606093db2094d7cd20f3c2f385c212750648bd6ea4fb2bf507a6a4c55506"
454
+ },
455
+ {
456
+ "path": "variants/events_brier_seed17/tokenizer/tokenizer_config.json",
457
+ "bytes": 692,
458
+ "sha256": "1cc816812993bff176eb4f7495433b736f06fba9b6e7b05cac7b4a1780650c95"
459
+ },
460
+ {
461
+ "path": "variants/events_brier_seed17/backbone_config/config.json",
462
+ "bytes": 1417,
463
+ "sha256": "30c011854471509747858ac07ffb8a4dab2bc6c04035b184729b69500c557c54"
464
+ }
465
+ ],
466
+ "initial_checkpoint_sha256": "fff62d1412685c1714eaa386acb603f9690371fb3cc8ad03dc41319302597c28",
467
+ "minimum_inference_patterns": [
468
+ "best.safetensors",
469
+ "config.json",
470
+ "tokenizer/*",
471
+ "backbone_config/*"
472
+ ]
473
+ },
474
+ {
475
+ "name": "events_paired_seed17",
476
+ "path": "variants/events_paired_seed17",
477
+ "source_checkpoint": "/home/rwang/nanojev_dev_v4_20260917/runs/events_paired_seed17",
478
+ "weight_sha256": "bbefcf829b9e36253ef46ab076f7d8fca00ffb33762e8358dbe780b3605bb777",
479
+ "best_step": 100,
480
+ "best_dev_target_ce": 0.45392147839038,
481
+ "training_seconds": 100.81704722624272,
482
+ "objective": "observed_outcome",
483
+ "loss": "paired_brier_pg",
484
+ "data_source_directory": "data/scaled_games_v4/events",
485
+ "data_files": [
486
+ {
487
+ "source_path": "data/scaled_games_v4/events/train.jsonl",
488
+ "sha256": "4cebf698e4b71dc37d4d0e985b6577dd565ac370ac82b6f47f375aa6af69f8c9",
489
+ "bytes": 4061794
490
+ },
491
+ {
492
+ "source_path": "data/scaled_games_v4/events/dev.jsonl",
493
+ "sha256": "0c352ee1064a5fb5d2aacad59cbea57afb897886d3d09e5ca108c866a5f893db",
494
+ "bytes": 1403407
495
+ },
496
+ {
497
+ "source_path": "data/scaled_games_v4/events/calibration.jsonl",
498
+ "sha256": "4a8a8a2b81c3c46df6bb4c8690e97374dd0165cbef40566bb7f87c1f425d5562",
499
+ "bytes": 1345231
500
+ },
501
+ {
502
+ "source_path": "data/scaled_games_v4/events/test.jsonl",
503
+ "sha256": "884bec71891668199925809d8708347f56f7bf06e8363e4a436347967b1423e3",
504
+ "bytes": 1299634
505
+ },
506
+ {
507
+ "source_path": "data/scaled_games_v4/events/ood.jsonl",
508
+ "sha256": "3c00d845524634ca3b34559761cdaf0c52b7b7633d5517d1e7ee7f31697cea48",
509
+ "bytes": 1154862
510
+ }
511
+ ],
512
+ "files": [
513
+ {
514
+ "path": "variants/events_paired_seed17/best.safetensors",
515
+ "bytes": 2385039280,
516
+ "sha256": "bbefcf829b9e36253ef46ab076f7d8fca00ffb33762e8358dbe780b3605bb777"
517
+ },
518
+ {
519
+ "path": "variants/events_paired_seed17/config.json",
520
+ "bytes": 2671,
521
+ "sha256": "9bdffa977eb28e4692ba19f53b05a7ad9afd9f37665848f9db24de4e829f1acb"
522
+ },
523
+ {
524
+ "path": "variants/events_paired_seed17/train_log.json",
525
+ "bytes": 28356,
526
+ "sha256": "33f337e289dab3afce8199fc7b72a6c766634986f49276ba8581ea2210b673cf"
527
+ },
528
+ {
529
+ "path": "variants/events_paired_seed17/summary.json",
530
+ "bytes": 3817,
531
+ "sha256": "82de7efe05e55bedc75e5b50b5b9ca867aa1f100cb4c4430b9b67524c6e82de1"
532
+ },
533
+ {
534
+ "path": "variants/events_paired_seed17/target_audit.json",
535
+ "bytes": 1334523,
536
+ "sha256": "7a9b2635cf87607e9f2a47fa1c2e037f1e5b13c18bcbbc5898cff09796222181"
537
+ },
538
+ {
539
+ "path": "variants/events_paired_seed17/tokenizer/chat_template.jinja",
540
+ "bytes": 4168,
541
+ "sha256": "a55ee1b1660128b7098723e0abcd92caa0788061051c62d51cbe87d9cf1974d8"
542
+ },
543
+ {
544
+ "path": "variants/events_paired_seed17/tokenizer/tokenizer.json",
545
+ "bytes": 11422650,
546
+ "sha256": "be75606093db2094d7cd20f3c2f385c212750648bd6ea4fb2bf507a6a4c55506"
547
+ },
548
+ {
549
+ "path": "variants/events_paired_seed17/tokenizer/tokenizer_config.json",
550
+ "bytes": 692,
551
+ "sha256": "1cc816812993bff176eb4f7495433b736f06fba9b6e7b05cac7b4a1780650c95"
552
+ },
553
+ {
554
+ "path": "variants/events_paired_seed17/backbone_config/config.json",
555
+ "bytes": 1417,
556
+ "sha256": "30c011854471509747858ac07ffb8a4dab2bc6c04035b184729b69500c557c54"
557
+ }
558
+ ],
559
+ "initial_checkpoint_sha256": "fff62d1412685c1714eaa386acb603f9690371fb3cc8ad03dc41319302597c28",
560
+ "minimum_inference_patterns": [
561
+ "best.safetensors",
562
+ "config.json",
563
+ "tokenizer/*",
564
+ "backbone_config/*"
565
+ ]
566
+ }
567
+ ],
568
+ "files": [
569
+ {
570
+ "path": "variants/local_atomic_seed17/best.safetensors",
571
+ "bytes": 2385039280,
572
+ "sha256": "2b06e4423861f47da7ca53ad9c23a26aebdbfe79de3a238b36c2954ae9c6abf0"
573
+ },
574
+ {
575
+ "path": "variants/local_atomic_seed17/config.json",
576
+ "bytes": 2498,
577
+ "sha256": "3a6b014a77eaa3a79155a9c0c84ce2c005b5a5c66fdf88445067657ac9def176"
578
+ },
579
+ {
580
+ "path": "variants/local_atomic_seed17/train_log.json",
581
+ "bytes": 85584,
582
+ "sha256": "5e2a39713c30d714dddd8b7d989aecf1a8105fee24e74b42ef8943a1fd4dff09"
583
+ },
584
+ {
585
+ "path": "variants/local_atomic_seed17/summary.json",
586
+ "bytes": 4341,
587
+ "sha256": "3da659f939b0cc8d44614d9258f8660482db548a3f62e6f5c8604f15a9770980"
588
+ },
589
+ {
590
+ "path": "variants/local_atomic_seed17/target_audit.json",
591
+ "bytes": 701867,
592
+ "sha256": "ae1cde9eeedf6e84266205fba29b32bcb58489ca7da4f17704d74d2964d6cc79"
593
+ },
594
+ {
595
+ "path": "variants/local_atomic_seed17/tokenizer/chat_template.jinja",
596
+ "bytes": 4168,
597
+ "sha256": "a55ee1b1660128b7098723e0abcd92caa0788061051c62d51cbe87d9cf1974d8"
598
+ },
599
+ {
600
+ "path": "variants/local_atomic_seed17/tokenizer/tokenizer.json",
601
+ "bytes": 11422650,
602
+ "sha256": "be75606093db2094d7cd20f3c2f385c212750648bd6ea4fb2bf507a6a4c55506"
603
+ },
604
+ {
605
+ "path": "variants/local_atomic_seed17/tokenizer/tokenizer_config.json",
606
+ "bytes": 692,
607
+ "sha256": "1cc816812993bff176eb4f7495433b736f06fba9b6e7b05cac7b4a1780650c95"
608
+ },
609
+ {
610
+ "path": "variants/local_atomic_seed17/backbone_config/config.json",
611
+ "bytes": 1417,
612
+ "sha256": "30c011854471509747858ac07ffb8a4dab2bc6c04035b184729b69500c557c54"
613
+ },
614
+ {
615
+ "path": "variants/games_gold_seed17/best.safetensors",
616
+ "bytes": 2385039280,
617
+ "sha256": "5a395c56fad2a9bd4fb664ef78e1e06c19102886d97dd56dfe2676e5896c7e3a"
618
+ },
619
+ {
620
+ "path": "variants/games_gold_seed17/config.json",
621
+ "bytes": 2562,
622
+ "sha256": "965c5905315315ff591d6c2eb581146b749104e65afdd463e3a3118f70baaaff"
623
+ },
624
+ {
625
+ "path": "variants/games_gold_seed17/train_log.json",
626
+ "bytes": 42887,
627
+ "sha256": "a1c2de04bf59f1da31ab4255be69a5b5960fde183b8a459d2cd8b1f7c618ae28"
628
+ },
629
+ {
630
+ "path": "variants/games_gold_seed17/summary.json",
631
+ "bytes": 4530,
632
+ "sha256": "f7de836505d40a4f779258155455fc1d3d8194e468cae269bcb942c13441ff5c"
633
+ },
634
+ {
635
+ "path": "variants/games_gold_seed17/target_audit.json",
636
+ "bytes": 1831844,
637
+ "sha256": "62e50d044875e9914b7edce5d1321c873eb923a31fefaf0a32d8a13e0e755b1b"
638
+ },
639
+ {
640
+ "path": "variants/games_gold_seed17/tokenizer/chat_template.jinja",
641
+ "bytes": 4168,
642
+ "sha256": "a55ee1b1660128b7098723e0abcd92caa0788061051c62d51cbe87d9cf1974d8"
643
+ },
644
+ {
645
+ "path": "variants/games_gold_seed17/tokenizer/tokenizer.json",
646
+ "bytes": 11422650,
647
+ "sha256": "be75606093db2094d7cd20f3c2f385c212750648bd6ea4fb2bf507a6a4c55506"
648
+ },
649
+ {
650
+ "path": "variants/games_gold_seed17/tokenizer/tokenizer_config.json",
651
+ "bytes": 692,
652
+ "sha256": "1cc816812993bff176eb4f7495433b736f06fba9b6e7b05cac7b4a1780650c95"
653
+ },
654
+ {
655
+ "path": "variants/games_gold_seed17/backbone_config/config.json",
656
+ "bytes": 1417,
657
+ "sha256": "30c011854471509747858ac07ffb8a4dab2bc6c04035b184729b69500c557c54"
658
+ },
659
+ {
660
+ "path": "variants/games_api_seed17/best.safetensors",
661
+ "bytes": 2385039280,
662
+ "sha256": "45829fc896391847220b67dcf6fc0702427d2f184be384649bcf9343c1552477"
663
+ },
664
+ {
665
+ "path": "variants/games_api_seed17/config.json",
666
+ "bytes": 2693,
667
+ "sha256": "509c449f3088e167f95ea284370763489ca182e787d7765efbcffc377b767931"
668
+ },
669
+ {
670
+ "path": "variants/games_api_seed17/train_log.json",
671
+ "bytes": 42439,
672
+ "sha256": "aca1693ee0a2c1c1c574416329e28a089adaa93014a071b09ac4350b62e12050"
673
+ },
674
+ {
675
+ "path": "variants/games_api_seed17/summary.json",
676
+ "bytes": 3996,
677
+ "sha256": "304fde47ce4cef20bd06ed088c55f4a8bc2c8da3f68e3f5a9520a5bdbf14dbdb"
678
+ },
679
+ {
680
+ "path": "variants/games_api_seed17/target_audit.json",
681
+ "bytes": 2012764,
682
+ "sha256": "28a1666127b407bbfdaf8bcf99ca386c66a814a37ef78221bda12cdce9f8ab38"
683
+ },
684
+ {
685
+ "path": "variants/games_api_seed17/tokenizer/chat_template.jinja",
686
+ "bytes": 4168,
687
+ "sha256": "a55ee1b1660128b7098723e0abcd92caa0788061051c62d51cbe87d9cf1974d8"
688
+ },
689
+ {
690
+ "path": "variants/games_api_seed17/tokenizer/tokenizer.json",
691
+ "bytes": 11422650,
692
+ "sha256": "be75606093db2094d7cd20f3c2f385c212750648bd6ea4fb2bf507a6a4c55506"
693
+ },
694
+ {
695
+ "path": "variants/games_api_seed17/tokenizer/tokenizer_config.json",
696
+ "bytes": 692,
697
+ "sha256": "1cc816812993bff176eb4f7495433b736f06fba9b6e7b05cac7b4a1780650c95"
698
+ },
699
+ {
700
+ "path": "variants/games_api_seed17/backbone_config/config.json",
701
+ "bytes": 1417,
702
+ "sha256": "30c011854471509747858ac07ffb8a4dab2bc6c04035b184729b69500c557c54"
703
+ },
704
+ {
705
+ "path": "variants/events_ce_seed17/best.safetensors",
706
+ "bytes": 2385039280,
707
+ "sha256": "582c01139e29807b5f2690335271cf5647583cbfc3dd963d5967792bb3963992"
708
+ },
709
+ {
710
+ "path": "variants/events_ce_seed17/config.json",
711
+ "bytes": 2555,
712
+ "sha256": "52f0309ceeefe51d8cadda95b6f0985978b2d523d758194f682932892d82a566"
713
+ },
714
+ {
715
+ "path": "variants/events_ce_seed17/train_log.json",
716
+ "bytes": 28195,
717
+ "sha256": "c5e20da3ea98546ac71a9d2a264b53d52bcfb3387a7cd5a57d65c7cc03c9623e"
718
+ },
719
+ {
720
+ "path": "variants/events_ce_seed17/summary.json",
721
+ "bytes": 3840,
722
+ "sha256": "b5f1a6592da085ea88ca7676fbaf11b39874b097cdc70a3424c992010b27c3a5"
723
+ },
724
+ {
725
+ "path": "variants/events_ce_seed17/target_audit.json",
726
+ "bytes": 1334523,
727
+ "sha256": "7a9b2635cf87607e9f2a47fa1c2e037f1e5b13c18bcbbc5898cff09796222181"
728
+ },
729
+ {
730
+ "path": "variants/events_ce_seed17/tokenizer/chat_template.jinja",
731
+ "bytes": 4168,
732
+ "sha256": "a55ee1b1660128b7098723e0abcd92caa0788061051c62d51cbe87d9cf1974d8"
733
+ },
734
+ {
735
+ "path": "variants/events_ce_seed17/tokenizer/tokenizer.json",
736
+ "bytes": 11422650,
737
+ "sha256": "be75606093db2094d7cd20f3c2f385c212750648bd6ea4fb2bf507a6a4c55506"
738
+ },
739
+ {
740
+ "path": "variants/events_ce_seed17/tokenizer/tokenizer_config.json",
741
+ "bytes": 692,
742
+ "sha256": "1cc816812993bff176eb4f7495433b736f06fba9b6e7b05cac7b4a1780650c95"
743
+ },
744
+ {
745
+ "path": "variants/events_ce_seed17/backbone_config/config.json",
746
+ "bytes": 1417,
747
+ "sha256": "30c011854471509747858ac07ffb8a4dab2bc6c04035b184729b69500c557c54"
748
+ },
749
+ {
750
+ "path": "variants/events_brier_seed17/best.safetensors",
751
+ "bytes": 2385039280,
752
+ "sha256": "0a8d8b14c97743eed4310d461a7ae5cff0400cb87eaa6dc1f8c5dc3c59988651"
753
+ },
754
+ {
755
+ "path": "variants/events_brier_seed17/config.json",
756
+ "bytes": 2561,
757
+ "sha256": "bca3a9b31a8f528bc144433f1b2f4f9f27f78f50ceb5957c99949e246c850ac2"
758
+ },
759
+ {
760
+ "path": "variants/events_brier_seed17/train_log.json",
761
+ "bytes": 28219,
762
+ "sha256": "e09fdb39c212589b56af8f5cbae6f51fa1d1f91de73a63636e8661a02f2312f9"
763
+ },
764
+ {
765
+ "path": "variants/events_brier_seed17/summary.json",
766
+ "bytes": 3829,
767
+ "sha256": "daaf0e3739fb244368662944e26b939278a31beac7c1c538ce3e72aa70536def"
768
+ },
769
+ {
770
+ "path": "variants/events_brier_seed17/target_audit.json",
771
+ "bytes": 1334523,
772
+ "sha256": "7a9b2635cf87607e9f2a47fa1c2e037f1e5b13c18bcbbc5898cff09796222181"
773
+ },
774
+ {
775
+ "path": "variants/events_brier_seed17/tokenizer/chat_template.jinja",
776
+ "bytes": 4168,
777
+ "sha256": "a55ee1b1660128b7098723e0abcd92caa0788061051c62d51cbe87d9cf1974d8"
778
+ },
779
+ {
780
+ "path": "variants/events_brier_seed17/tokenizer/tokenizer.json",
781
+ "bytes": 11422650,
782
+ "sha256": "be75606093db2094d7cd20f3c2f385c212750648bd6ea4fb2bf507a6a4c55506"
783
+ },
784
+ {
785
+ "path": "variants/events_brier_seed17/tokenizer/tokenizer_config.json",
786
+ "bytes": 692,
787
+ "sha256": "1cc816812993bff176eb4f7495433b736f06fba9b6e7b05cac7b4a1780650c95"
788
+ },
789
+ {
790
+ "path": "variants/events_brier_seed17/backbone_config/config.json",
791
+ "bytes": 1417,
792
+ "sha256": "30c011854471509747858ac07ffb8a4dab2bc6c04035b184729b69500c557c54"
793
+ },
794
+ {
795
+ "path": "variants/events_paired_seed17/best.safetensors",
796
+ "bytes": 2385039280,
797
+ "sha256": "bbefcf829b9e36253ef46ab076f7d8fca00ffb33762e8358dbe780b3605bb777"
798
+ },
799
+ {
800
+ "path": "variants/events_paired_seed17/config.json",
801
+ "bytes": 2671,
802
+ "sha256": "9bdffa977eb28e4692ba19f53b05a7ad9afd9f37665848f9db24de4e829f1acb"
803
+ },
804
+ {
805
+ "path": "variants/events_paired_seed17/train_log.json",
806
+ "bytes": 28356,
807
+ "sha256": "33f337e289dab3afce8199fc7b72a6c766634986f49276ba8581ea2210b673cf"
808
+ },
809
+ {
810
+ "path": "variants/events_paired_seed17/summary.json",
811
+ "bytes": 3817,
812
+ "sha256": "82de7efe05e55bedc75e5b50b5b9ca867aa1f100cb4c4430b9b67524c6e82de1"
813
+ },
814
+ {
815
+ "path": "variants/events_paired_seed17/target_audit.json",
816
+ "bytes": 1334523,
817
+ "sha256": "7a9b2635cf87607e9f2a47fa1c2e037f1e5b13c18bcbbc5898cff09796222181"
818
+ },
819
+ {
820
+ "path": "variants/events_paired_seed17/tokenizer/chat_template.jinja",
821
+ "bytes": 4168,
822
+ "sha256": "a55ee1b1660128b7098723e0abcd92caa0788061051c62d51cbe87d9cf1974d8"
823
+ },
824
+ {
825
+ "path": "variants/events_paired_seed17/tokenizer/tokenizer.json",
826
+ "bytes": 11422650,
827
+ "sha256": "be75606093db2094d7cd20f3c2f385c212750648bd6ea4fb2bf507a6a4c55506"
828
+ },
829
+ {
830
+ "path": "variants/events_paired_seed17/tokenizer/tokenizer_config.json",
831
+ "bytes": 692,
832
+ "sha256": "1cc816812993bff176eb4f7495433b736f06fba9b6e7b05cac7b4a1780650c95"
833
+ },
834
+ {
835
+ "path": "variants/events_paired_seed17/backbone_config/config.json",
836
+ "bytes": 1417,
837
+ "sha256": "30c011854471509747858ac07ffb8a4dab2bc6c04035b184729b69500c557c54"
838
+ }
839
+ ],
840
+ "total_variant_files": 54,
841
+ "total_variant_bytes": 14387654859
842
+ }
MODEL_MANIFEST.json ADDED
@@ -0,0 +1,383 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ {
2
+ "schema_version": "nanojev-model-release-v1",
3
+ "model_id": "C-Tianyu/NanoJev",
4
+ "base_model": "Qwen/Qwen3-0.6B",
5
+ "base_revision": "c1899de289a04d12100db370d81485cdf75e47ca",
6
+ "dataset": "C-Tianyu/NanoJev-Data",
7
+ "files": [
8
+ {
9
+ "path": "README.md",
10
+ "bytes": 8030,
11
+ "sha256": "6f76023d0ad20ba48f501506a48d2b5df2136e22d8e73c1cc981e16295499983"
12
+ },
13
+ {
14
+ "path": "backbone_config/config.json",
15
+ "bytes": 1417,
16
+ "sha256": "30c011854471509747858ac07ffb8a4dab2bc6c04035b184729b69500c557c54"
17
+ },
18
+ {
19
+ "path": "best.safetensors",
20
+ "bytes": 2385039280,
21
+ "sha256": "fff62d1412685c1714eaa386acb603f9690371fb3cc8ad03dc41319302597c28"
22
+ },
23
+ {
24
+ "path": "config.json",
25
+ "bytes": 2332,
26
+ "sha256": "8139aff38992e200c92de038169b1924412dc7a73c72fc90d86d3db4734fc6c1"
27
+ },
28
+ {
29
+ "path": "stage1/backbone_config/config.json",
30
+ "bytes": 1417,
31
+ "sha256": "30c011854471509747858ac07ffb8a4dab2bc6c04035b184729b69500c557c54"
32
+ },
33
+ {
34
+ "path": "stage1/best.safetensors",
35
+ "bytes": 2385039280,
36
+ "sha256": "231b5178098477d9f82ae2cf38786d9ab8e0a80a019f009e55b772a66882cec4"
37
+ },
38
+ {
39
+ "path": "stage1/config.json",
40
+ "bytes": 2229,
41
+ "sha256": "eb010e7a9a4dc3cc46a9fc63c3c516a495667c81697c4b80d54059e91fc69cca"
42
+ },
43
+ {
44
+ "path": "stage1/tokenizer/chat_template.jinja",
45
+ "bytes": 4168,
46
+ "sha256": "a55ee1b1660128b7098723e0abcd92caa0788061051c62d51cbe87d9cf1974d8"
47
+ },
48
+ {
49
+ "path": "stage1/tokenizer/tokenizer.json",
50
+ "bytes": 11422650,
51
+ "sha256": "be75606093db2094d7cd20f3c2f385c212750648bd6ea4fb2bf507a6a4c55506"
52
+ },
53
+ {
54
+ "path": "stage1/tokenizer/tokenizer_config.json",
55
+ "bytes": 693,
56
+ "sha256": "c4f4c62b741ab2940841480f9942f378d7d3878d47c2babb862af7381c59b952"
57
+ },
58
+ {
59
+ "path": "tokenizer/chat_template.jinja",
60
+ "bytes": 4168,
61
+ "sha256": "a55ee1b1660128b7098723e0abcd92caa0788061051c62d51cbe87d9cf1974d8"
62
+ },
63
+ {
64
+ "path": "tokenizer/tokenizer.json",
65
+ "bytes": 11422650,
66
+ "sha256": "be75606093db2094d7cd20f3c2f385c212750648bd6ea4fb2bf507a6a4c55506"
67
+ },
68
+ {
69
+ "path": "tokenizer/tokenizer_config.json",
70
+ "bytes": 692,
71
+ "sha256": "1cc816812993bff176eb4f7495433b736f06fba9b6e7b05cac7b4a1780650c95"
72
+ },
73
+ {
74
+ "path": "variants/local_atomic_seed17/best.safetensors",
75
+ "bytes": 2385039280,
76
+ "sha256": "2b06e4423861f47da7ca53ad9c23a26aebdbfe79de3a238b36c2954ae9c6abf0"
77
+ },
78
+ {
79
+ "path": "variants/local_atomic_seed17/config.json",
80
+ "bytes": 2498,
81
+ "sha256": "3a6b014a77eaa3a79155a9c0c84ce2c005b5a5c66fdf88445067657ac9def176"
82
+ },
83
+ {
84
+ "path": "variants/local_atomic_seed17/train_log.json",
85
+ "bytes": 85584,
86
+ "sha256": "5e2a39713c30d714dddd8b7d989aecf1a8105fee24e74b42ef8943a1fd4dff09"
87
+ },
88
+ {
89
+ "path": "variants/local_atomic_seed17/summary.json",
90
+ "bytes": 4341,
91
+ "sha256": "3da659f939b0cc8d44614d9258f8660482db548a3f62e6f5c8604f15a9770980"
92
+ },
93
+ {
94
+ "path": "variants/local_atomic_seed17/target_audit.json",
95
+ "bytes": 701867,
96
+ "sha256": "ae1cde9eeedf6e84266205fba29b32bcb58489ca7da4f17704d74d2964d6cc79"
97
+ },
98
+ {
99
+ "path": "variants/local_atomic_seed17/tokenizer/chat_template.jinja",
100
+ "bytes": 4168,
101
+ "sha256": "a55ee1b1660128b7098723e0abcd92caa0788061051c62d51cbe87d9cf1974d8"
102
+ },
103
+ {
104
+ "path": "variants/local_atomic_seed17/tokenizer/tokenizer.json",
105
+ "bytes": 11422650,
106
+ "sha256": "be75606093db2094d7cd20f3c2f385c212750648bd6ea4fb2bf507a6a4c55506"
107
+ },
108
+ {
109
+ "path": "variants/local_atomic_seed17/tokenizer/tokenizer_config.json",
110
+ "bytes": 692,
111
+ "sha256": "1cc816812993bff176eb4f7495433b736f06fba9b6e7b05cac7b4a1780650c95"
112
+ },
113
+ {
114
+ "path": "variants/local_atomic_seed17/backbone_config/config.json",
115
+ "bytes": 1417,
116
+ "sha256": "30c011854471509747858ac07ffb8a4dab2bc6c04035b184729b69500c557c54"
117
+ },
118
+ {
119
+ "path": "variants/games_gold_seed17/best.safetensors",
120
+ "bytes": 2385039280,
121
+ "sha256": "5a395c56fad2a9bd4fb664ef78e1e06c19102886d97dd56dfe2676e5896c7e3a"
122
+ },
123
+ {
124
+ "path": "variants/games_gold_seed17/config.json",
125
+ "bytes": 2562,
126
+ "sha256": "965c5905315315ff591d6c2eb581146b749104e65afdd463e3a3118f70baaaff"
127
+ },
128
+ {
129
+ "path": "variants/games_gold_seed17/train_log.json",
130
+ "bytes": 42887,
131
+ "sha256": "a1c2de04bf59f1da31ab4255be69a5b5960fde183b8a459d2cd8b1f7c618ae28"
132
+ },
133
+ {
134
+ "path": "variants/games_gold_seed17/summary.json",
135
+ "bytes": 4530,
136
+ "sha256": "f7de836505d40a4f779258155455fc1d3d8194e468cae269bcb942c13441ff5c"
137
+ },
138
+ {
139
+ "path": "variants/games_gold_seed17/target_audit.json",
140
+ "bytes": 1831844,
141
+ "sha256": "62e50d044875e9914b7edce5d1321c873eb923a31fefaf0a32d8a13e0e755b1b"
142
+ },
143
+ {
144
+ "path": "variants/games_gold_seed17/tokenizer/chat_template.jinja",
145
+ "bytes": 4168,
146
+ "sha256": "a55ee1b1660128b7098723e0abcd92caa0788061051c62d51cbe87d9cf1974d8"
147
+ },
148
+ {
149
+ "path": "variants/games_gold_seed17/tokenizer/tokenizer.json",
150
+ "bytes": 11422650,
151
+ "sha256": "be75606093db2094d7cd20f3c2f385c212750648bd6ea4fb2bf507a6a4c55506"
152
+ },
153
+ {
154
+ "path": "variants/games_gold_seed17/tokenizer/tokenizer_config.json",
155
+ "bytes": 692,
156
+ "sha256": "1cc816812993bff176eb4f7495433b736f06fba9b6e7b05cac7b4a1780650c95"
157
+ },
158
+ {
159
+ "path": "variants/games_gold_seed17/backbone_config/config.json",
160
+ "bytes": 1417,
161
+ "sha256": "30c011854471509747858ac07ffb8a4dab2bc6c04035b184729b69500c557c54"
162
+ },
163
+ {
164
+ "path": "variants/games_api_seed17/best.safetensors",
165
+ "bytes": 2385039280,
166
+ "sha256": "45829fc896391847220b67dcf6fc0702427d2f184be384649bcf9343c1552477"
167
+ },
168
+ {
169
+ "path": "variants/games_api_seed17/config.json",
170
+ "bytes": 2693,
171
+ "sha256": "509c449f3088e167f95ea284370763489ca182e787d7765efbcffc377b767931"
172
+ },
173
+ {
174
+ "path": "variants/games_api_seed17/train_log.json",
175
+ "bytes": 42439,
176
+ "sha256": "aca1693ee0a2c1c1c574416329e28a089adaa93014a071b09ac4350b62e12050"
177
+ },
178
+ {
179
+ "path": "variants/games_api_seed17/summary.json",
180
+ "bytes": 3996,
181
+ "sha256": "304fde47ce4cef20bd06ed088c55f4a8bc2c8da3f68e3f5a9520a5bdbf14dbdb"
182
+ },
183
+ {
184
+ "path": "variants/games_api_seed17/target_audit.json",
185
+ "bytes": 2012764,
186
+ "sha256": "28a1666127b407bbfdaf8bcf99ca386c66a814a37ef78221bda12cdce9f8ab38"
187
+ },
188
+ {
189
+ "path": "variants/games_api_seed17/tokenizer/chat_template.jinja",
190
+ "bytes": 4168,
191
+ "sha256": "a55ee1b1660128b7098723e0abcd92caa0788061051c62d51cbe87d9cf1974d8"
192
+ },
193
+ {
194
+ "path": "variants/games_api_seed17/tokenizer/tokenizer.json",
195
+ "bytes": 11422650,
196
+ "sha256": "be75606093db2094d7cd20f3c2f385c212750648bd6ea4fb2bf507a6a4c55506"
197
+ },
198
+ {
199
+ "path": "variants/games_api_seed17/tokenizer/tokenizer_config.json",
200
+ "bytes": 692,
201
+ "sha256": "1cc816812993bff176eb4f7495433b736f06fba9b6e7b05cac7b4a1780650c95"
202
+ },
203
+ {
204
+ "path": "variants/games_api_seed17/backbone_config/config.json",
205
+ "bytes": 1417,
206
+ "sha256": "30c011854471509747858ac07ffb8a4dab2bc6c04035b184729b69500c557c54"
207
+ },
208
+ {
209
+ "path": "variants/events_ce_seed17/best.safetensors",
210
+ "bytes": 2385039280,
211
+ "sha256": "582c01139e29807b5f2690335271cf5647583cbfc3dd963d5967792bb3963992"
212
+ },
213
+ {
214
+ "path": "variants/events_ce_seed17/config.json",
215
+ "bytes": 2555,
216
+ "sha256": "52f0309ceeefe51d8cadda95b6f0985978b2d523d758194f682932892d82a566"
217
+ },
218
+ {
219
+ "path": "variants/events_ce_seed17/train_log.json",
220
+ "bytes": 28195,
221
+ "sha256": "c5e20da3ea98546ac71a9d2a264b53d52bcfb3387a7cd5a57d65c7cc03c9623e"
222
+ },
223
+ {
224
+ "path": "variants/events_ce_seed17/summary.json",
225
+ "bytes": 3840,
226
+ "sha256": "b5f1a6592da085ea88ca7676fbaf11b39874b097cdc70a3424c992010b27c3a5"
227
+ },
228
+ {
229
+ "path": "variants/events_ce_seed17/target_audit.json",
230
+ "bytes": 1334523,
231
+ "sha256": "7a9b2635cf87607e9f2a47fa1c2e037f1e5b13c18bcbbc5898cff09796222181"
232
+ },
233
+ {
234
+ "path": "variants/events_ce_seed17/tokenizer/chat_template.jinja",
235
+ "bytes": 4168,
236
+ "sha256": "a55ee1b1660128b7098723e0abcd92caa0788061051c62d51cbe87d9cf1974d8"
237
+ },
238
+ {
239
+ "path": "variants/events_ce_seed17/tokenizer/tokenizer.json",
240
+ "bytes": 11422650,
241
+ "sha256": "be75606093db2094d7cd20f3c2f385c212750648bd6ea4fb2bf507a6a4c55506"
242
+ },
243
+ {
244
+ "path": "variants/events_ce_seed17/tokenizer/tokenizer_config.json",
245
+ "bytes": 692,
246
+ "sha256": "1cc816812993bff176eb4f7495433b736f06fba9b6e7b05cac7b4a1780650c95"
247
+ },
248
+ {
249
+ "path": "variants/events_ce_seed17/backbone_config/config.json",
250
+ "bytes": 1417,
251
+ "sha256": "30c011854471509747858ac07ffb8a4dab2bc6c04035b184729b69500c557c54"
252
+ },
253
+ {
254
+ "path": "variants/events_brier_seed17/best.safetensors",
255
+ "bytes": 2385039280,
256
+ "sha256": "0a8d8b14c97743eed4310d461a7ae5cff0400cb87eaa6dc1f8c5dc3c59988651"
257
+ },
258
+ {
259
+ "path": "variants/events_brier_seed17/config.json",
260
+ "bytes": 2561,
261
+ "sha256": "bca3a9b31a8f528bc144433f1b2f4f9f27f78f50ceb5957c99949e246c850ac2"
262
+ },
263
+ {
264
+ "path": "variants/events_brier_seed17/train_log.json",
265
+ "bytes": 28219,
266
+ "sha256": "e09fdb39c212589b56af8f5cbae6f51fa1d1f91de73a63636e8661a02f2312f9"
267
+ },
268
+ {
269
+ "path": "variants/events_brier_seed17/summary.json",
270
+ "bytes": 3829,
271
+ "sha256": "daaf0e3739fb244368662944e26b939278a31beac7c1c538ce3e72aa70536def"
272
+ },
273
+ {
274
+ "path": "variants/events_brier_seed17/target_audit.json",
275
+ "bytes": 1334523,
276
+ "sha256": "7a9b2635cf87607e9f2a47fa1c2e037f1e5b13c18bcbbc5898cff09796222181"
277
+ },
278
+ {
279
+ "path": "variants/events_brier_seed17/tokenizer/chat_template.jinja",
280
+ "bytes": 4168,
281
+ "sha256": "a55ee1b1660128b7098723e0abcd92caa0788061051c62d51cbe87d9cf1974d8"
282
+ },
283
+ {
284
+ "path": "variants/events_brier_seed17/tokenizer/tokenizer.json",
285
+ "bytes": 11422650,
286
+ "sha256": "be75606093db2094d7cd20f3c2f385c212750648bd6ea4fb2bf507a6a4c55506"
287
+ },
288
+ {
289
+ "path": "variants/events_brier_seed17/tokenizer/tokenizer_config.json",
290
+ "bytes": 692,
291
+ "sha256": "1cc816812993bff176eb4f7495433b736f06fba9b6e7b05cac7b4a1780650c95"
292
+ },
293
+ {
294
+ "path": "variants/events_brier_seed17/backbone_config/config.json",
295
+ "bytes": 1417,
296
+ "sha256": "30c011854471509747858ac07ffb8a4dab2bc6c04035b184729b69500c557c54"
297
+ },
298
+ {
299
+ "path": "variants/events_paired_seed17/best.safetensors",
300
+ "bytes": 2385039280,
301
+ "sha256": "bbefcf829b9e36253ef46ab076f7d8fca00ffb33762e8358dbe780b3605bb777"
302
+ },
303
+ {
304
+ "path": "variants/events_paired_seed17/config.json",
305
+ "bytes": 2671,
306
+ "sha256": "9bdffa977eb28e4692ba19f53b05a7ad9afd9f37665848f9db24de4e829f1acb"
307
+ },
308
+ {
309
+ "path": "variants/events_paired_seed17/train_log.json",
310
+ "bytes": 28356,
311
+ "sha256": "33f337e289dab3afce8199fc7b72a6c766634986f49276ba8581ea2210b673cf"
312
+ },
313
+ {
314
+ "path": "variants/events_paired_seed17/summary.json",
315
+ "bytes": 3817,
316
+ "sha256": "82de7efe05e55bedc75e5b50b5b9ca867aa1f100cb4c4430b9b67524c6e82de1"
317
+ },
318
+ {
319
+ "path": "variants/events_paired_seed17/target_audit.json",
320
+ "bytes": 1334523,
321
+ "sha256": "7a9b2635cf87607e9f2a47fa1c2e037f1e5b13c18bcbbc5898cff09796222181"
322
+ },
323
+ {
324
+ "path": "variants/events_paired_seed17/tokenizer/chat_template.jinja",
325
+ "bytes": 4168,
326
+ "sha256": "a55ee1b1660128b7098723e0abcd92caa0788061051c62d51cbe87d9cf1974d8"
327
+ },
328
+ {
329
+ "path": "variants/events_paired_seed17/tokenizer/tokenizer.json",
330
+ "bytes": 11422650,
331
+ "sha256": "be75606093db2094d7cd20f3c2f385c212750648bd6ea4fb2bf507a6a4c55506"
332
+ },
333
+ {
334
+ "path": "variants/events_paired_seed17/tokenizer/tokenizer_config.json",
335
+ "bytes": 692,
336
+ "sha256": "1cc816812993bff176eb4f7495433b736f06fba9b6e7b05cac7b4a1780650c95"
337
+ },
338
+ {
339
+ "path": "variants/events_paired_seed17/backbone_config/config.json",
340
+ "bytes": 1417,
341
+ "sha256": "30c011854471509747858ac07ffb8a4dab2bc6c04035b184729b69500c557c54"
342
+ },
343
+ {
344
+ "path": "GAMES_MODEL_MANIFEST.json",
345
+ "bytes": 32465,
346
+ "sha256": "ed8180264c7052c7806c11b6f2d6606b673aa41eb1d636597393ec89843b94f9"
347
+ }
348
+ ],
349
+ "stage1_input_sha256": "7294765b80e751fc5aee7aba906b28a8ea80d6f147253d3f6c3e4f491de0b2d7",
350
+ "stage2_input_sha256": "3a638666206e8ffacadd8beeaf0246687e546716deb8b06adf5e23a228890cdc",
351
+ "game_variants": [
352
+ {
353
+ "name": "local_atomic_seed17",
354
+ "path": "variants/local_atomic_seed17",
355
+ "weight_sha256": "2b06e4423861f47da7ca53ad9c23a26aebdbfe79de3a238b36c2954ae9c6abf0"
356
+ },
357
+ {
358
+ "name": "games_gold_seed17",
359
+ "path": "variants/games_gold_seed17",
360
+ "weight_sha256": "5a395c56fad2a9bd4fb664ef78e1e06c19102886d97dd56dfe2676e5896c7e3a"
361
+ },
362
+ {
363
+ "name": "games_api_seed17",
364
+ "path": "variants/games_api_seed17",
365
+ "weight_sha256": "45829fc896391847220b67dcf6fc0702427d2f184be384649bcf9343c1552477"
366
+ },
367
+ {
368
+ "name": "events_ce_seed17",
369
+ "path": "variants/events_ce_seed17",
370
+ "weight_sha256": "582c01139e29807b5f2690335271cf5647583cbfc3dd963d5967792bb3963992"
371
+ },
372
+ {
373
+ "name": "events_brier_seed17",
374
+ "path": "variants/events_brier_seed17",
375
+ "weight_sha256": "0a8d8b14c97743eed4310d461a7ae5cff0400cb87eaa6dc1f8c5dc3c59988651"
376
+ },
377
+ {
378
+ "name": "events_paired_seed17",
379
+ "path": "variants/events_paired_seed17",
380
+ "weight_sha256": "bbefcf829b9e36253ef46ab076f7d8fca00ffb33762e8358dbe780b3605bb777"
381
+ }
382
+ ]
383
+ }
README.md ADDED
@@ -0,0 +1,224 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ ---
2
+ language:
3
+ - en
4
+ - zh
5
+ base_model: Qwen/Qwen3-0.6B
6
+ library_name: pytorch
7
+ tags:
8
+ - nanojev
9
+ - decision-model
10
+ - probability-distributions
11
+ - parallel-inference
12
+ - navigation
13
+ ---
14
+
15
+ # NanoJev — A nano replica of [Jev](https://typesafe.ai/blog/introducing-system-one-models-and-jev)
16
+
17
+ **A 0.6B parallel decision model that returns complete probability distributions over dynamic candidates.**
18
+
19
+ Give NanoJev multiple states and questions in one request. It evaluates the candidate paths in one batched backbone forward pass and returns structured decisions with **zero output-token decoding**.
20
+
21
+ [Source code](https://github.com/TianyuCodings/NanoJev) · [Dataset](https://huggingface.co/datasets/C-Tianyu/NanoJev-Data) · [Pipeline commands](https://github.com/TianyuCodings/NanoJev/blob/main/research/pipeline_runbook.md)
22
+
23
+ ## Features
24
+
25
+ - **Qwen3-0.6B backbone** with structured decision heads.
26
+ - **Multiple states and questions** evaluated in one batched forward.
27
+ - **Choice:** 2–255 dynamic candidates with a probability for each candidate.
28
+ - **Boolean:** the probability that a proposition is true.
29
+ - **Score:** 2–10 ordered levels, their complete distribution, and an expected score.
30
+ - **Full distributions** for greedy selection or probability sampling.
31
+ - **Persistent serving:** load a checkpoint once and reuse it across requests.
32
+
33
+ Measured in the running service: **6 states · 18 questions · 44 candidate paths · 1 backbone forward**.
34
+
35
+ ## Maze, Snake, and calibrated-decision checkpoints
36
+
37
+ The game showcase uses two specialized checkpoints. Every variant below includes complete weights, configuration, tokenizer, and backbone configuration, plus its training summary.
38
+
39
+ | Variant directory | Use |
40
+ |---|---|
41
+ | `variants/local_atomic_seed17` | Local safety judgments used in the 50×50 maze showcase |
42
+ | `variants/games_gold_seed17` | Dynamic action choices used in the 12×12 Snake showcase |
43
+ | `variants/games_api_seed17` | Full-map game-question comparison |
44
+ | `variants/events_ce_seed17` | Observed-event cross-entropy control |
45
+ | `variants/events_brier_seed17` | Observed-event Brier control |
46
+ | `variants/events_paired_seed17` | RLCD-inspired paired proper-reward experiment |
47
+
48
+ [Watch the game demos](https://github.com/TianyuCodings/NanoJev#recorded-showcase-runs) · [Game data](https://huggingface.co/datasets/C-Tianyu/NanoJev-Data/tree/main/games_v4) · [Variant manifest](GAMES_MODEL_MANIFEST.json)
49
+
50
+ The recorded 50×50 maze run reaches its goal in 244 attempts. The selected 12×12 Snake run collects 27 food and survives 256 steps with greedy control. These are model decisions composed with the shared planning code described in the repository.
51
+
52
+ Download the exact model for either showcase:
53
+
54
+ ```python
55
+ from pathlib import Path
56
+ from huggingface_hub import snapshot_download
57
+
58
+ variant = "local_atomic_seed17" # Use "games_gold_seed17" for Snake.
59
+ snapshot = snapshot_download(
60
+ repo_id="C-Tianyu/NanoJev",
61
+ allow_patterns=[f"variants/{variant}/*"],
62
+ )
63
+ checkpoint_dir = Path(snapshot) / "variants" / variant
64
+ print(checkpoint_dir)
65
+ ```
66
+
67
+ Pass the printed directory to the matching evaluator's `--checkpoint` argument, or the persistent service's `--checkpoint-dir`. Full commands are in the [game release guide](https://github.com/TianyuCodings/NanoJev/blob/main/docs/GAME_RELEASE.md).
68
+
69
+ ## Earlier navigation performance
70
+
71
+ **Controller: T=1 probability sampling.** The complete benchmark contains 20 test maps and 20 OOD maps.
72
+
73
+ | System | 4×4 test | 6×6 OOD |
74
+ |---|---:|---:|
75
+ | **trained NanoJev** | **19/20 — 95%** | **18/20 — 90%** |
76
+ | Jev | 20/20 — 100% | 19/20 — 95% |
77
+ | Original Qwen3-0.6B | 7/20 — 35% | 3/20 — 15% |
78
+
79
+ Original Qwen is pretrained without task-specific fine-tuning. Its action distribution is obtained from the native language-model head, conditioned on the offered A–D answer tokens.
80
+
81
+ ## Download the base release checkpoint
82
+
83
+ The base release checkpoint for the earlier navigation benchmark is stored at the repository root:
84
+
85
+ ```text
86
+ best.safetensors
87
+ config.json
88
+ tokenizer/
89
+ backbone_config/
90
+ ```
91
+
92
+ Use `snapshot_download` to retrieve only the final checkpoint files:
93
+
94
+ ```python
95
+ from huggingface_hub import snapshot_download
96
+
97
+ checkpoint_dir = snapshot_download(
98
+ repo_id="C-Tianyu/NanoJev",
99
+ allow_patterns=[
100
+ "best.safetensors",
101
+ "config.json",
102
+ "tokenizer/*",
103
+ "backbone_config/*",
104
+ ],
105
+ )
106
+ print(checkpoint_dir)
107
+ ```
108
+
109
+ Use the printed snapshot directory directly as `--checkpoint-dir`.
110
+
111
+ ## Run inference
112
+
113
+ Clone the implementation and install its recorded Python dependencies in a compatible CUDA environment:
114
+
115
+ ```bash
116
+ git clone https://github.com/TianyuCodings/NanoJev.git
117
+ cd NanoJev
118
+ python -m pip install -r requirements-toy.txt
119
+ ```
120
+
121
+ Save this request as `request.json`:
122
+
123
+ ```json
124
+ {
125
+ "states": [{
126
+ "id": "living_room",
127
+ "state": "The room is 29 degrees, the target is 24 degrees, and someone is home.",
128
+ "questions": {
129
+ "action": {
130
+ "type": "choice",
131
+ "instructions": "Choose the action that most directly lowers the room temperature.",
132
+ "criteria": {
133
+ "cool": "Turn on air conditioning",
134
+ "light": "Turn on the lights",
135
+ "wait": "Keep the current settings"
136
+ }
137
+ },
138
+ "occupied": {
139
+ "type": "boolean",
140
+ "instructions": "Someone is home."
141
+ },
142
+ "heat": {
143
+ "type": "score",
144
+ "instructions": "Classify how far the room temperature exceeds the target.",
145
+ "criteria": [
146
+ "At or below the target",
147
+ "Above the target by at most 3 degrees",
148
+ "Above the target by more than 3 degrees"
149
+ ]
150
+ }
151
+ }
152
+ }]
153
+ }
154
+ ```
155
+
156
+ Run all questions in one forward pass:
157
+
158
+ ```bash
159
+ CUDA_VISIBLE_DEVICES=0 python scripts/predict_toy_decisions.py \
160
+ --checkpoint-dir path/to/downloaded/snapshot \
161
+ --input request.json \
162
+ --output predictions.json \
163
+ --batch-questions 0 --precision bf16 --temperature 1
164
+ ```
165
+
166
+ Add more objects to `states` to evaluate multiple states together. The response contains the distributions, selected choices, expected scores, and execution counters.
167
+
168
+ For a persistent HTTP service:
169
+
170
+ ```bash
171
+ CUDA_VISIBLE_DEVICES=0 python scripts/serve_decisions.py \
172
+ --checkpoint-dir path/to/downloaded/snapshot \
173
+ --web-root web --port 8765 --precision bf16
174
+ ```
175
+
176
+ Open `http://127.0.0.1:8765` or send requests to `POST /api/evaluate`.
177
+
178
+ ## Stage 1 initialization checkpoint
179
+
180
+ The `stage1/` directory contains an initialization checkpoint for further training with the same model architecture. Download it separately:
181
+
182
+ ```python
183
+ from pathlib import Path
184
+ from huggingface_hub import snapshot_download
185
+
186
+ snapshot_dir = snapshot_download(
187
+ repo_id="C-Tianyu/NanoJev",
188
+ allow_patterns=[
189
+ "stage1/best.safetensors",
190
+ "stage1/config.json",
191
+ "stage1/tokenizer/*",
192
+ "stage1/backbone_config/*",
193
+ ],
194
+ )
195
+ stage1_dir = Path(snapshot_dir) / "stage1"
196
+ print(stage1_dir)
197
+ ```
198
+
199
+ Pass this directory to `scripts/train_pipeline_decisions.py` with `--init-checkpoint path/to/stage1`. The trainer loads that checkpoint's architecture and parameters and starts a new optimizer. Use the root checkpoint for the earlier navigation benchmark; select a game variant for the new showcases.
200
+
201
+ ## Download the data
202
+
203
+ ```python
204
+ from huggingface_hub import snapshot_download
205
+
206
+ data_dir = snapshot_download(
207
+ repo_id="C-Tianyu/NanoJev-Data",
208
+ repo_type="dataset",
209
+ )
210
+ print(data_dir)
211
+ ```
212
+
213
+ The [pipeline guide](https://github.com/TianyuCodings/NanoJev/blob/main/research/pipeline_runbook.md) covers the JSONL target-distribution interface, data generation, training, evaluation, and serving.
214
+
215
+ ## Checkpoint details
216
+
217
+ | Item | Value |
218
+ |---|---|
219
+ | Backbone | `Qwen/Qwen3-0.6B` |
220
+ | Backbone revision | `c1899de289a04d12100db370d81485cdf75e47ca` |
221
+ | Final weights | `best.safetensors` |
222
+ | Final SHA256 | `fff62d1412685c1714eaa386acb603f9690371fb3cc8ad03dc41319302597c28` |
223
+ | Stage 1 weights | `stage1/best.safetensors` |
224
+ | Stage 1 SHA256 | `231b5178098477d9f82ae2cf38786d9ab8e0a80a019f009e55b772a66882cec4` |
backbone_config/config.json ADDED
@@ -0,0 +1,63 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ {
2
+ "architectures": [
3
+ "Qwen3ForCausalLM"
4
+ ],
5
+ "attention_bias": false,
6
+ "attention_dropout": 0.0,
7
+ "bos_token_id": 151643,
8
+ "dtype": "float32",
9
+ "eos_token_id": 151645,
10
+ "head_dim": 128,
11
+ "hidden_act": "silu",
12
+ "hidden_size": 1024,
13
+ "initializer_range": 0.02,
14
+ "intermediate_size": 3072,
15
+ "layer_types": [
16
+ "full_attention",
17
+ "full_attention",
18
+ "full_attention",
19
+ "full_attention",
20
+ "full_attention",
21
+ "full_attention",
22
+ "full_attention",
23
+ "full_attention",
24
+ "full_attention",
25
+ "full_attention",
26
+ "full_attention",
27
+ "full_attention",
28
+ "full_attention",
29
+ "full_attention",
30
+ "full_attention",
31
+ "full_attention",
32
+ "full_attention",
33
+ "full_attention",
34
+ "full_attention",
35
+ "full_attention",
36
+ "full_attention",
37
+ "full_attention",
38
+ "full_attention",
39
+ "full_attention",
40
+ "full_attention",
41
+ "full_attention",
42
+ "full_attention",
43
+ "full_attention"
44
+ ],
45
+ "max_position_embeddings": 40960,
46
+ "max_window_layers": 28,
47
+ "model_type": "qwen3",
48
+ "num_attention_heads": 16,
49
+ "num_hidden_layers": 28,
50
+ "num_key_value_heads": 8,
51
+ "pad_token_id": null,
52
+ "rms_norm_eps": 1e-06,
53
+ "rope_parameters": {
54
+ "rope_theta": 1000000,
55
+ "rope_type": "default"
56
+ },
57
+ "sliding_window": null,
58
+ "tie_word_embeddings": true,
59
+ "transformers_version": "5.17.0",
60
+ "use_cache": false,
61
+ "use_sliding_window": false,
62
+ "vocab_size": 151936
63
+ }
best.safetensors ADDED
@@ -0,0 +1,3 @@
 
 
 
 
1
+ version https://git-lfs.github.com/spec/v1
2
+ oid sha256:fff62d1412685c1714eaa386acb603f9690371fb3cc8ad03dc41319302597c28
3
+ size 2385039280
config.json ADDED
@@ -0,0 +1,66 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ {
2
+ "input": "/home/rwang/openjev_codex_20260917/research/private_navigation_v3/views/coords_multi.jsonl",
3
+ "output_dir": "/home/rwang/openjev_codex_20260917/runs/v3_teacher_coords_multi_seed17",
4
+ "model": "Qwen/Qwen3-0.6B",
5
+ "revision": "c1899de289a04d12100db370d81485cdf75e47ca",
6
+ "init_checkpoint": "/home/rwang/openjev_codex_20260917/runs/v2_teacher_seed17",
7
+ "objective": "teacher",
8
+ "set_head": "attention",
9
+ "steps": 1200,
10
+ "head_steps": 0,
11
+ "batch_questions": 12,
12
+ "microbatch_questions": 4,
13
+ "max_microbatch_tokens": 6000,
14
+ "eval_every": 50,
15
+ "max_length": 512,
16
+ "seed": 17,
17
+ "backbone_lr": 2e-05,
18
+ "head_lr": 0.0002,
19
+ "head_warmup_lr": 0.001,
20
+ "precision": "bf16",
21
+ "disable_native_triton": true,
22
+ "validate_only": false,
23
+ "self_check": false,
24
+ "schema_version": "openjev-decision-pipeline-v1",
25
+ "resolved_model_revision": "c1899de289a04d12100db370d81485cdf75e47ca",
26
+ "initialization": "local DecisionModel warm start, fresh optimizer",
27
+ "data_sha256": {
28
+ "/home/rwang/openjev_codex_20260917/research/private_navigation_v3/views/coords_multi.jsonl": "3a638666206e8ffacadd8beeaf0246687e546716deb8b06adf5e23a228890cdc"
29
+ },
30
+ "deps": {
31
+ "torch": "2.14.0",
32
+ "transformers": "5.17.0",
33
+ "safetensors": "0.8.0"
34
+ },
35
+ "gpu": "NVIDIA A100-SXM4-80GB",
36
+ "parameter_storage": "float32",
37
+ "forward_autocast": "bfloat16",
38
+ "parameter_count": 596250498,
39
+ "train_questions": 10105,
40
+ "all_train_questions": 10152,
41
+ "schema_counts": {
42
+ "records": 3984,
43
+ "questions_by_split": {
44
+ "train": 10152,
45
+ "dev": 360,
46
+ "calibration": 360,
47
+ "test": 720,
48
+ "ood": 360
49
+ },
50
+ "eligible_by_split_objective": {
51
+ "calibration/gold_distribution": 360,
52
+ "calibration/teacher": 359,
53
+ "dev/gold_distribution": 360,
54
+ "dev/teacher": 357,
55
+ "ood/gold_distribution": 360,
56
+ "ood/teacher": 357,
57
+ "test/gold_distribution": 720,
58
+ "test/teacher": 715,
59
+ "train/gold_distribution": 10152,
60
+ "train/teacher": 10105
61
+ }
62
+ },
63
+ "loss": "per-question forward soft CE, equal question weight; complete candidate normalization",
64
+ "selection": "minimum dev target CE; held-out test first evaluated after training and checkpoint selection",
65
+ "parallelism": "complete candidate paths batched in one backbone call per microbatch; no prefix sharing"
66
+ }
stage1/backbone_config/config.json ADDED
@@ -0,0 +1,63 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ {
2
+ "architectures": [
3
+ "Qwen3ForCausalLM"
4
+ ],
5
+ "attention_bias": false,
6
+ "attention_dropout": 0.0,
7
+ "bos_token_id": 151643,
8
+ "dtype": "float32",
9
+ "eos_token_id": 151645,
10
+ "head_dim": 128,
11
+ "hidden_act": "silu",
12
+ "hidden_size": 1024,
13
+ "initializer_range": 0.02,
14
+ "intermediate_size": 3072,
15
+ "layer_types": [
16
+ "full_attention",
17
+ "full_attention",
18
+ "full_attention",
19
+ "full_attention",
20
+ "full_attention",
21
+ "full_attention",
22
+ "full_attention",
23
+ "full_attention",
24
+ "full_attention",
25
+ "full_attention",
26
+ "full_attention",
27
+ "full_attention",
28
+ "full_attention",
29
+ "full_attention",
30
+ "full_attention",
31
+ "full_attention",
32
+ "full_attention",
33
+ "full_attention",
34
+ "full_attention",
35
+ "full_attention",
36
+ "full_attention",
37
+ "full_attention",
38
+ "full_attention",
39
+ "full_attention",
40
+ "full_attention",
41
+ "full_attention",
42
+ "full_attention",
43
+ "full_attention"
44
+ ],
45
+ "max_position_embeddings": 40960,
46
+ "max_window_layers": 28,
47
+ "model_type": "qwen3",
48
+ "num_attention_heads": 16,
49
+ "num_hidden_layers": 28,
50
+ "num_key_value_heads": 8,
51
+ "pad_token_id": null,
52
+ "rms_norm_eps": 1e-06,
53
+ "rope_parameters": {
54
+ "rope_theta": 1000000,
55
+ "rope_type": "default"
56
+ },
57
+ "sliding_window": null,
58
+ "tie_word_embeddings": true,
59
+ "transformers_version": "5.17.0",
60
+ "use_cache": false,
61
+ "use_sliding_window": false,
62
+ "vocab_size": 151936
63
+ }
stage1/best.safetensors ADDED
@@ -0,0 +1,3 @@
 
 
 
 
1
+ version https://git-lfs.github.com/spec/v1
2
+ oid sha256:231b5178098477d9f82ae2cf38786d9ab8e0a80a019f009e55b772a66882cec4
3
+ size 2385039280
stage1/config.json ADDED
@@ -0,0 +1,66 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ {
2
+ "input": "/home/rwang/openjev_codex_20260917/research/private_pipeline_v2/merged.jsonl",
3
+ "output_dir": "/home/rwang/openjev_codex_20260917/runs/v2_teacher_seed17",
4
+ "model": "Qwen/Qwen3-0.6B",
5
+ "revision": "c1899de289a04d12100db370d81485cdf75e47ca",
6
+ "init_checkpoint": null,
7
+ "objective": "teacher",
8
+ "set_head": "attention",
9
+ "steps": 600,
10
+ "head_steps": 12,
11
+ "batch_questions": 12,
12
+ "microbatch_questions": 4,
13
+ "max_microbatch_tokens": 6000,
14
+ "eval_every": 50,
15
+ "max_length": 512,
16
+ "seed": 17,
17
+ "backbone_lr": 2e-05,
18
+ "head_lr": 0.0002,
19
+ "head_warmup_lr": 0.001,
20
+ "precision": "bf16",
21
+ "disable_native_triton": true,
22
+ "validate_only": false,
23
+ "self_check": false,
24
+ "schema_version": "openjev-decision-pipeline-v1",
25
+ "resolved_model_revision": "c1899de289a04d12100db370d81485cdf75e47ca",
26
+ "initialization": "pretrained backbone with new decision head",
27
+ "data_sha256": {
28
+ "/home/rwang/openjev_codex_20260917/research/private_pipeline_v2/merged.jsonl": "7294765b80e751fc5aee7aba906b28a8ea80d6f147253d3f6c3e4f491de0b2d7"
29
+ },
30
+ "deps": {
31
+ "torch": "2.14.0",
32
+ "transformers": "5.17.0",
33
+ "safetensors": "0.8.0"
34
+ },
35
+ "gpu": "NVIDIA A100-SXM4-80GB",
36
+ "parameter_storage": "float32",
37
+ "forward_autocast": "bfloat16",
38
+ "parameter_count": 596250498,
39
+ "train_questions": 4133,
40
+ "all_train_questions": 4152,
41
+ "schema_counts": {
42
+ "records": 2312,
43
+ "questions_by_split": {
44
+ "train": 4152,
45
+ "dev": 552,
46
+ "calibration": 552,
47
+ "test": 1104,
48
+ "ood": 576
49
+ },
50
+ "eligible_by_split_objective": {
51
+ "calibration/gold_distribution": 552,
52
+ "calibration/teacher": 550,
53
+ "dev/gold_distribution": 552,
54
+ "dev/teacher": 548,
55
+ "ood/gold_distribution": 576,
56
+ "ood/teacher": 573,
57
+ "test/gold_distribution": 1104,
58
+ "test/teacher": 1100,
59
+ "train/gold_distribution": 4152,
60
+ "train/teacher": 4133
61
+ }
62
+ },
63
+ "loss": "per-question forward soft CE, equal question weight; complete candidate normalization",
64
+ "selection": "minimum dev target CE; held-out test first evaluated after training and checkpoint selection",
65
+ "parallelism": "complete candidate paths batched in one backbone call per microbatch; no prefix sharing"
66
+ }
stage1/tokenizer/chat_template.jinja ADDED
@@ -0,0 +1,89 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ {%- if tools %}
2
+ {{- '<|im_start|>system\n' }}
3
+ {%- if messages[0].role == 'system' %}
4
+ {{- messages[0].content + '\n\n' }}
5
+ {%- endif %}
6
+ {{- "# Tools\n\nYou may call one or more functions to assist with the user query.\n\nYou are provided with function signatures within <tools></tools> XML tags:\n<tools>" }}
7
+ {%- for tool in tools %}
8
+ {{- "\n" }}
9
+ {{- tool | tojson }}
10
+ {%- endfor %}
11
+ {{- "\n</tools>\n\nFor each function call, return a json object with function name and arguments within <tool_call></tool_call> XML tags:\n<tool_call>\n{\"name\": <function-name>, \"arguments\": <args-json-object>}\n</tool_call><|im_end|>\n" }}
12
+ {%- else %}
13
+ {%- if messages[0].role == 'system' %}
14
+ {{- '<|im_start|>system\n' + messages[0].content + '<|im_end|>\n' }}
15
+ {%- endif %}
16
+ {%- endif %}
17
+ {%- set ns = namespace(multi_step_tool=true, last_query_index=messages|length - 1) %}
18
+ {%- for message in messages[::-1] %}
19
+ {%- set index = (messages|length - 1) - loop.index0 %}
20
+ {%- if ns.multi_step_tool and message.role == "user" and message.content is string and not(message.content.startswith('<tool_response>') and message.content.endswith('</tool_response>')) %}
21
+ {%- set ns.multi_step_tool = false %}
22
+ {%- set ns.last_query_index = index %}
23
+ {%- endif %}
24
+ {%- endfor %}
25
+ {%- for message in messages %}
26
+ {%- if message.content is string %}
27
+ {%- set content = message.content %}
28
+ {%- else %}
29
+ {%- set content = '' %}
30
+ {%- endif %}
31
+ {%- if (message.role == "user") or (message.role == "system" and not loop.first) %}
32
+ {{- '<|im_start|>' + message.role + '\n' + content + '<|im_end|>' + '\n' }}
33
+ {%- elif message.role == "assistant" %}
34
+ {%- set reasoning_content = '' %}
35
+ {%- if message.reasoning_content is string %}
36
+ {%- set reasoning_content = message.reasoning_content %}
37
+ {%- else %}
38
+ {%- if '</think>' in content %}
39
+ {%- set reasoning_content = content.split('</think>')[0].rstrip('\n').split('<think>')[-1].lstrip('\n') %}
40
+ {%- set content = content.split('</think>')[-1].lstrip('\n') %}
41
+ {%- endif %}
42
+ {%- endif %}
43
+ {%- if loop.index0 > ns.last_query_index %}
44
+ {%- if loop.last or (not loop.last and reasoning_content) %}
45
+ {{- '<|im_start|>' + message.role + '\n<think>\n' + reasoning_content.strip('\n') + '\n</think>\n\n' + content.lstrip('\n') }}
46
+ {%- else %}
47
+ {{- '<|im_start|>' + message.role + '\n' + content }}
48
+ {%- endif %}
49
+ {%- else %}
50
+ {{- '<|im_start|>' + message.role + '\n' + content }}
51
+ {%- endif %}
52
+ {%- if message.tool_calls %}
53
+ {%- for tool_call in message.tool_calls %}
54
+ {%- if (loop.first and content) or (not loop.first) %}
55
+ {{- '\n' }}
56
+ {%- endif %}
57
+ {%- if tool_call.function %}
58
+ {%- set tool_call = tool_call.function %}
59
+ {%- endif %}
60
+ {{- '<tool_call>\n{"name": "' }}
61
+ {{- tool_call.name }}
62
+ {{- '", "arguments": ' }}
63
+ {%- if tool_call.arguments is string %}
64
+ {{- tool_call.arguments }}
65
+ {%- else %}
66
+ {{- tool_call.arguments | tojson }}
67
+ {%- endif %}
68
+ {{- '}\n</tool_call>' }}
69
+ {%- endfor %}
70
+ {%- endif %}
71
+ {{- '<|im_end|>\n' }}
72
+ {%- elif message.role == "tool" %}
73
+ {%- if loop.first or (messages[loop.index0 - 1].role != "tool") %}
74
+ {{- '<|im_start|>user' }}
75
+ {%- endif %}
76
+ {{- '\n<tool_response>\n' }}
77
+ {{- content }}
78
+ {{- '\n</tool_response>' }}
79
+ {%- if loop.last or (messages[loop.index0 + 1].role != "tool") %}
80
+ {{- '<|im_end|>\n' }}
81
+ {%- endif %}
82
+ {%- endif %}
83
+ {%- endfor %}
84
+ {%- if add_generation_prompt %}
85
+ {{- '<|im_start|>assistant\n' }}
86
+ {%- if enable_thinking is defined and enable_thinking is false %}
87
+ {{- '<think>\n\n</think>\n\n' }}
88
+ {%- endif %}
89
+ {%- endif %}
stage1/tokenizer/tokenizer.json ADDED
@@ -0,0 +1,3 @@
 
 
 
 
1
+ version https://git-lfs.github.com/spec/v1
2
+ oid sha256:be75606093db2094d7cd20f3c2f385c212750648bd6ea4fb2bf507a6a4c55506
3
+ size 11422650
stage1/tokenizer/tokenizer_config.json ADDED
@@ -0,0 +1,30 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ {
2
+ "add_prefix_space": false,
3
+ "backend": "tokenizers",
4
+ "bos_token": null,
5
+ "clean_up_tokenization_spaces": false,
6
+ "eos_token": "<|im_end|>",
7
+ "errors": "replace",
8
+ "extra_special_tokens": [
9
+ "<|im_start|>",
10
+ "<|im_end|>",
11
+ "<|object_ref_start|>",
12
+ "<|object_ref_end|>",
13
+ "<|box_start|>",
14
+ "<|box_end|>",
15
+ "<|quad_start|>",
16
+ "<|quad_end|>",
17
+ "<|vision_start|>",
18
+ "<|vision_end|>",
19
+ "<|vision_pad|>",
20
+ "<|image_pad|>",
21
+ "<|video_pad|>"
22
+ ],
23
+ "is_local": false,
24
+ "local_files_only": true,
25
+ "model_max_length": 131072,
26
+ "pad_token": "<|endoftext|>",
27
+ "split_special_tokens": false,
28
+ "tokenizer_class": "Qwen2Tokenizer",
29
+ "unk_token": null
30
+ }
tokenizer/chat_template.jinja ADDED
@@ -0,0 +1,89 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ {%- if tools %}
2
+ {{- '<|im_start|>system\n' }}
3
+ {%- if messages[0].role == 'system' %}
4
+ {{- messages[0].content + '\n\n' }}
5
+ {%- endif %}
6
+ {{- "# Tools\n\nYou may call one or more functions to assist with the user query.\n\nYou are provided with function signatures within <tools></tools> XML tags:\n<tools>" }}
7
+ {%- for tool in tools %}
8
+ {{- "\n" }}
9
+ {{- tool | tojson }}
10
+ {%- endfor %}
11
+ {{- "\n</tools>\n\nFor each function call, return a json object with function name and arguments within <tool_call></tool_call> XML tags:\n<tool_call>\n{\"name\": <function-name>, \"arguments\": <args-json-object>}\n</tool_call><|im_end|>\n" }}
12
+ {%- else %}
13
+ {%- if messages[0].role == 'system' %}
14
+ {{- '<|im_start|>system\n' + messages[0].content + '<|im_end|>\n' }}
15
+ {%- endif %}
16
+ {%- endif %}
17
+ {%- set ns = namespace(multi_step_tool=true, last_query_index=messages|length - 1) %}
18
+ {%- for message in messages[::-1] %}
19
+ {%- set index = (messages|length - 1) - loop.index0 %}
20
+ {%- if ns.multi_step_tool and message.role == "user" and message.content is string and not(message.content.startswith('<tool_response>') and message.content.endswith('</tool_response>')) %}
21
+ {%- set ns.multi_step_tool = false %}
22
+ {%- set ns.last_query_index = index %}
23
+ {%- endif %}
24
+ {%- endfor %}
25
+ {%- for message in messages %}
26
+ {%- if message.content is string %}
27
+ {%- set content = message.content %}
28
+ {%- else %}
29
+ {%- set content = '' %}
30
+ {%- endif %}
31
+ {%- if (message.role == "user") or (message.role == "system" and not loop.first) %}
32
+ {{- '<|im_start|>' + message.role + '\n' + content + '<|im_end|>' + '\n' }}
33
+ {%- elif message.role == "assistant" %}
34
+ {%- set reasoning_content = '' %}
35
+ {%- if message.reasoning_content is string %}
36
+ {%- set reasoning_content = message.reasoning_content %}
37
+ {%- else %}
38
+ {%- if '</think>' in content %}
39
+ {%- set reasoning_content = content.split('</think>')[0].rstrip('\n').split('<think>')[-1].lstrip('\n') %}
40
+ {%- set content = content.split('</think>')[-1].lstrip('\n') %}
41
+ {%- endif %}
42
+ {%- endif %}
43
+ {%- if loop.index0 > ns.last_query_index %}
44
+ {%- if loop.last or (not loop.last and reasoning_content) %}
45
+ {{- '<|im_start|>' + message.role + '\n<think>\n' + reasoning_content.strip('\n') + '\n</think>\n\n' + content.lstrip('\n') }}
46
+ {%- else %}
47
+ {{- '<|im_start|>' + message.role + '\n' + content }}
48
+ {%- endif %}
49
+ {%- else %}
50
+ {{- '<|im_start|>' + message.role + '\n' + content }}
51
+ {%- endif %}
52
+ {%- if message.tool_calls %}
53
+ {%- for tool_call in message.tool_calls %}
54
+ {%- if (loop.first and content) or (not loop.first) %}
55
+ {{- '\n' }}
56
+ {%- endif %}
57
+ {%- if tool_call.function %}
58
+ {%- set tool_call = tool_call.function %}
59
+ {%- endif %}
60
+ {{- '<tool_call>\n{"name": "' }}
61
+ {{- tool_call.name }}
62
+ {{- '", "arguments": ' }}
63
+ {%- if tool_call.arguments is string %}
64
+ {{- tool_call.arguments }}
65
+ {%- else %}
66
+ {{- tool_call.arguments | tojson }}
67
+ {%- endif %}
68
+ {{- '}\n</tool_call>' }}
69
+ {%- endfor %}
70
+ {%- endif %}
71
+ {{- '<|im_end|>\n' }}
72
+ {%- elif message.role == "tool" %}
73
+ {%- if loop.first or (messages[loop.index0 - 1].role != "tool") %}
74
+ {{- '<|im_start|>user' }}
75
+ {%- endif %}
76
+ {{- '\n<tool_response>\n' }}
77
+ {{- content }}
78
+ {{- '\n</tool_response>' }}
79
+ {%- if loop.last or (messages[loop.index0 + 1].role != "tool") %}
80
+ {{- '<|im_end|>\n' }}
81
+ {%- endif %}
82
+ {%- endif %}
83
+ {%- endfor %}
84
+ {%- if add_generation_prompt %}
85
+ {{- '<|im_start|>assistant\n' }}
86
+ {%- if enable_thinking is defined and enable_thinking is false %}
87
+ {{- '<think>\n\n</think>\n\n' }}
88
+ {%- endif %}
89
+ {%- endif %}
tokenizer/tokenizer.json ADDED
@@ -0,0 +1,3 @@
 
 
 
 
1
+ version https://git-lfs.github.com/spec/v1
2
+ oid sha256:be75606093db2094d7cd20f3c2f385c212750648bd6ea4fb2bf507a6a4c55506
3
+ size 11422650
tokenizer/tokenizer_config.json ADDED
@@ -0,0 +1,30 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ {
2
+ "add_prefix_space": false,
3
+ "backend": "tokenizers",
4
+ "bos_token": null,
5
+ "clean_up_tokenization_spaces": false,
6
+ "eos_token": "<|im_end|>",
7
+ "errors": "replace",
8
+ "extra_special_tokens": [
9
+ "<|im_start|>",
10
+ "<|im_end|>",
11
+ "<|object_ref_start|>",
12
+ "<|object_ref_end|>",
13
+ "<|box_start|>",
14
+ "<|box_end|>",
15
+ "<|quad_start|>",
16
+ "<|quad_end|>",
17
+ "<|vision_start|>",
18
+ "<|vision_end|>",
19
+ "<|vision_pad|>",
20
+ "<|image_pad|>",
21
+ "<|video_pad|>"
22
+ ],
23
+ "is_local": true,
24
+ "local_files_only": true,
25
+ "model_max_length": 131072,
26
+ "pad_token": "<|endoftext|>",
27
+ "split_special_tokens": false,
28
+ "tokenizer_class": "Qwen2Tokenizer",
29
+ "unk_token": null
30
+ }
variants/events_brier_seed17/backbone_config/config.json ADDED
@@ -0,0 +1,63 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ {
2
+ "architectures": [
3
+ "Qwen3ForCausalLM"
4
+ ],
5
+ "attention_bias": false,
6
+ "attention_dropout": 0.0,
7
+ "bos_token_id": 151643,
8
+ "dtype": "float32",
9
+ "eos_token_id": 151645,
10
+ "head_dim": 128,
11
+ "hidden_act": "silu",
12
+ "hidden_size": 1024,
13
+ "initializer_range": 0.02,
14
+ "intermediate_size": 3072,
15
+ "layer_types": [
16
+ "full_attention",
17
+ "full_attention",
18
+ "full_attention",
19
+ "full_attention",
20
+ "full_attention",
21
+ "full_attention",
22
+ "full_attention",
23
+ "full_attention",
24
+ "full_attention",
25
+ "full_attention",
26
+ "full_attention",
27
+ "full_attention",
28
+ "full_attention",
29
+ "full_attention",
30
+ "full_attention",
31
+ "full_attention",
32
+ "full_attention",
33
+ "full_attention",
34
+ "full_attention",
35
+ "full_attention",
36
+ "full_attention",
37
+ "full_attention",
38
+ "full_attention",
39
+ "full_attention",
40
+ "full_attention",
41
+ "full_attention",
42
+ "full_attention",
43
+ "full_attention"
44
+ ],
45
+ "max_position_embeddings": 40960,
46
+ "max_window_layers": 28,
47
+ "model_type": "qwen3",
48
+ "num_attention_heads": 16,
49
+ "num_hidden_layers": 28,
50
+ "num_key_value_heads": 8,
51
+ "pad_token_id": null,
52
+ "rms_norm_eps": 1e-06,
53
+ "rope_parameters": {
54
+ "rope_theta": 1000000,
55
+ "rope_type": "default"
56
+ },
57
+ "sliding_window": null,
58
+ "tie_word_embeddings": true,
59
+ "transformers_version": "5.17.0",
60
+ "use_cache": false,
61
+ "use_sliding_window": false,
62
+ "vocab_size": 151936
63
+ }
variants/events_brier_seed17/best.safetensors ADDED
@@ -0,0 +1,3 @@
 
 
 
 
1
+ version https://git-lfs.github.com/spec/v1
2
+ oid sha256:0a8d8b14c97743eed4310d461a7ae5cff0400cb87eaa6dc1f8c5dc3c59988651
3
+ size 2385039280
variants/events_brier_seed17/config.json ADDED
@@ -0,0 +1,69 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ {
2
+ "input": "data/scaled_games_v4/events",
3
+ "output_dir": "runs/events_brier_seed17",
4
+ "model": "Qwen/Qwen3-0.6B",
5
+ "revision": "main",
6
+ "init_checkpoint": "/home/rwang/openjev_codex_20260917/runs/v3_teacher_coords_multi_seed17",
7
+ "objective": "observed_outcome",
8
+ "loss": "brier",
9
+ "reward_samples": 32,
10
+ "gradient_checkpointing": true,
11
+ "set_head": "attention",
12
+ "steps": 100,
13
+ "head_steps": 0,
14
+ "batch_questions": 16,
15
+ "microbatch_questions": 4,
16
+ "max_microbatch_tokens": 16384,
17
+ "eval_every": 50,
18
+ "max_length": 8192,
19
+ "seed": 17,
20
+ "backbone_lr": 2e-05,
21
+ "head_lr": 0.0002,
22
+ "head_warmup_lr": 0.001,
23
+ "precision": "bf16",
24
+ "disable_native_triton": true,
25
+ "validate_only": false,
26
+ "self_check": false,
27
+ "schema_version": "openjev-decision-pipeline-v1",
28
+ "resolved_model_revision": "c1899de289a04d12100db370d81485cdf75e47ca",
29
+ "initialization": "local DecisionModel warm start, fresh optimizer",
30
+ "data_sha256": {
31
+ "data/scaled_games_v4/events/train.jsonl": "4cebf698e4b71dc37d4d0e985b6577dd565ac370ac82b6f47f375aa6af69f8c9",
32
+ "data/scaled_games_v4/events/dev.jsonl": "0c352ee1064a5fb5d2aacad59cbea57afb897886d3d09e5ca108c866a5f893db",
33
+ "data/scaled_games_v4/events/calibration.jsonl": "4a8a8a2b81c3c46df6bb4c8690e97374dd0165cbef40566bb7f87c1f425d5562",
34
+ "data/scaled_games_v4/events/test.jsonl": "884bec71891668199925809d8708347f56f7bf06e8363e4a436347967b1423e3",
35
+ "data/scaled_games_v4/events/ood.jsonl": "3c00d845524634ca3b34559761cdaf0c52b7b7633d5517d1e7ee7f31697cea48"
36
+ },
37
+ "deps": {
38
+ "torch": "2.14.0",
39
+ "transformers": "5.17.0",
40
+ "safetensors": "0.8.0"
41
+ },
42
+ "gpu": "NVIDIA A100-SXM4-80GB",
43
+ "parameter_storage": "float32",
44
+ "forward_autocast": "bfloat16",
45
+ "parameter_count": 596250498,
46
+ "train_questions": 1124,
47
+ "all_train_questions": 1124,
48
+ "schema_counts": {
49
+ "records": 2356,
50
+ "questions_by_split": {
51
+ "train": 1124,
52
+ "dev": 372,
53
+ "calibration": 368,
54
+ "test": 364,
55
+ "ood": 128
56
+ },
57
+ "eligible_by_split_objective": {
58
+ "calibration/gold_distribution": 368,
59
+ "dev/gold_distribution": 372,
60
+ "ood/gold_distribution": 128,
61
+ "test/gold_distribution": 364,
62
+ "train/gold_distribution": 1124
63
+ }
64
+ },
65
+ "loss_description": "per-question proper loss; complete candidate normalization",
66
+ "rlcd_candidate": null,
67
+ "selection": "minimum dev target CE; held-out test first evaluated after training and checkpoint selection",
68
+ "parallelism": "complete candidate paths batched in one backbone call per microbatch; no prefix sharing"
69
+ }
variants/events_brier_seed17/summary.json ADDED
@@ -0,0 +1,136 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ {
2
+ "best_step": 100,
3
+ "best_dev_target_ce": 0.4421807345122086,
4
+ "selected_on": "dev target CE",
5
+ "objective": "observed_outcome",
6
+ "metrics_by_split": {
7
+ "dev": {
8
+ "questions": 372,
9
+ "objective": "observed_outcome",
10
+ "eligible_questions": 372,
11
+ "target_ce": 0.4421807345122086,
12
+ "target_kl": 0.4421807345122086,
13
+ "target_tv": 0.2684592128219396,
14
+ "by_target_kind": {
15
+ "observed_outcome": {
16
+ "n": 372,
17
+ "ce": 0.4421807345122086,
18
+ "kl": 0.4421807345122086,
19
+ "tv": 0.2684592128219396
20
+ }
21
+ },
22
+ "independent_targets": {
23
+ "teacher": {
24
+ "n": 0,
25
+ "ce": 0.0,
26
+ "kl": 0.0,
27
+ "tv": 0.0
28
+ },
29
+ "gold_distribution": {
30
+ "n": 372,
31
+ "ce": 0.46751975308439736,
32
+ "kl": 0.19751716047686577,
33
+ "tv": 0.17476332997830224
34
+ }
35
+ },
36
+ "note": "question-mean distribution recovery; policy targets and reference argmax labels are not observed outcomes"
37
+ },
38
+ "calibration": {
39
+ "questions": 368,
40
+ "objective": "observed_outcome",
41
+ "eligible_questions": 368,
42
+ "target_ce": 0.4847698677293702,
43
+ "target_kl": 0.4847698677293702,
44
+ "target_tv": 0.28490406827388554,
45
+ "by_target_kind": {
46
+ "observed_outcome": {
47
+ "n": 368,
48
+ "ce": 0.4847698677293702,
49
+ "kl": 0.4847698677293702,
50
+ "tv": 0.28490406827388554
51
+ }
52
+ },
53
+ "independent_targets": {
54
+ "teacher": {
55
+ "n": 0,
56
+ "ce": 0.0,
57
+ "kl": 0.0,
58
+ "tv": 0.0
59
+ },
60
+ "gold_distribution": {
61
+ "n": 368,
62
+ "ce": 0.4790668897887141,
63
+ "kl": 0.19277057070800374,
64
+ "tv": 0.17281581351777625
65
+ }
66
+ },
67
+ "note": "question-mean distribution recovery; policy targets and reference argmax labels are not observed outcomes"
68
+ },
69
+ "test": {
70
+ "questions": 364,
71
+ "objective": "observed_outcome",
72
+ "eligible_questions": 364,
73
+ "target_ce": 0.45052489631248066,
74
+ "target_kl": 0.45052489631248066,
75
+ "target_tv": 0.2634734042022083,
76
+ "by_target_kind": {
77
+ "observed_outcome": {
78
+ "n": 364,
79
+ "ce": 0.45052489631248066,
80
+ "kl": 0.45052489631248066,
81
+ "tv": 0.2634734042022083
82
+ }
83
+ },
84
+ "independent_targets": {
85
+ "teacher": {
86
+ "n": 0,
87
+ "ce": 0.0,
88
+ "kl": 0.0,
89
+ "tv": 0.0
90
+ },
91
+ "gold_distribution": {
92
+ "n": 364,
93
+ "ce": 0.46949283429897337,
94
+ "kl": 0.208722033972855,
95
+ "tv": 0.1715940341122606
96
+ }
97
+ },
98
+ "note": "question-mean distribution recovery; policy targets and reference argmax labels are not observed outcomes"
99
+ },
100
+ "ood": {
101
+ "questions": 128,
102
+ "objective": "observed_outcome",
103
+ "eligible_questions": 128,
104
+ "target_ce": 0.3166816048577138,
105
+ "target_kl": 0.3166816048577138,
106
+ "target_tv": 0.21465602967029523,
107
+ "by_target_kind": {
108
+ "observed_outcome": {
109
+ "n": 128,
110
+ "ce": 0.3166816048577138,
111
+ "kl": 0.3166816048577138,
112
+ "tv": 0.21465602967029523
113
+ }
114
+ },
115
+ "independent_targets": {
116
+ "teacher": {
117
+ "n": 0,
118
+ "ce": 0.0,
119
+ "kl": 0.0,
120
+ "tv": 0.0
121
+ },
122
+ "gold_distribution": {
123
+ "n": 128,
124
+ "ce": 0.3313783618564118,
125
+ "kl": 0.0978911402197844,
126
+ "tv": 0.11505404497567416
127
+ }
128
+ },
129
+ "note": "question-mean distribution recovery; policy targets and reference argmax labels are not observed outcomes"
130
+ }
131
+ },
132
+ "temperature": 1.0,
133
+ "temperature_fitted": false,
134
+ "training_seconds": 94.21064267028123,
135
+ "max_gpu_allocated_gb": 11.964137984
136
+ }
variants/events_brier_seed17/target_audit.json ADDED
The diff for this file is too large to render. See raw diff
 
variants/events_brier_seed17/tokenizer/chat_template.jinja ADDED
@@ -0,0 +1,89 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ {%- if tools %}
2
+ {{- '<|im_start|>system\n' }}
3
+ {%- if messages[0].role == 'system' %}
4
+ {{- messages[0].content + '\n\n' }}
5
+ {%- endif %}
6
+ {{- "# Tools\n\nYou may call one or more functions to assist with the user query.\n\nYou are provided with function signatures within <tools></tools> XML tags:\n<tools>" }}
7
+ {%- for tool in tools %}
8
+ {{- "\n" }}
9
+ {{- tool | tojson }}
10
+ {%- endfor %}
11
+ {{- "\n</tools>\n\nFor each function call, return a json object with function name and arguments within <tool_call></tool_call> XML tags:\n<tool_call>\n{\"name\": <function-name>, \"arguments\": <args-json-object>}\n</tool_call><|im_end|>\n" }}
12
+ {%- else %}
13
+ {%- if messages[0].role == 'system' %}
14
+ {{- '<|im_start|>system\n' + messages[0].content + '<|im_end|>\n' }}
15
+ {%- endif %}
16
+ {%- endif %}
17
+ {%- set ns = namespace(multi_step_tool=true, last_query_index=messages|length - 1) %}
18
+ {%- for message in messages[::-1] %}
19
+ {%- set index = (messages|length - 1) - loop.index0 %}
20
+ {%- if ns.multi_step_tool and message.role == "user" and message.content is string and not(message.content.startswith('<tool_response>') and message.content.endswith('</tool_response>')) %}
21
+ {%- set ns.multi_step_tool = false %}
22
+ {%- set ns.last_query_index = index %}
23
+ {%- endif %}
24
+ {%- endfor %}
25
+ {%- for message in messages %}
26
+ {%- if message.content is string %}
27
+ {%- set content = message.content %}
28
+ {%- else %}
29
+ {%- set content = '' %}
30
+ {%- endif %}
31
+ {%- if (message.role == "user") or (message.role == "system" and not loop.first) %}
32
+ {{- '<|im_start|>' + message.role + '\n' + content + '<|im_end|>' + '\n' }}
33
+ {%- elif message.role == "assistant" %}
34
+ {%- set reasoning_content = '' %}
35
+ {%- if message.reasoning_content is string %}
36
+ {%- set reasoning_content = message.reasoning_content %}
37
+ {%- else %}
38
+ {%- if '</think>' in content %}
39
+ {%- set reasoning_content = content.split('</think>')[0].rstrip('\n').split('<think>')[-1].lstrip('\n') %}
40
+ {%- set content = content.split('</think>')[-1].lstrip('\n') %}
41
+ {%- endif %}
42
+ {%- endif %}
43
+ {%- if loop.index0 > ns.last_query_index %}
44
+ {%- if loop.last or (not loop.last and reasoning_content) %}
45
+ {{- '<|im_start|>' + message.role + '\n<think>\n' + reasoning_content.strip('\n') + '\n</think>\n\n' + content.lstrip('\n') }}
46
+ {%- else %}
47
+ {{- '<|im_start|>' + message.role + '\n' + content }}
48
+ {%- endif %}
49
+ {%- else %}
50
+ {{- '<|im_start|>' + message.role + '\n' + content }}
51
+ {%- endif %}
52
+ {%- if message.tool_calls %}
53
+ {%- for tool_call in message.tool_calls %}
54
+ {%- if (loop.first and content) or (not loop.first) %}
55
+ {{- '\n' }}
56
+ {%- endif %}
57
+ {%- if tool_call.function %}
58
+ {%- set tool_call = tool_call.function %}
59
+ {%- endif %}
60
+ {{- '<tool_call>\n{"name": "' }}
61
+ {{- tool_call.name }}
62
+ {{- '", "arguments": ' }}
63
+ {%- if tool_call.arguments is string %}
64
+ {{- tool_call.arguments }}
65
+ {%- else %}
66
+ {{- tool_call.arguments | tojson }}
67
+ {%- endif %}
68
+ {{- '}\n</tool_call>' }}
69
+ {%- endfor %}
70
+ {%- endif %}
71
+ {{- '<|im_end|>\n' }}
72
+ {%- elif message.role == "tool" %}
73
+ {%- if loop.first or (messages[loop.index0 - 1].role != "tool") %}
74
+ {{- '<|im_start|>user' }}
75
+ {%- endif %}
76
+ {{- '\n<tool_response>\n' }}
77
+ {{- content }}
78
+ {{- '\n</tool_response>' }}
79
+ {%- if loop.last or (messages[loop.index0 + 1].role != "tool") %}
80
+ {{- '<|im_end|>\n' }}
81
+ {%- endif %}
82
+ {%- endif %}
83
+ {%- endfor %}
84
+ {%- if add_generation_prompt %}
85
+ {{- '<|im_start|>assistant\n' }}
86
+ {%- if enable_thinking is defined and enable_thinking is false %}
87
+ {{- '<think>\n\n</think>\n\n' }}
88
+ {%- endif %}
89
+ {%- endif %}
variants/events_brier_seed17/tokenizer/tokenizer.json ADDED
@@ -0,0 +1,3 @@
 
 
 
 
1
+ version https://git-lfs.github.com/spec/v1
2
+ oid sha256:be75606093db2094d7cd20f3c2f385c212750648bd6ea4fb2bf507a6a4c55506
3
+ size 11422650
variants/events_brier_seed17/tokenizer/tokenizer_config.json ADDED
@@ -0,0 +1,30 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ {
2
+ "add_prefix_space": false,
3
+ "backend": "tokenizers",
4
+ "bos_token": null,
5
+ "clean_up_tokenization_spaces": false,
6
+ "eos_token": "<|im_end|>",
7
+ "errors": "replace",
8
+ "extra_special_tokens": [
9
+ "<|im_start|>",
10
+ "<|im_end|>",
11
+ "<|object_ref_start|>",
12
+ "<|object_ref_end|>",
13
+ "<|box_start|>",
14
+ "<|box_end|>",
15
+ "<|quad_start|>",
16
+ "<|quad_end|>",
17
+ "<|vision_start|>",
18
+ "<|vision_end|>",
19
+ "<|vision_pad|>",
20
+ "<|image_pad|>",
21
+ "<|video_pad|>"
22
+ ],
23
+ "is_local": true,
24
+ "local_files_only": true,
25
+ "model_max_length": 131072,
26
+ "pad_token": "<|endoftext|>",
27
+ "split_special_tokens": false,
28
+ "tokenizer_class": "Qwen2Tokenizer",
29
+ "unk_token": null
30
+ }
variants/events_brier_seed17/train_log.json ADDED
@@ -0,0 +1,964 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ [
2
+ {
3
+ "step": 1,
4
+ "phase": "full",
5
+ "loss": 0.47767234593629837,
6
+ "questions": 16,
7
+ "microbatches": 4,
8
+ "elapsed_seconds": 1.299889337271452,
9
+ "batch_question_ids_sha256": "9815e3084b0060ca1bf3f53e09caa2e8ba2ecfbfa18cd5a90c1b6ca1dc29fbcb"
10
+ },
11
+ {
12
+ "step": 2,
13
+ "phase": "full",
14
+ "loss": 0.4745664047077298,
15
+ "questions": 16,
16
+ "microbatches": 4,
17
+ "elapsed_seconds": 2.020658330991864,
18
+ "batch_question_ids_sha256": "0766e00bccd0b9d63535226a2a7454d88d591d6a1adb32befcd64d40174bd87c"
19
+ },
20
+ {
21
+ "step": 3,
22
+ "phase": "full",
23
+ "loss": 0.3339335899800062,
24
+ "questions": 16,
25
+ "microbatches": 4,
26
+ "elapsed_seconds": 2.7444814248010516,
27
+ "batch_question_ids_sha256": "3eb8fb253ed8d59b98fd13fb12ae3aa08473ff98bf07f26c7537d8d4c6999318"
28
+ },
29
+ {
30
+ "step": 4,
31
+ "phase": "full",
32
+ "loss": 0.30728491954505444,
33
+ "questions": 16,
34
+ "microbatches": 4,
35
+ "elapsed_seconds": 3.4489749232307076,
36
+ "batch_question_ids_sha256": "b72b9555459b9660b8cfea9c061b0d69e21b213181cf19973fa9312c1d740e4c"
37
+ },
38
+ {
39
+ "step": 5,
40
+ "phase": "full",
41
+ "loss": 0.3180481344461441,
42
+ "questions": 16,
43
+ "microbatches": 4,
44
+ "elapsed_seconds": 4.1599197732284665,
45
+ "batch_question_ids_sha256": "851370879eecc2bb2b872bf9f26a90b0c422dbb910faa7ab99625462dd70126c"
46
+ },
47
+ {
48
+ "step": 6,
49
+ "phase": "full",
50
+ "loss": 0.31872232258319855,
51
+ "questions": 16,
52
+ "microbatches": 4,
53
+ "elapsed_seconds": 4.87426084280014,
54
+ "batch_question_ids_sha256": "a44ac9fef40754316a48915525e55ee3c8c5f613d2a3a206a70c53f6a360d6a1"
55
+ },
56
+ {
57
+ "step": 7,
58
+ "phase": "full",
59
+ "loss": 0.2807605154812336,
60
+ "questions": 16,
61
+ "microbatches": 4,
62
+ "elapsed_seconds": 5.582096985541284,
63
+ "batch_question_ids_sha256": "dd15d5df027d50d8bc83cc139ce58f36adf2613954b88024320afffaf4f275ce"
64
+ },
65
+ {
66
+ "step": 8,
67
+ "phase": "full",
68
+ "loss": 0.3625841060420498,
69
+ "questions": 16,
70
+ "microbatches": 4,
71
+ "elapsed_seconds": 6.32407046854496,
72
+ "batch_question_ids_sha256": "b16e21ce00030d172555d2ed47f3935798c5293130655f2d4bb114fab8d4f82e"
73
+ },
74
+ {
75
+ "step": 9,
76
+ "phase": "full",
77
+ "loss": 0.1536292051896453,
78
+ "questions": 16,
79
+ "microbatches": 4,
80
+ "elapsed_seconds": 7.142426532693207,
81
+ "batch_question_ids_sha256": "db13a8f636762d4e5c56f60dcecdc4a04f7acccee36ed4846fe242c110813e33"
82
+ },
83
+ {
84
+ "step": 10,
85
+ "phase": "full",
86
+ "loss": 0.4065534844994545,
87
+ "questions": 16,
88
+ "microbatches": 4,
89
+ "elapsed_seconds": 7.96108222194016,
90
+ "batch_question_ids_sha256": "d29a2ab7bae9372a6460e6d41d2320b0a13b55633fcc21e2a06f06a64a2b8f22"
91
+ },
92
+ {
93
+ "step": 11,
94
+ "phase": "full",
95
+ "loss": 0.2941920473240316,
96
+ "questions": 16,
97
+ "microbatches": 4,
98
+ "elapsed_seconds": 8.778865645639598,
99
+ "batch_question_ids_sha256": "e435c49e7cce06b5c387742c1cf2eae1fb3f57349a03ff3229713bf8e9179b7f"
100
+ },
101
+ {
102
+ "step": 12,
103
+ "phase": "full",
104
+ "loss": 0.5255483770743012,
105
+ "questions": 16,
106
+ "microbatches": 4,
107
+ "elapsed_seconds": 9.569397198967636,
108
+ "batch_question_ids_sha256": "52067a44724ec88a3f85cace2ddba01f511bb396cbf27877697e85efd4f0c68e"
109
+ },
110
+ {
111
+ "step": 13,
112
+ "phase": "full",
113
+ "loss": 0.18976907199248672,
114
+ "questions": 16,
115
+ "microbatches": 4,
116
+ "elapsed_seconds": 10.306229672394693,
117
+ "batch_question_ids_sha256": "d01e765c3bf9b2b1edf10d82d3b784f51514a21682d040badd6d9c1c29c3069f"
118
+ },
119
+ {
120
+ "step": 14,
121
+ "phase": "full",
122
+ "loss": 0.30221015959978104,
123
+ "questions": 16,
124
+ "microbatches": 4,
125
+ "elapsed_seconds": 11.033171433024108,
126
+ "batch_question_ids_sha256": "f7d76860a7e7abbd079278179b2a6054aa8b1bf90c6e38db91f414cf906eff6a"
127
+ },
128
+ {
129
+ "step": 15,
130
+ "phase": "full",
131
+ "loss": 0.3888988718390465,
132
+ "questions": 16,
133
+ "microbatches": 4,
134
+ "elapsed_seconds": 11.780395239591599,
135
+ "batch_question_ids_sha256": "1d089d97e3e6836ec1771dba1776d38f6a8ca414f96cc3c25fd0053ed3e27a44"
136
+ },
137
+ {
138
+ "step": 16,
139
+ "phase": "full",
140
+ "loss": 0.2125754952430725,
141
+ "questions": 16,
142
+ "microbatches": 4,
143
+ "elapsed_seconds": 12.491722567006946,
144
+ "batch_question_ids_sha256": "abd204307499e62673daa334cd3a4533448a0ef971b517e7acbf23783905ae91"
145
+ },
146
+ {
147
+ "step": 17,
148
+ "phase": "full",
149
+ "loss": 0.2280114609748125,
150
+ "questions": 16,
151
+ "microbatches": 4,
152
+ "elapsed_seconds": 13.19165222439915,
153
+ "batch_question_ids_sha256": "6d4029fa8044653ff0ccbe80e4371cccd55b6fc7c6462fb81d77f565bde1e2f5"
154
+ },
155
+ {
156
+ "step": 18,
157
+ "phase": "full",
158
+ "loss": 0.37986432015895844,
159
+ "questions": 16,
160
+ "microbatches": 4,
161
+ "elapsed_seconds": 13.910702037625015,
162
+ "batch_question_ids_sha256": "88ddecb0e43b996acf8e321896454ceb6e79095126d68380cfdc7c74ba66d04a"
163
+ },
164
+ {
165
+ "step": 19,
166
+ "phase": "full",
167
+ "loss": 0.3772949390113354,
168
+ "questions": 16,
169
+ "microbatches": 4,
170
+ "elapsed_seconds": 14.622391032986343,
171
+ "batch_question_ids_sha256": "3ba3b72125f7617c5cb4dfc031843bc10773d8aa7747ef6eef76ec813abc7a11"
172
+ },
173
+ {
174
+ "step": 20,
175
+ "phase": "full",
176
+ "loss": 0.47917328029870987,
177
+ "questions": 16,
178
+ "microbatches": 4,
179
+ "elapsed_seconds": 15.332007034681737,
180
+ "batch_question_ids_sha256": "a8fbfc400a3dd61e2468ddde908d8e6914e8742e66df75fa209575bbd49f674f"
181
+ },
182
+ {
183
+ "step": 21,
184
+ "phase": "full",
185
+ "loss": 0.2207504389807582,
186
+ "questions": 16,
187
+ "microbatches": 4,
188
+ "elapsed_seconds": 16.107440183870494,
189
+ "batch_question_ids_sha256": "49a7b1f6bb6891cd06f7c495942bbdbd98cfbd7435f49d77922ab364b9c1c5f5"
190
+ },
191
+ {
192
+ "step": 22,
193
+ "phase": "full",
194
+ "loss": 0.2577735111117363,
195
+ "questions": 16,
196
+ "microbatches": 4,
197
+ "elapsed_seconds": 16.850639042444527,
198
+ "batch_question_ids_sha256": "d93e27a20a5666b8fe2ea5489dac0a58f99a770bfb0d4ee6870d36621ed78dee"
199
+ },
200
+ {
201
+ "step": 23,
202
+ "phase": "full",
203
+ "loss": 0.36209800839424133,
204
+ "questions": 16,
205
+ "microbatches": 4,
206
+ "elapsed_seconds": 17.691069215536118,
207
+ "batch_question_ids_sha256": "88845418c486e4c7d856761b8aca680a7ca52ec59c84219e708745af76a445c2"
208
+ },
209
+ {
210
+ "step": 24,
211
+ "phase": "full",
212
+ "loss": 0.2882579481229186,
213
+ "questions": 16,
214
+ "microbatches": 4,
215
+ "elapsed_seconds": 18.40779590792954,
216
+ "batch_question_ids_sha256": "b7fcb8d686d0a87ea4fd6ded60d595c7ccc4b132fc739e047564d72f5944577f"
217
+ },
218
+ {
219
+ "step": 25,
220
+ "phase": "full",
221
+ "loss": 0.269174687564373,
222
+ "questions": 16,
223
+ "microbatches": 4,
224
+ "elapsed_seconds": 19.135179361328483,
225
+ "batch_question_ids_sha256": "b2e6af990ea722e6cb276f57a736c23766306ae29b211bf00447d868b434c962"
226
+ },
227
+ {
228
+ "step": 26,
229
+ "phase": "full",
230
+ "loss": 0.34036025404930115,
231
+ "questions": 16,
232
+ "microbatches": 4,
233
+ "elapsed_seconds": 19.86950681824237,
234
+ "batch_question_ids_sha256": "67253956a0835f33db1fa53443fed54476a027338e0ffdd7726dcb48fbef1c9a"
235
+ },
236
+ {
237
+ "step": 27,
238
+ "phase": "full",
239
+ "loss": 0.33926529437303543,
240
+ "questions": 16,
241
+ "microbatches": 4,
242
+ "elapsed_seconds": 20.584741784259677,
243
+ "batch_question_ids_sha256": "b37cd333ba07e45e13875704fb0cc42b2f44fc5317737590bb27d88a7ce61452"
244
+ },
245
+ {
246
+ "step": 28,
247
+ "phase": "full",
248
+ "loss": 0.3321617506444454,
249
+ "questions": 16,
250
+ "microbatches": 4,
251
+ "elapsed_seconds": 21.315836602821946,
252
+ "batch_question_ids_sha256": "d2a5a95e218ea89d7469dfdd8ff320de68f88ff52171675c0ccf7c9c3d95a891"
253
+ },
254
+ {
255
+ "step": 29,
256
+ "phase": "full",
257
+ "loss": 0.2353851981461048,
258
+ "questions": 16,
259
+ "microbatches": 4,
260
+ "elapsed_seconds": 22.03636402450502,
261
+ "batch_question_ids_sha256": "1c6aa40b463139c53b7a93df96c3a1db9c53fac3123be7e8dda4fc8f53bc7996"
262
+ },
263
+ {
264
+ "step": 30,
265
+ "phase": "full",
266
+ "loss": 0.11482883733697236,
267
+ "questions": 16,
268
+ "microbatches": 4,
269
+ "elapsed_seconds": 22.730047752149403,
270
+ "batch_question_ids_sha256": "be09ca5b88c61e0617249157b2d75545b12e403abaf2ff570be1e6d30856d742"
271
+ },
272
+ {
273
+ "step": 31,
274
+ "phase": "full",
275
+ "loss": 0.41507381945848465,
276
+ "questions": 16,
277
+ "microbatches": 4,
278
+ "elapsed_seconds": 23.457241214811802,
279
+ "batch_question_ids_sha256": "08f4544d1dfa624e78df3609c0059204a16146f8421bf85c14d0b10b5926b448"
280
+ },
281
+ {
282
+ "step": 32,
283
+ "phase": "full",
284
+ "loss": 0.330725422129035,
285
+ "questions": 16,
286
+ "microbatches": 4,
287
+ "elapsed_seconds": 24.159707821905613,
288
+ "batch_question_ids_sha256": "f8a0c95d4ce5a21d3b7bdd381eb7c9d15a251da10b33f0a8521a3e81eeb1450e"
289
+ },
290
+ {
291
+ "step": 33,
292
+ "phase": "full",
293
+ "loss": 0.2639482840895653,
294
+ "questions": 16,
295
+ "microbatches": 4,
296
+ "elapsed_seconds": 24.896084915846586,
297
+ "batch_question_ids_sha256": "4c396fcb15e0a31d52d00473dc2f3bee3767da23146fb5038b6e63d9ad5016c8"
298
+ },
299
+ {
300
+ "step": 34,
301
+ "phase": "full",
302
+ "loss": 0.26776977452391293,
303
+ "questions": 16,
304
+ "microbatches": 4,
305
+ "elapsed_seconds": 25.612479377537966,
306
+ "batch_question_ids_sha256": "c59ffed718f7ded62b1028779c65e8759de67070b26019e8642f75dd2e43a512"
307
+ },
308
+ {
309
+ "step": 35,
310
+ "phase": "full",
311
+ "loss": 0.43445201084250584,
312
+ "questions": 16,
313
+ "microbatches": 4,
314
+ "elapsed_seconds": 26.31534642726183,
315
+ "batch_question_ids_sha256": "0e54ce4db6221fac98c15ce77d50701153708881dda86adbcb2776637215c9bd"
316
+ },
317
+ {
318
+ "step": 36,
319
+ "phase": "full",
320
+ "loss": 0.5263467519544065,
321
+ "questions": 16,
322
+ "microbatches": 4,
323
+ "elapsed_seconds": 27.034690125845373,
324
+ "batch_question_ids_sha256": "24c257b893015e139aaec823c9fa8ed630a7d2df239552f229ebb832a7746723"
325
+ },
326
+ {
327
+ "step": 37,
328
+ "phase": "full",
329
+ "loss": 0.32995557998947334,
330
+ "questions": 16,
331
+ "microbatches": 4,
332
+ "elapsed_seconds": 27.750695827417076,
333
+ "batch_question_ids_sha256": "bc22d73d0ac85948fd9bbd0813992b093f9beb230652c6234447dbbbfb4add35"
334
+ },
335
+ {
336
+ "step": 38,
337
+ "phase": "full",
338
+ "loss": 0.28109178878366947,
339
+ "questions": 16,
340
+ "microbatches": 4,
341
+ "elapsed_seconds": 28.491634809412062,
342
+ "batch_question_ids_sha256": "6e47f7951fb42ce413ed2f4ce54d8832b534a3e9d0f2e35829a1608286227ef4"
343
+ },
344
+ {
345
+ "step": 39,
346
+ "phase": "full",
347
+ "loss": 0.256468391045928,
348
+ "questions": 16,
349
+ "microbatches": 4,
350
+ "elapsed_seconds": 29.180634811520576,
351
+ "batch_question_ids_sha256": "b5bfa6d3b4f6544cb42bc8b82b7c88ad3430367ce2f47dd0c3182d7f55e26796"
352
+ },
353
+ {
354
+ "step": 40,
355
+ "phase": "full",
356
+ "loss": 0.2841529995203018,
357
+ "questions": 16,
358
+ "microbatches": 4,
359
+ "elapsed_seconds": 29.919747012667358,
360
+ "batch_question_ids_sha256": "e356d8692763396b1f305f967bc42bbea222b3adcfd7355c8ffd489d7c4a8702"
361
+ },
362
+ {
363
+ "step": 41,
364
+ "phase": "full",
365
+ "loss": 0.3729940578341484,
366
+ "questions": 16,
367
+ "microbatches": 4,
368
+ "elapsed_seconds": 30.638639828190207,
369
+ "batch_question_ids_sha256": "eddf9bafc965ad35c8a9189e6a7b19eb336e5f2736ab490e5d0f7547f885b883"
370
+ },
371
+ {
372
+ "step": 42,
373
+ "phase": "full",
374
+ "loss": 0.20839229971170425,
375
+ "questions": 16,
376
+ "microbatches": 4,
377
+ "elapsed_seconds": 31.362661720253527,
378
+ "batch_question_ids_sha256": "93f507828c7fa8715902d87edef912b05fe2920c25f400745b5d4b1835ff1fcd"
379
+ },
380
+ {
381
+ "step": 43,
382
+ "phase": "full",
383
+ "loss": 0.27495690272189677,
384
+ "questions": 16,
385
+ "microbatches": 4,
386
+ "elapsed_seconds": 32.06642537377775,
387
+ "batch_question_ids_sha256": "50f72e83de9b1b9daf154f013aa536b1454d8d8c13b503ee6efa9007c721faa6"
388
+ },
389
+ {
390
+ "step": 44,
391
+ "phase": "full",
392
+ "loss": 0.48349061608314514,
393
+ "questions": 16,
394
+ "microbatches": 4,
395
+ "elapsed_seconds": 32.812288109213114,
396
+ "batch_question_ids_sha256": "71067cb1ca0d666b5a16e53227ae1840e468c0c39e879ef8004ac73520b396f9"
397
+ },
398
+ {
399
+ "step": 45,
400
+ "phase": "full",
401
+ "loss": 0.19684232398867607,
402
+ "questions": 16,
403
+ "microbatches": 4,
404
+ "elapsed_seconds": 33.537176745943725,
405
+ "batch_question_ids_sha256": "b9270e713e4fca02ecc7205c7ac1e320d3c87286eb7237ed0b02a4686b626bc4"
406
+ },
407
+ {
408
+ "step": 46,
409
+ "phase": "full",
410
+ "loss": 0.2759946044534445,
411
+ "questions": 16,
412
+ "microbatches": 4,
413
+ "elapsed_seconds": 34.26172788254917,
414
+ "batch_question_ids_sha256": "fbecc4d2e806b5b42f4f82a7e99465e1fceb5b7b61dfc3cd167654deeef897d0"
415
+ },
416
+ {
417
+ "step": 47,
418
+ "phase": "full",
419
+ "loss": 0.22207287698984146,
420
+ "questions": 16,
421
+ "microbatches": 4,
422
+ "elapsed_seconds": 34.97563951276243,
423
+ "batch_question_ids_sha256": "9fb1e3f8e3eb222e0e377eeedad5b0df90d11389465803b67a9b5090d2e79ca0"
424
+ },
425
+ {
426
+ "step": 48,
427
+ "phase": "full",
428
+ "loss": 0.2445692487526685,
429
+ "questions": 16,
430
+ "microbatches": 4,
431
+ "elapsed_seconds": 35.7044816268608,
432
+ "batch_question_ids_sha256": "8ec776449fb8174bd55b829feff0711ca000a9e09d903e741da6f91e2d21a136"
433
+ },
434
+ {
435
+ "step": 49,
436
+ "phase": "full",
437
+ "loss": 0.09901037975214422,
438
+ "questions": 16,
439
+ "microbatches": 4,
440
+ "elapsed_seconds": 36.49502827972174,
441
+ "batch_question_ids_sha256": "a20669c6cee5e157b8b54729bb88921f4ff6df89b4dda467ec25abb10f0e78da"
442
+ },
443
+ {
444
+ "step": 50,
445
+ "phase": "full",
446
+ "loss": 0.14326033811084926,
447
+ "questions": 16,
448
+ "microbatches": 4,
449
+ "elapsed_seconds": 37.1994747761637,
450
+ "batch_question_ids_sha256": "cab2a2f34731e8a959300614046a8b8fc4c926971dd5909a1a6e3de0ad3acb26",
451
+ "dev": {
452
+ "questions": 372,
453
+ "objective": "observed_outcome",
454
+ "eligible_questions": 372,
455
+ "target_ce": 0.5348790703109285,
456
+ "target_kl": 0.5348790703109285,
457
+ "target_tv": 0.30709812289697785,
458
+ "by_target_kind": {
459
+ "observed_outcome": {
460
+ "n": 372,
461
+ "ce": 0.5348790703109285,
462
+ "kl": 0.5348790703109285,
463
+ "tv": 0.30709812289697785
464
+ }
465
+ },
466
+ "independent_targets": {
467
+ "teacher": {
468
+ "n": 0,
469
+ "ce": 0.0,
470
+ "kl": 0.0,
471
+ "tv": 0.0
472
+ },
473
+ "gold_distribution": {
474
+ "n": 372,
475
+ "ce": 0.5338139083905662,
476
+ "kl": 0.26381131578303585,
477
+ "tv": 0.20872586212331487
478
+ }
479
+ },
480
+ "note": "question-mean distribution recovery; policy targets and reference argmax labels are not observed outcomes"
481
+ }
482
+ },
483
+ {
484
+ "step": 51,
485
+ "phase": "full",
486
+ "loss": 0.3137478260323405,
487
+ "questions": 16,
488
+ "microbatches": 4,
489
+ "elapsed_seconds": 43.16678562480956,
490
+ "batch_question_ids_sha256": "46758e1c6996a3f1a932a11e8372539949f8ac8f4c0415b36f65c713e195096f"
491
+ },
492
+ {
493
+ "step": 52,
494
+ "phase": "full",
495
+ "loss": 0.2717863286379725,
496
+ "questions": 16,
497
+ "microbatches": 4,
498
+ "elapsed_seconds": 43.874052685685456,
499
+ "batch_question_ids_sha256": "d87fcd81d136c657c13e49aa033db193bfbb121e4a721fa86c16aad4c0a7c39d"
500
+ },
501
+ {
502
+ "step": 53,
503
+ "phase": "full",
504
+ "loss": 0.28308067098259926,
505
+ "questions": 16,
506
+ "microbatches": 4,
507
+ "elapsed_seconds": 44.6021969569847,
508
+ "batch_question_ids_sha256": "4fae217aef8823502570d85ef34334cdfe2994ba567eef62990a60ac5e3ede4a"
509
+ },
510
+ {
511
+ "step": 54,
512
+ "phase": "full",
513
+ "loss": 0.6188238859176636,
514
+ "questions": 16,
515
+ "microbatches": 4,
516
+ "elapsed_seconds": 45.31346751283854,
517
+ "batch_question_ids_sha256": "edbcfec53ad977392c69172be6901bdb4eee1725bf8fd83582f64ac6320e77cd"
518
+ },
519
+ {
520
+ "step": 55,
521
+ "phase": "full",
522
+ "loss": 0.42757737822830677,
523
+ "questions": 16,
524
+ "microbatches": 4,
525
+ "elapsed_seconds": 46.016700938344,
526
+ "batch_question_ids_sha256": "7af8e2cece43ae333d4687839255ee144e5968fc9aa063efa2ff6dd96c42d1b2"
527
+ },
528
+ {
529
+ "step": 56,
530
+ "phase": "full",
531
+ "loss": 0.30507429130375385,
532
+ "questions": 16,
533
+ "microbatches": 4,
534
+ "elapsed_seconds": 46.7371188858524,
535
+ "batch_question_ids_sha256": "6d579036d49f401569b6e29499a444725fe7a64de0907bc07bd389fd352e615b"
536
+ },
537
+ {
538
+ "step": 57,
539
+ "phase": "full",
540
+ "loss": 0.23919871263206005,
541
+ "questions": 16,
542
+ "microbatches": 4,
543
+ "elapsed_seconds": 47.45829302165657,
544
+ "batch_question_ids_sha256": "848946e51c61b43a984fc49047cb8c9f0fae9f28c6447d8b4ee2c74dafe198bc"
545
+ },
546
+ {
547
+ "step": 58,
548
+ "phase": "full",
549
+ "loss": 0.3340693097561598,
550
+ "questions": 16,
551
+ "microbatches": 4,
552
+ "elapsed_seconds": 48.18450126796961,
553
+ "batch_question_ids_sha256": "7a207cb4e3ff6157802e3b560ec0b63a488a8ab0eabaa5bcdea45eeeb107d3da"
554
+ },
555
+ {
556
+ "step": 59,
557
+ "phase": "full",
558
+ "loss": 0.3097129948437214,
559
+ "questions": 16,
560
+ "microbatches": 4,
561
+ "elapsed_seconds": 48.893560361117125,
562
+ "batch_question_ids_sha256": "6ed68210f517203266a035e4dc739a5e91ad1773fb37b2d954aabbc97ce44cc5"
563
+ },
564
+ {
565
+ "step": 60,
566
+ "phase": "full",
567
+ "loss": 0.2643849514424801,
568
+ "questions": 16,
569
+ "microbatches": 4,
570
+ "elapsed_seconds": 49.75948201492429,
571
+ "batch_question_ids_sha256": "cbce464330f46f2f323539fb9d7962cede14e7074be588012e1dfa684881e6d4"
572
+ },
573
+ {
574
+ "step": 61,
575
+ "phase": "full",
576
+ "loss": 0.22575335949659348,
577
+ "questions": 16,
578
+ "microbatches": 4,
579
+ "elapsed_seconds": 50.51938905380666,
580
+ "batch_question_ids_sha256": "7212e07d655e85ec9b791cd6cde65ec3e9ebfa5224d1854c76d9a8e98cc07be3"
581
+ },
582
+ {
583
+ "step": 62,
584
+ "phase": "full",
585
+ "loss": 0.3371951300650835,
586
+ "questions": 16,
587
+ "microbatches": 4,
588
+ "elapsed_seconds": 51.23941683862358,
589
+ "batch_question_ids_sha256": "768e8960dc65d1f0cc4b13cc2c90dbfe8db2c685cceb3a8ed4f834e59a13a2f5"
590
+ },
591
+ {
592
+ "step": 63,
593
+ "phase": "full",
594
+ "loss": 0.4498938750475645,
595
+ "questions": 16,
596
+ "microbatches": 4,
597
+ "elapsed_seconds": 51.95728688593954,
598
+ "batch_question_ids_sha256": "ce0e1bf7240822fc1f373178c20604e9dba0409aa63f2b51415a5f4db012ea78"
599
+ },
600
+ {
601
+ "step": 64,
602
+ "phase": "full",
603
+ "loss": 0.3990790694952011,
604
+ "questions": 16,
605
+ "microbatches": 4,
606
+ "elapsed_seconds": 52.69237087201327,
607
+ "batch_question_ids_sha256": "6b8c351645894c06829c8dffb3045f3a3081c0fcb0b24c924b9576410c78d47c"
608
+ },
609
+ {
610
+ "step": 65,
611
+ "phase": "full",
612
+ "loss": 0.15212762681767344,
613
+ "questions": 16,
614
+ "microbatches": 4,
615
+ "elapsed_seconds": 53.410304352641106,
616
+ "batch_question_ids_sha256": "06de39fb8b5de5ce866344dcb10447ed159378857789cc67daf23061f43e3e7f"
617
+ },
618
+ {
619
+ "step": 66,
620
+ "phase": "full",
621
+ "loss": 0.25516666239127517,
622
+ "questions": 16,
623
+ "microbatches": 4,
624
+ "elapsed_seconds": 54.1299799894914,
625
+ "batch_question_ids_sha256": "1728e00f1c3d6bfc533f2891740148ba51b40dcb766709fed2a21437282f6ee1"
626
+ },
627
+ {
628
+ "step": 67,
629
+ "phase": "full",
630
+ "loss": 0.15473376447334886,
631
+ "questions": 16,
632
+ "microbatches": 4,
633
+ "elapsed_seconds": 54.84718861337751,
634
+ "batch_question_ids_sha256": "26b40bee228ef411ff536ea5c58c9a2fe4790fcc250d1ec97719f5c2b6051983"
635
+ },
636
+ {
637
+ "step": 68,
638
+ "phase": "full",
639
+ "loss": 0.5024608373641968,
640
+ "questions": 16,
641
+ "microbatches": 4,
642
+ "elapsed_seconds": 55.56789685878903,
643
+ "batch_question_ids_sha256": "ebcd0c069c1aad6db15af178bfb86b95232c842c724cae991202c127dddf4df4"
644
+ },
645
+ {
646
+ "step": 69,
647
+ "phase": "full",
648
+ "loss": 0.40726556070148945,
649
+ "questions": 16,
650
+ "microbatches": 4,
651
+ "elapsed_seconds": 56.28566949442029,
652
+ "batch_question_ids_sha256": "021f795640c8ebe5b45f63378c3fb38ba2c849f5d70b747f653482f046de3fba"
653
+ },
654
+ {
655
+ "step": 70,
656
+ "phase": "full",
657
+ "loss": 0.3981996960937977,
658
+ "questions": 16,
659
+ "microbatches": 4,
660
+ "elapsed_seconds": 57.01161473803222,
661
+ "batch_question_ids_sha256": "968367da65a754ce0d7eb72abb8cbb64b558c86aa9d8adda10e119723caa2bb9"
662
+ },
663
+ {
664
+ "step": 71,
665
+ "phase": "full",
666
+ "loss": 0.2851921394467354,
667
+ "questions": 16,
668
+ "microbatches": 4,
669
+ "elapsed_seconds": 57.74261846765876,
670
+ "batch_question_ids_sha256": "5f2d81f647af85f4a48f30464c10a61c6a85b7d5d35f687d527ade870b0c1e7e"
671
+ },
672
+ {
673
+ "step": 72,
674
+ "phase": "full",
675
+ "loss": 0.517420694231987,
676
+ "questions": 16,
677
+ "microbatches": 4,
678
+ "elapsed_seconds": 58.47589023038745,
679
+ "batch_question_ids_sha256": "6780ef3f330f6cd45a377b3f3c501ecc3ef2ae65d08769f207f4a68c152e3dbe"
680
+ },
681
+ {
682
+ "step": 73,
683
+ "phase": "full",
684
+ "loss": 0.22744558565318584,
685
+ "questions": 16,
686
+ "microbatches": 4,
687
+ "elapsed_seconds": 59.20974231977016,
688
+ "batch_question_ids_sha256": "d521c674f97a615228053a262b84518189f9580e2c73b32eee51bcef5ed76d29"
689
+ },
690
+ {
691
+ "step": 74,
692
+ "phase": "full",
693
+ "loss": 0.31863199174404144,
694
+ "questions": 16,
695
+ "microbatches": 4,
696
+ "elapsed_seconds": 60.064713726751506,
697
+ "batch_question_ids_sha256": "6c26d6bf906bf556525d65a2709f2f59a7cee21e2f62c8bad8c7015fcde12f89"
698
+ },
699
+ {
700
+ "step": 75,
701
+ "phase": "full",
702
+ "loss": 0.30666476860642433,
703
+ "questions": 16,
704
+ "microbatches": 4,
705
+ "elapsed_seconds": 60.784851206466556,
706
+ "batch_question_ids_sha256": "afabf643cddd0692428615c4aa81345acbbcdab1c52a5c9a48c815f9e2700755"
707
+ },
708
+ {
709
+ "step": 76,
710
+ "phase": "full",
711
+ "loss": 0.3913169279694557,
712
+ "questions": 16,
713
+ "microbatches": 4,
714
+ "elapsed_seconds": 61.525260652415454,
715
+ "batch_question_ids_sha256": "ba8ce0da2f777532f12642355dcce1c4d29f8e0865010f26b4bf03c56d7723d1"
716
+ },
717
+ {
718
+ "step": 77,
719
+ "phase": "full",
720
+ "loss": 0.2860779892653227,
721
+ "questions": 16,
722
+ "microbatches": 4,
723
+ "elapsed_seconds": 62.26424302812666,
724
+ "batch_question_ids_sha256": "9e0a4687654abac2570a41f8674ab95fcade2ac8e5e238d4661499dff7d58834"
725
+ },
726
+ {
727
+ "step": 78,
728
+ "phase": "full",
729
+ "loss": 0.34327320381999016,
730
+ "questions": 16,
731
+ "microbatches": 4,
732
+ "elapsed_seconds": 62.992976791225374,
733
+ "batch_question_ids_sha256": "e36ae8ba0c5eaffb1a534d69421fb9118e4f45453e90b83cbbf5f29ef284cafc"
734
+ },
735
+ {
736
+ "step": 79,
737
+ "phase": "full",
738
+ "loss": 0.39875966031104326,
739
+ "questions": 16,
740
+ "microbatches": 4,
741
+ "elapsed_seconds": 63.6979681737721,
742
+ "batch_question_ids_sha256": "171d14c1b2acfe9895a9353fe3dc68dcde6a998182a601a2746bf8fb10de20e3"
743
+ },
744
+ {
745
+ "step": 80,
746
+ "phase": "full",
747
+ "loss": 0.24940739758312702,
748
+ "questions": 16,
749
+ "microbatches": 4,
750
+ "elapsed_seconds": 64.45239959936589,
751
+ "batch_question_ids_sha256": "5091eab21841dd9384edd2c719867375591caf9d084ba03c065878eb451a5674"
752
+ },
753
+ {
754
+ "step": 81,
755
+ "phase": "full",
756
+ "loss": 0.34674376994371414,
757
+ "questions": 16,
758
+ "microbatches": 4,
759
+ "elapsed_seconds": 65.18292573653162,
760
+ "batch_question_ids_sha256": "f94cae69521603c0b60eb1cc6bb5370d2ea3d089f5b6bfd9a260b41a766ff974"
761
+ },
762
+ {
763
+ "step": 82,
764
+ "phase": "full",
765
+ "loss": 0.24664756562560797,
766
+ "questions": 16,
767
+ "microbatches": 4,
768
+ "elapsed_seconds": 65.89143331162632,
769
+ "batch_question_ids_sha256": "f804d3278fd4930c24d09af7cc2e1fc70170c35cac45cbed9779d59f2d765e08"
770
+ },
771
+ {
772
+ "step": 83,
773
+ "phase": "full",
774
+ "loss": 0.242635240778327,
775
+ "questions": 16,
776
+ "microbatches": 4,
777
+ "elapsed_seconds": 66.62519173696637,
778
+ "batch_question_ids_sha256": "19b60c4f93e1ba7368b1f066c22c4ba32b8b66981562b2cbe2485da7433c07aa"
779
+ },
780
+ {
781
+ "step": 84,
782
+ "phase": "full",
783
+ "loss": 0.24826265033334494,
784
+ "questions": 16,
785
+ "microbatches": 4,
786
+ "elapsed_seconds": 67.33410876244307,
787
+ "batch_question_ids_sha256": "ebdf0b28c90afc00f6691d1a09af18f9f66ebaf71377d0ba210fc4892869014c"
788
+ },
789
+ {
790
+ "step": 85,
791
+ "phase": "full",
792
+ "loss": 0.2863457463681698,
793
+ "questions": 16,
794
+ "microbatches": 4,
795
+ "elapsed_seconds": 68.0715844547376,
796
+ "batch_question_ids_sha256": "4f04bf635d321e106f9856ba99a47fd6c2c033ae0450aa32c6dfe595a74ae4ed"
797
+ },
798
+ {
799
+ "step": 86,
800
+ "phase": "full",
801
+ "loss": 0.18521394208073616,
802
+ "questions": 16,
803
+ "microbatches": 4,
804
+ "elapsed_seconds": 68.79707364458591,
805
+ "batch_question_ids_sha256": "ba9c22de6939f9c021221fe5a43ff5f50b15a70b21ed4effe8d97caf79b049ca"
806
+ },
807
+ {
808
+ "step": 87,
809
+ "phase": "full",
810
+ "loss": 0.18474309891462326,
811
+ "questions": 16,
812
+ "microbatches": 4,
813
+ "elapsed_seconds": 69.51761453226209,
814
+ "batch_question_ids_sha256": "c9f35e67c5374cdfd4ecee262a8723f0f688d0387f073072324c9d259d9f4323"
815
+ },
816
+ {
817
+ "step": 88,
818
+ "phase": "full",
819
+ "loss": 0.3266881722956896,
820
+ "questions": 16,
821
+ "microbatches": 4,
822
+ "elapsed_seconds": 70.239103092812,
823
+ "batch_question_ids_sha256": "64ab5437c436a5c934bb05696c9b73f14c9838d3b210eeeb1114438303808400"
824
+ },
825
+ {
826
+ "step": 89,
827
+ "phase": "full",
828
+ "loss": 0.16573298256844282,
829
+ "questions": 16,
830
+ "microbatches": 4,
831
+ "elapsed_seconds": 70.9650420229882,
832
+ "batch_question_ids_sha256": "6ff06537c7c8217ebac7a64762ff5ee340f367a0f7652d25d5cd799c0d15cbac"
833
+ },
834
+ {
835
+ "step": 90,
836
+ "phase": "full",
837
+ "loss": 0.26366132125258446,
838
+ "questions": 16,
839
+ "microbatches": 4,
840
+ "elapsed_seconds": 71.69922364503145,
841
+ "batch_question_ids_sha256": "8fd09d36031bf8ee878da3b38a0694c5e416636fcd74e603f03dd205a9a3ab33"
842
+ },
843
+ {
844
+ "step": 91,
845
+ "phase": "full",
846
+ "loss": 0.2215642500668764,
847
+ "questions": 16,
848
+ "microbatches": 4,
849
+ "elapsed_seconds": 72.42362531926483,
850
+ "batch_question_ids_sha256": "322fb22dffd39fb986e5bfd20ac78a9963d2d5b01978598e3ea9ff285c9b09d3"
851
+ },
852
+ {
853
+ "step": 92,
854
+ "phase": "full",
855
+ "loss": 0.2945997603237629,
856
+ "questions": 16,
857
+ "microbatches": 4,
858
+ "elapsed_seconds": 73.12151438836008,
859
+ "batch_question_ids_sha256": "d0417dd44fb2608169bd7f2ab17d57acd203be7f8242f3c2e89ad317009a2e56"
860
+ },
861
+ {
862
+ "step": 93,
863
+ "phase": "full",
864
+ "loss": 0.1663272229488939,
865
+ "questions": 16,
866
+ "microbatches": 4,
867
+ "elapsed_seconds": 73.81322857644409,
868
+ "batch_question_ids_sha256": "57be67994833e0a92fac3ba6c7b6b1fe3b8383d53284136aad0b9a926f82aa1f"
869
+ },
870
+ {
871
+ "step": 94,
872
+ "phase": "full",
873
+ "loss": 0.17362246103584766,
874
+ "questions": 16,
875
+ "microbatches": 4,
876
+ "elapsed_seconds": 74.53701240383089,
877
+ "batch_question_ids_sha256": "9c3cbfa9806ac2c618d0a1e3ca3ef41c91b3a86c8aa45a36071383124ee71262"
878
+ },
879
+ {
880
+ "step": 95,
881
+ "phase": "full",
882
+ "loss": 0.23721570521593094,
883
+ "questions": 16,
884
+ "microbatches": 4,
885
+ "elapsed_seconds": 75.25338559411466,
886
+ "batch_question_ids_sha256": "100fb8785fb8c78e98988f555871d3d9a4ed2716c9bff93597e8a64de5e2db46"
887
+ },
888
+ {
889
+ "step": 96,
890
+ "phase": "full",
891
+ "loss": 0.19234935194253922,
892
+ "questions": 16,
893
+ "microbatches": 4,
894
+ "elapsed_seconds": 75.98261929582804,
895
+ "batch_question_ids_sha256": "c781c81a212d6bc7153d4ecdfb49ff5f6570c3bde61d9f7babaa983e433d7348"
896
+ },
897
+ {
898
+ "step": 97,
899
+ "phase": "full",
900
+ "loss": 0.271314961835742,
901
+ "questions": 16,
902
+ "microbatches": 4,
903
+ "elapsed_seconds": 76.7082808027044,
904
+ "batch_question_ids_sha256": "8b19c5c7751e07c22469266bfa40bb1eaac4670fe203a8c3e8a289350956893c"
905
+ },
906
+ {
907
+ "step": 98,
908
+ "phase": "full",
909
+ "loss": 0.346242094412446,
910
+ "questions": 16,
911
+ "microbatches": 4,
912
+ "elapsed_seconds": 77.43025597464293,
913
+ "batch_question_ids_sha256": "043573f3a457c1f8cfbcefabfef11e430a1dd573e779dd7e19f7dc26a4a1ebad"
914
+ },
915
+ {
916
+ "step": 99,
917
+ "phase": "full",
918
+ "loss": 0.2261855776887387,
919
+ "questions": 16,
920
+ "microbatches": 4,
921
+ "elapsed_seconds": 78.14046130888164,
922
+ "batch_question_ids_sha256": "933b002848b57b1ae58768d4d38da9823ca19226389dff7c41f6ba3e58ff14b6"
923
+ },
924
+ {
925
+ "step": 100,
926
+ "phase": "full",
927
+ "loss": 0.15166849168599583,
928
+ "questions": 16,
929
+ "microbatches": 4,
930
+ "elapsed_seconds": 78.94848389271647,
931
+ "batch_question_ids_sha256": "2a2b48544fe4380e81979729c21f00410faff377b407c4afdee1bef7530a71af",
932
+ "dev": {
933
+ "questions": 372,
934
+ "objective": "observed_outcome",
935
+ "eligible_questions": 372,
936
+ "target_ce": 0.4421807345122086,
937
+ "target_kl": 0.4421807345122086,
938
+ "target_tv": 0.2684592128219396,
939
+ "by_target_kind": {
940
+ "observed_outcome": {
941
+ "n": 372,
942
+ "ce": 0.4421807345122086,
943
+ "kl": 0.4421807345122086,
944
+ "tv": 0.2684592128219396
945
+ }
946
+ },
947
+ "independent_targets": {
948
+ "teacher": {
949
+ "n": 0,
950
+ "ce": 0.0,
951
+ "kl": 0.0,
952
+ "tv": 0.0
953
+ },
954
+ "gold_distribution": {
955
+ "n": 372,
956
+ "ce": 0.46751975308439736,
957
+ "kl": 0.19751716047686577,
958
+ "tv": 0.17476332997830224
959
+ }
960
+ },
961
+ "note": "question-mean distribution recovery; policy targets and reference argmax labels are not observed outcomes"
962
+ }
963
+ }
964
+ ]
variants/events_ce_seed17/backbone_config/config.json ADDED
@@ -0,0 +1,63 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ {
2
+ "architectures": [
3
+ "Qwen3ForCausalLM"
4
+ ],
5
+ "attention_bias": false,
6
+ "attention_dropout": 0.0,
7
+ "bos_token_id": 151643,
8
+ "dtype": "float32",
9
+ "eos_token_id": 151645,
10
+ "head_dim": 128,
11
+ "hidden_act": "silu",
12
+ "hidden_size": 1024,
13
+ "initializer_range": 0.02,
14
+ "intermediate_size": 3072,
15
+ "layer_types": [
16
+ "full_attention",
17
+ "full_attention",
18
+ "full_attention",
19
+ "full_attention",
20
+ "full_attention",
21
+ "full_attention",
22
+ "full_attention",
23
+ "full_attention",
24
+ "full_attention",
25
+ "full_attention",
26
+ "full_attention",
27
+ "full_attention",
28
+ "full_attention",
29
+ "full_attention",
30
+ "full_attention",
31
+ "full_attention",
32
+ "full_attention",
33
+ "full_attention",
34
+ "full_attention",
35
+ "full_attention",
36
+ "full_attention",
37
+ "full_attention",
38
+ "full_attention",
39
+ "full_attention",
40
+ "full_attention",
41
+ "full_attention",
42
+ "full_attention",
43
+ "full_attention"
44
+ ],
45
+ "max_position_embeddings": 40960,
46
+ "max_window_layers": 28,
47
+ "model_type": "qwen3",
48
+ "num_attention_heads": 16,
49
+ "num_hidden_layers": 28,
50
+ "num_key_value_heads": 8,
51
+ "pad_token_id": null,
52
+ "rms_norm_eps": 1e-06,
53
+ "rope_parameters": {
54
+ "rope_theta": 1000000,
55
+ "rope_type": "default"
56
+ },
57
+ "sliding_window": null,
58
+ "tie_word_embeddings": true,
59
+ "transformers_version": "5.17.0",
60
+ "use_cache": false,
61
+ "use_sliding_window": false,
62
+ "vocab_size": 151936
63
+ }
variants/events_ce_seed17/best.safetensors ADDED
@@ -0,0 +1,3 @@
 
 
 
 
1
+ version https://git-lfs.github.com/spec/v1
2
+ oid sha256:582c01139e29807b5f2690335271cf5647583cbfc3dd963d5967792bb3963992
3
+ size 2385039280
variants/events_ce_seed17/config.json ADDED
@@ -0,0 +1,69 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ {
2
+ "input": "data/scaled_games_v4/events",
3
+ "output_dir": "runs/events_ce_seed17",
4
+ "model": "Qwen/Qwen3-0.6B",
5
+ "revision": "main",
6
+ "init_checkpoint": "/home/rwang/openjev_codex_20260917/runs/v3_teacher_coords_multi_seed17",
7
+ "objective": "observed_outcome",
8
+ "loss": "ce",
9
+ "reward_samples": 32,
10
+ "gradient_checkpointing": true,
11
+ "set_head": "attention",
12
+ "steps": 100,
13
+ "head_steps": 0,
14
+ "batch_questions": 16,
15
+ "microbatch_questions": 4,
16
+ "max_microbatch_tokens": 16384,
17
+ "eval_every": 50,
18
+ "max_length": 8192,
19
+ "seed": 17,
20
+ "backbone_lr": 2e-05,
21
+ "head_lr": 0.0002,
22
+ "head_warmup_lr": 0.001,
23
+ "precision": "bf16",
24
+ "disable_native_triton": true,
25
+ "validate_only": false,
26
+ "self_check": false,
27
+ "schema_version": "openjev-decision-pipeline-v1",
28
+ "resolved_model_revision": "c1899de289a04d12100db370d81485cdf75e47ca",
29
+ "initialization": "local DecisionModel warm start, fresh optimizer",
30
+ "data_sha256": {
31
+ "data/scaled_games_v4/events/train.jsonl": "4cebf698e4b71dc37d4d0e985b6577dd565ac370ac82b6f47f375aa6af69f8c9",
32
+ "data/scaled_games_v4/events/dev.jsonl": "0c352ee1064a5fb5d2aacad59cbea57afb897886d3d09e5ca108c866a5f893db",
33
+ "data/scaled_games_v4/events/calibration.jsonl": "4a8a8a2b81c3c46df6bb4c8690e97374dd0165cbef40566bb7f87c1f425d5562",
34
+ "data/scaled_games_v4/events/test.jsonl": "884bec71891668199925809d8708347f56f7bf06e8363e4a436347967b1423e3",
35
+ "data/scaled_games_v4/events/ood.jsonl": "3c00d845524634ca3b34559761cdaf0c52b7b7633d5517d1e7ee7f31697cea48"
36
+ },
37
+ "deps": {
38
+ "torch": "2.14.0",
39
+ "transformers": "5.17.0",
40
+ "safetensors": "0.8.0"
41
+ },
42
+ "gpu": "NVIDIA A100-SXM4-80GB",
43
+ "parameter_storage": "float32",
44
+ "forward_autocast": "bfloat16",
45
+ "parameter_count": 596250498,
46
+ "train_questions": 1124,
47
+ "all_train_questions": 1124,
48
+ "schema_counts": {
49
+ "records": 2356,
50
+ "questions_by_split": {
51
+ "train": 1124,
52
+ "dev": 372,
53
+ "calibration": 368,
54
+ "test": 364,
55
+ "ood": 128
56
+ },
57
+ "eligible_by_split_objective": {
58
+ "calibration/gold_distribution": 368,
59
+ "dev/gold_distribution": 372,
60
+ "ood/gold_distribution": 128,
61
+ "test/gold_distribution": 364,
62
+ "train/gold_distribution": 1124
63
+ }
64
+ },
65
+ "loss_description": "per-question proper loss; complete candidate normalization",
66
+ "rlcd_candidate": null,
67
+ "selection": "minimum dev target CE; held-out test first evaluated after training and checkpoint selection",
68
+ "parallelism": "complete candidate paths batched in one backbone call per microbatch; no prefix sharing"
69
+ }
variants/events_ce_seed17/summary.json ADDED
@@ -0,0 +1,136 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ {
2
+ "best_step": 100,
3
+ "best_dev_target_ce": 0.47277030621569244,
4
+ "selected_on": "dev target CE",
5
+ "objective": "observed_outcome",
6
+ "metrics_by_split": {
7
+ "dev": {
8
+ "questions": 372,
9
+ "objective": "observed_outcome",
10
+ "eligible_questions": 372,
11
+ "target_ce": 0.47277030621569244,
12
+ "target_kl": 0.47277030621569244,
13
+ "target_tv": 0.26819442928540077,
14
+ "by_target_kind": {
15
+ "observed_outcome": {
16
+ "n": 372,
17
+ "ce": 0.47277030621569244,
18
+ "kl": 0.47277030621569244,
19
+ "tv": 0.26819442928540077
20
+ }
21
+ },
22
+ "independent_targets": {
23
+ "teacher": {
24
+ "n": 0,
25
+ "ce": 0.0,
26
+ "kl": 0.0,
27
+ "tv": 0.0
28
+ },
29
+ "gold_distribution": {
30
+ "n": 372,
31
+ "ce": 0.4926936688108985,
32
+ "kl": 0.22269107620336712,
33
+ "tv": 0.1714566637835109
34
+ }
35
+ },
36
+ "note": "question-mean distribution recovery; policy targets and reference argmax labels are not observed outcomes"
37
+ },
38
+ "calibration": {
39
+ "questions": 368,
40
+ "objective": "observed_outcome",
41
+ "eligible_questions": 368,
42
+ "target_ce": 0.5122280796241342,
43
+ "target_kl": 0.5122280796241342,
44
+ "target_tv": 0.2856829761147523,
45
+ "by_target_kind": {
46
+ "observed_outcome": {
47
+ "n": 368,
48
+ "ce": 0.5122280796241342,
49
+ "kl": 0.5122280796241342,
50
+ "tv": 0.2856829761147523
51
+ }
52
+ },
53
+ "independent_targets": {
54
+ "teacher": {
55
+ "n": 0,
56
+ "ce": 0.0,
57
+ "kl": 0.0,
58
+ "tv": 0.0
59
+ },
60
+ "gold_distribution": {
61
+ "n": 368,
62
+ "ce": 0.4798722818150309,
63
+ "kl": 0.1935759627343198,
64
+ "tv": 0.16869317512676552
65
+ }
66
+ },
67
+ "note": "question-mean distribution recovery; policy targets and reference argmax labels are not observed outcomes"
68
+ },
69
+ "test": {
70
+ "questions": 364,
71
+ "objective": "observed_outcome",
72
+ "eligible_questions": 364,
73
+ "target_ce": 0.46045028819154266,
74
+ "target_kl": 0.46045028819154266,
75
+ "target_tv": 0.2658522378700697,
76
+ "by_target_kind": {
77
+ "observed_outcome": {
78
+ "n": 364,
79
+ "ce": 0.46045028819154266,
80
+ "kl": 0.46045028819154266,
81
+ "tv": 0.2658522378700697
82
+ }
83
+ },
84
+ "independent_targets": {
85
+ "teacher": {
86
+ "n": 0,
87
+ "ce": 0.0,
88
+ "kl": 0.0,
89
+ "tv": 0.0
90
+ },
91
+ "gold_distribution": {
92
+ "n": 364,
93
+ "ce": 0.46643101807363885,
94
+ "kl": 0.20566021774752138,
95
+ "tv": 0.16303083589195544
96
+ }
97
+ },
98
+ "note": "question-mean distribution recovery; policy targets and reference argmax labels are not observed outcomes"
99
+ },
100
+ "ood": {
101
+ "questions": 128,
102
+ "objective": "observed_outcome",
103
+ "eligible_questions": 128,
104
+ "target_ce": 0.33238140555610585,
105
+ "target_kl": 0.33238140555610585,
106
+ "target_tv": 0.23786875248045813,
107
+ "by_target_kind": {
108
+ "observed_outcome": {
109
+ "n": 128,
110
+ "ce": 0.33238140555610585,
111
+ "kl": 0.33238140555610585,
112
+ "tv": 0.23786875248045813
113
+ }
114
+ },
115
+ "independent_targets": {
116
+ "teacher": {
117
+ "n": 0,
118
+ "ce": 0.0,
119
+ "kl": 0.0,
120
+ "tv": 0.0
121
+ },
122
+ "gold_distribution": {
123
+ "n": 128,
124
+ "ce": 0.32782799216499303,
125
+ "kl": 0.09434077052836559,
126
+ "tv": 0.12276145397536038
127
+ }
128
+ },
129
+ "note": "question-mean distribution recovery; policy targets and reference argmax labels are not observed outcomes"
130
+ }
131
+ },
132
+ "temperature": 1.0,
133
+ "temperature_fitted": false,
134
+ "training_seconds": 96.74807532411069,
135
+ "max_gpu_allocated_gb": 11.964137984
136
+ }
variants/events_ce_seed17/target_audit.json ADDED
The diff for this file is too large to render. See raw diff
 
variants/events_ce_seed17/tokenizer/chat_template.jinja ADDED
@@ -0,0 +1,89 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ {%- if tools %}
2
+ {{- '<|im_start|>system\n' }}
3
+ {%- if messages[0].role == 'system' %}
4
+ {{- messages[0].content + '\n\n' }}
5
+ {%- endif %}
6
+ {{- "# Tools\n\nYou may call one or more functions to assist with the user query.\n\nYou are provided with function signatures within <tools></tools> XML tags:\n<tools>" }}
7
+ {%- for tool in tools %}
8
+ {{- "\n" }}
9
+ {{- tool | tojson }}
10
+ {%- endfor %}
11
+ {{- "\n</tools>\n\nFor each function call, return a json object with function name and arguments within <tool_call></tool_call> XML tags:\n<tool_call>\n{\"name\": <function-name>, \"arguments\": <args-json-object>}\n</tool_call><|im_end|>\n" }}
12
+ {%- else %}
13
+ {%- if messages[0].role == 'system' %}
14
+ {{- '<|im_start|>system\n' + messages[0].content + '<|im_end|>\n' }}
15
+ {%- endif %}
16
+ {%- endif %}
17
+ {%- set ns = namespace(multi_step_tool=true, last_query_index=messages|length - 1) %}
18
+ {%- for message in messages[::-1] %}
19
+ {%- set index = (messages|length - 1) - loop.index0 %}
20
+ {%- if ns.multi_step_tool and message.role == "user" and message.content is string and not(message.content.startswith('<tool_response>') and message.content.endswith('</tool_response>')) %}
21
+ {%- set ns.multi_step_tool = false %}
22
+ {%- set ns.last_query_index = index %}
23
+ {%- endif %}
24
+ {%- endfor %}
25
+ {%- for message in messages %}
26
+ {%- if message.content is string %}
27
+ {%- set content = message.content %}
28
+ {%- else %}
29
+ {%- set content = '' %}
30
+ {%- endif %}
31
+ {%- if (message.role == "user") or (message.role == "system" and not loop.first) %}
32
+ {{- '<|im_start|>' + message.role + '\n' + content + '<|im_end|>' + '\n' }}
33
+ {%- elif message.role == "assistant" %}
34
+ {%- set reasoning_content = '' %}
35
+ {%- if message.reasoning_content is string %}
36
+ {%- set reasoning_content = message.reasoning_content %}
37
+ {%- else %}
38
+ {%- if '</think>' in content %}
39
+ {%- set reasoning_content = content.split('</think>')[0].rstrip('\n').split('<think>')[-1].lstrip('\n') %}
40
+ {%- set content = content.split('</think>')[-1].lstrip('\n') %}
41
+ {%- endif %}
42
+ {%- endif %}
43
+ {%- if loop.index0 > ns.last_query_index %}
44
+ {%- if loop.last or (not loop.last and reasoning_content) %}
45
+ {{- '<|im_start|>' + message.role + '\n<think>\n' + reasoning_content.strip('\n') + '\n</think>\n\n' + content.lstrip('\n') }}
46
+ {%- else %}
47
+ {{- '<|im_start|>' + message.role + '\n' + content }}
48
+ {%- endif %}
49
+ {%- else %}
50
+ {{- '<|im_start|>' + message.role + '\n' + content }}
51
+ {%- endif %}
52
+ {%- if message.tool_calls %}
53
+ {%- for tool_call in message.tool_calls %}
54
+ {%- if (loop.first and content) or (not loop.first) %}
55
+ {{- '\n' }}
56
+ {%- endif %}
57
+ {%- if tool_call.function %}
58
+ {%- set tool_call = tool_call.function %}
59
+ {%- endif %}
60
+ {{- '<tool_call>\n{"name": "' }}
61
+ {{- tool_call.name }}
62
+ {{- '", "arguments": ' }}
63
+ {%- if tool_call.arguments is string %}
64
+ {{- tool_call.arguments }}
65
+ {%- else %}
66
+ {{- tool_call.arguments | tojson }}
67
+ {%- endif %}
68
+ {{- '}\n</tool_call>' }}
69
+ {%- endfor %}
70
+ {%- endif %}
71
+ {{- '<|im_end|>\n' }}
72
+ {%- elif message.role == "tool" %}
73
+ {%- if loop.first or (messages[loop.index0 - 1].role != "tool") %}
74
+ {{- '<|im_start|>user' }}
75
+ {%- endif %}
76
+ {{- '\n<tool_response>\n' }}
77
+ {{- content }}
78
+ {{- '\n</tool_response>' }}
79
+ {%- if loop.last or (messages[loop.index0 + 1].role != "tool") %}
80
+ {{- '<|im_end|>\n' }}
81
+ {%- endif %}
82
+ {%- endif %}
83
+ {%- endfor %}
84
+ {%- if add_generation_prompt %}
85
+ {{- '<|im_start|>assistant\n' }}
86
+ {%- if enable_thinking is defined and enable_thinking is false %}
87
+ {{- '<think>\n\n</think>\n\n' }}
88
+ {%- endif %}
89
+ {%- endif %}
variants/events_ce_seed17/tokenizer/tokenizer.json ADDED
@@ -0,0 +1,3 @@
 
 
 
 
1
+ version https://git-lfs.github.com/spec/v1
2
+ oid sha256:be75606093db2094d7cd20f3c2f385c212750648bd6ea4fb2bf507a6a4c55506
3
+ size 11422650
variants/events_ce_seed17/tokenizer/tokenizer_config.json ADDED
@@ -0,0 +1,30 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ {
2
+ "add_prefix_space": false,
3
+ "backend": "tokenizers",
4
+ "bos_token": null,
5
+ "clean_up_tokenization_spaces": false,
6
+ "eos_token": "<|im_end|>",
7
+ "errors": "replace",
8
+ "extra_special_tokens": [
9
+ "<|im_start|>",
10
+ "<|im_end|>",
11
+ "<|object_ref_start|>",
12
+ "<|object_ref_end|>",
13
+ "<|box_start|>",
14
+ "<|box_end|>",
15
+ "<|quad_start|>",
16
+ "<|quad_end|>",
17
+ "<|vision_start|>",
18
+ "<|vision_end|>",
19
+ "<|vision_pad|>",
20
+ "<|image_pad|>",
21
+ "<|video_pad|>"
22
+ ],
23
+ "is_local": true,
24
+ "local_files_only": true,
25
+ "model_max_length": 131072,
26
+ "pad_token": "<|endoftext|>",
27
+ "split_special_tokens": false,
28
+ "tokenizer_class": "Qwen2Tokenizer",
29
+ "unk_token": null
30
+ }
variants/events_ce_seed17/train_log.json ADDED
@@ -0,0 +1,964 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ [
2
+ {
3
+ "step": 1,
4
+ "phase": "full",
5
+ "loss": 0.6698931157588959,
6
+ "questions": 16,
7
+ "microbatches": 4,
8
+ "elapsed_seconds": 1.206429815851152,
9
+ "batch_question_ids_sha256": "9815e3084b0060ca1bf3f53e09caa2e8ba2ecfbfa18cd5a90c1b6ca1dc29fbcb"
10
+ },
11
+ {
12
+ "step": 2,
13
+ "phase": "full",
14
+ "loss": 0.6899416968226433,
15
+ "questions": 16,
16
+ "microbatches": 4,
17
+ "elapsed_seconds": 1.9398853788152337,
18
+ "batch_question_ids_sha256": "0766e00bccd0b9d63535226a2a7454d88d591d6a1adb32befcd64d40174bd87c"
19
+ },
20
+ {
21
+ "step": 3,
22
+ "phase": "full",
23
+ "loss": 0.4910094514489174,
24
+ "questions": 16,
25
+ "microbatches": 4,
26
+ "elapsed_seconds": 2.678210761398077,
27
+ "batch_question_ids_sha256": "3eb8fb253ed8d59b98fd13fb12ae3aa08473ff98bf07f26c7537d8d4c6999318"
28
+ },
29
+ {
30
+ "step": 4,
31
+ "phase": "full",
32
+ "loss": 0.47546009719371796,
33
+ "questions": 16,
34
+ "microbatches": 4,
35
+ "elapsed_seconds": 3.3923207791522145,
36
+ "batch_question_ids_sha256": "b72b9555459b9660b8cfea9c061b0d69e21b213181cf19973fa9312c1d740e4c"
37
+ },
38
+ {
39
+ "step": 5,
40
+ "phase": "full",
41
+ "loss": 0.472549706697464,
42
+ "questions": 16,
43
+ "microbatches": 4,
44
+ "elapsed_seconds": 4.11856244970113,
45
+ "batch_question_ids_sha256": "851370879eecc2bb2b872bf9f26a90b0c422dbb910faa7ab99625462dd70126c"
46
+ },
47
+ {
48
+ "step": 6,
49
+ "phase": "full",
50
+ "loss": 0.47603993117809296,
51
+ "questions": 16,
52
+ "microbatches": 4,
53
+ "elapsed_seconds": 4.842001353390515,
54
+ "batch_question_ids_sha256": "a44ac9fef40754316a48915525e55ee3c8c5f613d2a3a206a70c53f6a360d6a1"
55
+ },
56
+ {
57
+ "step": 7,
58
+ "phase": "full",
59
+ "loss": 0.4351997897028923,
60
+ "questions": 16,
61
+ "microbatches": 4,
62
+ "elapsed_seconds": 5.562285288237035,
63
+ "batch_question_ids_sha256": "dd15d5df027d50d8bc83cc139ce58f36adf2613954b88024320afffaf4f275ce"
64
+ },
65
+ {
66
+ "step": 8,
67
+ "phase": "full",
68
+ "loss": 0.5573603184893727,
69
+ "questions": 16,
70
+ "microbatches": 4,
71
+ "elapsed_seconds": 6.283805459737778,
72
+ "batch_question_ids_sha256": "b16e21ce00030d172555d2ed47f3935798c5293130655f2d4bb114fab8d4f82e"
73
+ },
74
+ {
75
+ "step": 9,
76
+ "phase": "full",
77
+ "loss": 0.25312627106904984,
78
+ "questions": 16,
79
+ "microbatches": 4,
80
+ "elapsed_seconds": 7.027102201245725,
81
+ "batch_question_ids_sha256": "db13a8f636762d4e5c56f60dcecdc4a04f7acccee36ed4846fe242c110813e33"
82
+ },
83
+ {
84
+ "step": 10,
85
+ "phase": "full",
86
+ "loss": 0.5761465504765511,
87
+ "questions": 16,
88
+ "microbatches": 4,
89
+ "elapsed_seconds": 7.779015112668276,
90
+ "batch_question_ids_sha256": "d29a2ab7bae9372a6460e6d41d2320b0a13b55633fcc21e2a06f06a64a2b8f22"
91
+ },
92
+ {
93
+ "step": 11,
94
+ "phase": "full",
95
+ "loss": 0.4777762070298195,
96
+ "questions": 16,
97
+ "microbatches": 4,
98
+ "elapsed_seconds": 8.493401571176946,
99
+ "batch_question_ids_sha256": "e435c49e7cce06b5c387742c1cf2eae1fb3f57349a03ff3229713bf8e9179b7f"
100
+ },
101
+ {
102
+ "step": 12,
103
+ "phase": "full",
104
+ "loss": 0.7341089807450771,
105
+ "questions": 16,
106
+ "microbatches": 4,
107
+ "elapsed_seconds": 9.231942504644394,
108
+ "batch_question_ids_sha256": "52067a44724ec88a3f85cace2ddba01f511bb396cbf27877697e85efd4f0c68e"
109
+ },
110
+ {
111
+ "step": 13,
112
+ "phase": "full",
113
+ "loss": 0.3001631014049053,
114
+ "questions": 16,
115
+ "microbatches": 4,
116
+ "elapsed_seconds": 9.975338879972696,
117
+ "batch_question_ids_sha256": "d01e765c3bf9b2b1edf10d82d3b784f51514a21682d040badd6d9c1c29c3069f"
118
+ },
119
+ {
120
+ "step": 14,
121
+ "phase": "full",
122
+ "loss": 0.43069224059581757,
123
+ "questions": 16,
124
+ "microbatches": 4,
125
+ "elapsed_seconds": 10.703695585951209,
126
+ "batch_question_ids_sha256": "f7d76860a7e7abbd079278179b2a6054aa8b1bf90c6e38db91f414cf906eff6a"
127
+ },
128
+ {
129
+ "step": 15,
130
+ "phase": "full",
131
+ "loss": 0.5127209424972534,
132
+ "questions": 16,
133
+ "microbatches": 4,
134
+ "elapsed_seconds": 11.451611735858023,
135
+ "batch_question_ids_sha256": "1d089d97e3e6836ec1771dba1776d38f6a8ca414f96cc3c25fd0053ed3e27a44"
136
+ },
137
+ {
138
+ "step": 16,
139
+ "phase": "full",
140
+ "loss": 0.3563181236386299,
141
+ "questions": 16,
142
+ "microbatches": 4,
143
+ "elapsed_seconds": 12.173952360637486,
144
+ "batch_question_ids_sha256": "abd204307499e62673daa334cd3a4533448a0ef971b517e7acbf23783905ae91"
145
+ },
146
+ {
147
+ "step": 17,
148
+ "phase": "full",
149
+ "loss": 0.36245644837617874,
150
+ "questions": 16,
151
+ "microbatches": 4,
152
+ "elapsed_seconds": 12.8896827660501,
153
+ "batch_question_ids_sha256": "6d4029fa8044653ff0ccbe80e4371cccd55b6fc7c6462fb81d77f565bde1e2f5"
154
+ },
155
+ {
156
+ "step": 18,
157
+ "phase": "full",
158
+ "loss": 0.5653930976986885,
159
+ "questions": 16,
160
+ "microbatches": 4,
161
+ "elapsed_seconds": 13.62013566866517,
162
+ "batch_question_ids_sha256": "88ddecb0e43b996acf8e321896454ceb6e79095126d68380cfdc7c74ba66d04a"
163
+ },
164
+ {
165
+ "step": 19,
166
+ "phase": "full",
167
+ "loss": 0.5002457201480865,
168
+ "questions": 16,
169
+ "microbatches": 4,
170
+ "elapsed_seconds": 14.346520440652966,
171
+ "batch_question_ids_sha256": "3ba3b72125f7617c5cb4dfc031843bc10773d8aa7747ef6eef76ec813abc7a11"
172
+ },
173
+ {
174
+ "step": 20,
175
+ "phase": "full",
176
+ "loss": 0.6882804483175278,
177
+ "questions": 16,
178
+ "microbatches": 4,
179
+ "elapsed_seconds": 15.069504925049841,
180
+ "batch_question_ids_sha256": "a8fbfc400a3dd61e2468ddde908d8e6914e8742e66df75fa209575bbd49f674f"
181
+ },
182
+ {
183
+ "step": 21,
184
+ "phase": "full",
185
+ "loss": 0.3362383544445038,
186
+ "questions": 16,
187
+ "microbatches": 4,
188
+ "elapsed_seconds": 15.863264437764883,
189
+ "batch_question_ids_sha256": "49a7b1f6bb6891cd06f7c495942bbdbd98cfbd7435f49d77922ab364b9c1c5f5"
190
+ },
191
+ {
192
+ "step": 22,
193
+ "phase": "full",
194
+ "loss": 0.4032861888408661,
195
+ "questions": 16,
196
+ "microbatches": 4,
197
+ "elapsed_seconds": 16.61500151734799,
198
+ "batch_question_ids_sha256": "d93e27a20a5666b8fe2ea5489dac0a58f99a770bfb0d4ee6870d36621ed78dee"
199
+ },
200
+ {
201
+ "step": 23,
202
+ "phase": "full",
203
+ "loss": 0.5400929227471352,
204
+ "questions": 16,
205
+ "microbatches": 4,
206
+ "elapsed_seconds": 17.49686106480658,
207
+ "batch_question_ids_sha256": "88845418c486e4c7d856761b8aca680a7ca52ec59c84219e708745af76a445c2"
208
+ },
209
+ {
210
+ "step": 24,
211
+ "phase": "full",
212
+ "loss": 0.4465572461485863,
213
+ "questions": 16,
214
+ "microbatches": 4,
215
+ "elapsed_seconds": 18.21982985921204,
216
+ "batch_question_ids_sha256": "b7fcb8d686d0a87ea4fd6ded60d595c7ccc4b132fc739e047564d72f5944577f"
217
+ },
218
+ {
219
+ "step": 25,
220
+ "phase": "full",
221
+ "loss": 0.38778524473309517,
222
+ "questions": 16,
223
+ "microbatches": 4,
224
+ "elapsed_seconds": 18.957347061485052,
225
+ "batch_question_ids_sha256": "b2e6af990ea722e6cb276f57a736c23766306ae29b211bf00447d868b434c962"
226
+ },
227
+ {
228
+ "step": 26,
229
+ "phase": "full",
230
+ "loss": 0.46882137656211853,
231
+ "questions": 16,
232
+ "microbatches": 4,
233
+ "elapsed_seconds": 19.701538900844753,
234
+ "batch_question_ids_sha256": "67253956a0835f33db1fa53443fed54476a027338e0ffdd7726dcb48fbef1c9a"
235
+ },
236
+ {
237
+ "step": 27,
238
+ "phase": "full",
239
+ "loss": 0.4425557181239128,
240
+ "questions": 16,
241
+ "microbatches": 4,
242
+ "elapsed_seconds": 20.427865900099277,
243
+ "batch_question_ids_sha256": "b37cd333ba07e45e13875704fb0cc42b2f44fc5317737590bb27d88a7ce61452"
244
+ },
245
+ {
246
+ "step": 28,
247
+ "phase": "full",
248
+ "loss": 0.537452906370163,
249
+ "questions": 16,
250
+ "microbatches": 4,
251
+ "elapsed_seconds": 21.164012663066387,
252
+ "batch_question_ids_sha256": "d2a5a95e218ea89d7469dfdd8ff320de68f88ff52171675c0ccf7c9c3d95a891"
253
+ },
254
+ {
255
+ "step": 29,
256
+ "phase": "full",
257
+ "loss": 0.33650635927915573,
258
+ "questions": 16,
259
+ "microbatches": 4,
260
+ "elapsed_seconds": 21.895237404853106,
261
+ "batch_question_ids_sha256": "1c6aa40b463139c53b7a93df96c3a1db9c53fac3123be7e8dda4fc8f53bc7996"
262
+ },
263
+ {
264
+ "step": 30,
265
+ "phase": "full",
266
+ "loss": 0.18629102176055312,
267
+ "questions": 16,
268
+ "microbatches": 4,
269
+ "elapsed_seconds": 22.596429711207747,
270
+ "batch_question_ids_sha256": "be09ca5b88c61e0617249157b2d75545b12e403abaf2ff570be1e6d30856d742"
271
+ },
272
+ {
273
+ "step": 31,
274
+ "phase": "full",
275
+ "loss": 0.575798325240612,
276
+ "questions": 16,
277
+ "microbatches": 4,
278
+ "elapsed_seconds": 23.33407491259277,
279
+ "batch_question_ids_sha256": "08f4544d1dfa624e78df3609c0059204a16146f8421bf85c14d0b10b5926b448"
280
+ },
281
+ {
282
+ "step": 32,
283
+ "phase": "full",
284
+ "loss": 0.5029938481748104,
285
+ "questions": 16,
286
+ "microbatches": 4,
287
+ "elapsed_seconds": 24.04410624690354,
288
+ "batch_question_ids_sha256": "f8a0c95d4ce5a21d3b7bdd381eb7c9d15a251da10b33f0a8521a3e81eeb1450e"
289
+ },
290
+ {
291
+ "step": 33,
292
+ "phase": "full",
293
+ "loss": 0.3700496293604374,
294
+ "questions": 16,
295
+ "microbatches": 4,
296
+ "elapsed_seconds": 24.790576100349426,
297
+ "batch_question_ids_sha256": "4c396fcb15e0a31d52d00473dc2f3bee3767da23146fb5038b6e63d9ad5016c8"
298
+ },
299
+ {
300
+ "step": 34,
301
+ "phase": "full",
302
+ "loss": 0.3722036397084594,
303
+ "questions": 16,
304
+ "microbatches": 4,
305
+ "elapsed_seconds": 25.522674680687487,
306
+ "batch_question_ids_sha256": "c59ffed718f7ded62b1028779c65e8759de67070b26019e8642f75dd2e43a512"
307
+ },
308
+ {
309
+ "step": 35,
310
+ "phase": "full",
311
+ "loss": 0.6066184220835567,
312
+ "questions": 16,
313
+ "microbatches": 4,
314
+ "elapsed_seconds": 26.259713889099658,
315
+ "batch_question_ids_sha256": "0e54ce4db6221fac98c15ce77d50701153708881dda86adbcb2776637215c9bd"
316
+ },
317
+ {
318
+ "step": 36,
319
+ "phase": "full",
320
+ "loss": 0.7472243523225188,
321
+ "questions": 16,
322
+ "microbatches": 4,
323
+ "elapsed_seconds": 26.99773812200874,
324
+ "batch_question_ids_sha256": "24c257b893015e139aaec823c9fa8ed630a7d2df239552f229ebb832a7746723"
325
+ },
326
+ {
327
+ "step": 37,
328
+ "phase": "full",
329
+ "loss": 0.5684906765818596,
330
+ "questions": 16,
331
+ "microbatches": 4,
332
+ "elapsed_seconds": 27.740370864979923,
333
+ "batch_question_ids_sha256": "bc22d73d0ac85948fd9bbd0813992b093f9beb230652c6234447dbbbfb4add35"
334
+ },
335
+ {
336
+ "step": 38,
337
+ "phase": "full",
338
+ "loss": 0.36942679062485695,
339
+ "questions": 16,
340
+ "microbatches": 4,
341
+ "elapsed_seconds": 28.50852065719664,
342
+ "batch_question_ids_sha256": "6e47f7951fb42ce413ed2f4ce54d8832b534a3e9d0f2e35829a1608286227ef4"
343
+ },
344
+ {
345
+ "step": 39,
346
+ "phase": "full",
347
+ "loss": 0.3635300546884537,
348
+ "questions": 16,
349
+ "microbatches": 4,
350
+ "elapsed_seconds": 29.22664304729551,
351
+ "batch_question_ids_sha256": "b5bfa6d3b4f6544cb42bc8b82b7c88ad3430367ce2f47dd0c3182d7f55e26796"
352
+ },
353
+ {
354
+ "step": 40,
355
+ "phase": "full",
356
+ "loss": 0.40619706362485886,
357
+ "questions": 16,
358
+ "microbatches": 4,
359
+ "elapsed_seconds": 29.978176869452,
360
+ "batch_question_ids_sha256": "e356d8692763396b1f305f967bc42bbea222b3adcfd7355c8ffd489d7c4a8702"
361
+ },
362
+ {
363
+ "step": 41,
364
+ "phase": "full",
365
+ "loss": 0.5313360318541527,
366
+ "questions": 16,
367
+ "microbatches": 4,
368
+ "elapsed_seconds": 30.714267833158374,
369
+ "batch_question_ids_sha256": "eddf9bafc965ad35c8a9189e6a7b19eb336e5f2736ab490e5d0f7547f885b883"
370
+ },
371
+ {
372
+ "step": 42,
373
+ "phase": "full",
374
+ "loss": 0.2942931465804577,
375
+ "questions": 16,
376
+ "microbatches": 4,
377
+ "elapsed_seconds": 31.4632102586329,
378
+ "batch_question_ids_sha256": "93f507828c7fa8715902d87edef912b05fe2920c25f400745b5d4b1835ff1fcd"
379
+ },
380
+ {
381
+ "step": 43,
382
+ "phase": "full",
383
+ "loss": 0.43418202362954617,
384
+ "questions": 16,
385
+ "microbatches": 4,
386
+ "elapsed_seconds": 32.20179224293679,
387
+ "batch_question_ids_sha256": "50f72e83de9b1b9daf154f013aa536b1454d8d8c13b503ee6efa9007c721faa6"
388
+ },
389
+ {
390
+ "step": 44,
391
+ "phase": "full",
392
+ "loss": 0.5574786812067032,
393
+ "questions": 16,
394
+ "microbatches": 4,
395
+ "elapsed_seconds": 32.9865256017074,
396
+ "batch_question_ids_sha256": "71067cb1ca0d666b5a16e53227ae1840e468c0c39e879ef8004ac73520b396f9"
397
+ },
398
+ {
399
+ "step": 45,
400
+ "phase": "full",
401
+ "loss": 0.31117148976773024,
402
+ "questions": 16,
403
+ "microbatches": 4,
404
+ "elapsed_seconds": 33.79001239966601,
405
+ "batch_question_ids_sha256": "b9270e713e4fca02ecc7205c7ac1e320d3c87286eb7237ed0b02a4686b626bc4"
406
+ },
407
+ {
408
+ "step": 46,
409
+ "phase": "full",
410
+ "loss": 0.4900943711400032,
411
+ "questions": 16,
412
+ "microbatches": 4,
413
+ "elapsed_seconds": 34.549148452468216,
414
+ "batch_question_ids_sha256": "fbecc4d2e806b5b42f4f82a7e99465e1fceb5b7b61dfc3cd167654deeef897d0"
415
+ },
416
+ {
417
+ "step": 47,
418
+ "phase": "full",
419
+ "loss": 0.35709232836961746,
420
+ "questions": 16,
421
+ "microbatches": 4,
422
+ "elapsed_seconds": 35.28404710628092,
423
+ "batch_question_ids_sha256": "9fb1e3f8e3eb222e0e377eeedad5b0df90d11389465803b67a9b5090d2e79ca0"
424
+ },
425
+ {
426
+ "step": 48,
427
+ "phase": "full",
428
+ "loss": 0.3858681209385395,
429
+ "questions": 16,
430
+ "microbatches": 4,
431
+ "elapsed_seconds": 36.02657262887806,
432
+ "batch_question_ids_sha256": "8ec776449fb8174bd55b829feff0711ca000a9e09d903e741da6f91e2d21a136"
433
+ },
434
+ {
435
+ "step": 49,
436
+ "phase": "full",
437
+ "loss": 0.16362383612431586,
438
+ "questions": 16,
439
+ "microbatches": 4,
440
+ "elapsed_seconds": 36.99050153326243,
441
+ "batch_question_ids_sha256": "a20669c6cee5e157b8b54729bb88921f4ff6df89b4dda467ec25abb10f0e78da"
442
+ },
443
+ {
444
+ "step": 50,
445
+ "phase": "full",
446
+ "loss": 0.2581261422019452,
447
+ "questions": 16,
448
+ "microbatches": 4,
449
+ "elapsed_seconds": 37.71963926590979,
450
+ "batch_question_ids_sha256": "cab2a2f34731e8a959300614046a8b8fc4c926971dd5909a1a6e3de0ad3acb26",
451
+ "dev": {
452
+ "questions": 372,
453
+ "objective": "observed_outcome",
454
+ "eligible_questions": 372,
455
+ "target_ce": 0.5806565442146586,
456
+ "target_kl": 0.5806565442146586,
457
+ "target_tv": 0.27957613891343785,
458
+ "by_target_kind": {
459
+ "observed_outcome": {
460
+ "n": 372,
461
+ "ce": 0.5806565442146586,
462
+ "kl": 0.5806565442146586,
463
+ "tv": 0.27957613891343785
464
+ }
465
+ },
466
+ "independent_targets": {
467
+ "teacher": {
468
+ "n": 0,
469
+ "ce": 0.0,
470
+ "kl": 0.0,
471
+ "tv": 0.0
472
+ },
473
+ "gold_distribution": {
474
+ "n": 372,
475
+ "ce": 0.5856488371589073,
476
+ "kl": 0.31564624455137563,
477
+ "tv": 0.18229928321525732
478
+ }
479
+ },
480
+ "note": "question-mean distribution recovery; policy targets and reference argmax labels are not observed outcomes"
481
+ }
482
+ },
483
+ {
484
+ "step": 51,
485
+ "phase": "full",
486
+ "loss": 0.5709785744547844,
487
+ "questions": 16,
488
+ "microbatches": 4,
489
+ "elapsed_seconds": 44.23742653615773,
490
+ "batch_question_ids_sha256": "46758e1c6996a3f1a932a11e8372539949f8ac8f4c0415b36f65c713e195096f"
491
+ },
492
+ {
493
+ "step": 52,
494
+ "phase": "full",
495
+ "loss": 0.22201280656736344,
496
+ "questions": 16,
497
+ "microbatches": 4,
498
+ "elapsed_seconds": 44.957907737232745,
499
+ "batch_question_ids_sha256": "d87fcd81d136c657c13e49aa033db193bfbb121e4a721fa86c16aad4c0a7c39d"
500
+ },
501
+ {
502
+ "step": 53,
503
+ "phase": "full",
504
+ "loss": 0.3642621487379074,
505
+ "questions": 16,
506
+ "microbatches": 4,
507
+ "elapsed_seconds": 45.691871273331344,
508
+ "batch_question_ids_sha256": "4fae217aef8823502570d85ef34334cdfe2994ba567eef62990a60ac5e3ede4a"
509
+ },
510
+ {
511
+ "step": 54,
512
+ "phase": "full",
513
+ "loss": 0.9462117776274681,
514
+ "questions": 16,
515
+ "microbatches": 4,
516
+ "elapsed_seconds": 46.41959032230079,
517
+ "batch_question_ids_sha256": "edbcfec53ad977392c69172be6901bdb4eee1725bf8fd83582f64ac6320e77cd"
518
+ },
519
+ {
520
+ "step": 55,
521
+ "phase": "full",
522
+ "loss": 0.618944376707077,
523
+ "questions": 16,
524
+ "microbatches": 4,
525
+ "elapsed_seconds": 47.144659521989524,
526
+ "batch_question_ids_sha256": "7af8e2cece43ae333d4687839255ee144e5968fc9aa063efa2ff6dd96c42d1b2"
527
+ },
528
+ {
529
+ "step": 56,
530
+ "phase": "full",
531
+ "loss": 0.520059403963387,
532
+ "questions": 16,
533
+ "microbatches": 4,
534
+ "elapsed_seconds": 47.87055600807071,
535
+ "batch_question_ids_sha256": "6d579036d49f401569b6e29499a444725fe7a64de0907bc07bd389fd352e615b"
536
+ },
537
+ {
538
+ "step": 57,
539
+ "phase": "full",
540
+ "loss": 0.292273685336113,
541
+ "questions": 16,
542
+ "microbatches": 4,
543
+ "elapsed_seconds": 48.57758554536849,
544
+ "batch_question_ids_sha256": "848946e51c61b43a984fc49047cb8c9f0fae9f28c6447d8b4ee2c74dafe198bc"
545
+ },
546
+ {
547
+ "step": 58,
548
+ "phase": "full",
549
+ "loss": 0.47000307217240334,
550
+ "questions": 16,
551
+ "microbatches": 4,
552
+ "elapsed_seconds": 49.29666138906032,
553
+ "batch_question_ids_sha256": "7a207cb4e3ff6157802e3b560ec0b63a488a8ab0eabaa5bcdea45eeeb107d3da"
554
+ },
555
+ {
556
+ "step": 59,
557
+ "phase": "full",
558
+ "loss": 0.4892919175326824,
559
+ "questions": 16,
560
+ "microbatches": 4,
561
+ "elapsed_seconds": 50.00856906082481,
562
+ "batch_question_ids_sha256": "6ed68210f517203266a035e4dc739a5e91ad1773fb37b2d954aabbc97ce44cc5"
563
+ },
564
+ {
565
+ "step": 60,
566
+ "phase": "full",
567
+ "loss": 0.3904850035905838,
568
+ "questions": 16,
569
+ "microbatches": 4,
570
+ "elapsed_seconds": 50.76625431701541,
571
+ "batch_question_ids_sha256": "cbce464330f46f2f323539fb9d7962cede14e7074be588012e1dfa684881e6d4"
572
+ },
573
+ {
574
+ "step": 61,
575
+ "phase": "full",
576
+ "loss": 0.3799959197640419,
577
+ "questions": 16,
578
+ "microbatches": 4,
579
+ "elapsed_seconds": 51.495480451732874,
580
+ "batch_question_ids_sha256": "7212e07d655e85ec9b791cd6cde65ec3e9ebfa5224d1854c76d9a8e98cc07be3"
581
+ },
582
+ {
583
+ "step": 62,
584
+ "phase": "full",
585
+ "loss": 0.37555741146206856,
586
+ "questions": 16,
587
+ "microbatches": 4,
588
+ "elapsed_seconds": 52.22409507166594,
589
+ "batch_question_ids_sha256": "768e8960dc65d1f0cc4b13cc2c90dbfe8db2c685cceb3a8ed4f834e59a13a2f5"
590
+ },
591
+ {
592
+ "step": 63,
593
+ "phase": "full",
594
+ "loss": 0.5196910500526428,
595
+ "questions": 16,
596
+ "microbatches": 4,
597
+ "elapsed_seconds": 52.95018903538585,
598
+ "batch_question_ids_sha256": "ce0e1bf7240822fc1f373178c20604e9dba0409aa63f2b51415a5f4db012ea78"
599
+ },
600
+ {
601
+ "step": 64,
602
+ "phase": "full",
603
+ "loss": 0.5655737966299057,
604
+ "questions": 16,
605
+ "microbatches": 4,
606
+ "elapsed_seconds": 53.69126392249018,
607
+ "batch_question_ids_sha256": "6b8c351645894c06829c8dffb3045f3a3081c0fcb0b24c924b9576410c78d47c"
608
+ },
609
+ {
610
+ "step": 65,
611
+ "phase": "full",
612
+ "loss": 0.2570880688726902,
613
+ "questions": 16,
614
+ "microbatches": 4,
615
+ "elapsed_seconds": 54.4230572860688,
616
+ "batch_question_ids_sha256": "06de39fb8b5de5ce866344dcb10447ed159378857789cc67daf23061f43e3e7f"
617
+ },
618
+ {
619
+ "step": 66,
620
+ "phase": "full",
621
+ "loss": 0.4608404226601124,
622
+ "questions": 16,
623
+ "microbatches": 4,
624
+ "elapsed_seconds": 55.15514244977385,
625
+ "batch_question_ids_sha256": "1728e00f1c3d6bfc533f2891740148ba51b40dcb766709fed2a21437282f6ee1"
626
+ },
627
+ {
628
+ "step": 67,
629
+ "phase": "full",
630
+ "loss": 0.2939162664115429,
631
+ "questions": 16,
632
+ "microbatches": 4,
633
+ "elapsed_seconds": 55.88841801136732,
634
+ "batch_question_ids_sha256": "26b40bee228ef411ff536ea5c58c9a2fe4790fcc250d1ec97719f5c2b6051983"
635
+ },
636
+ {
637
+ "step": 68,
638
+ "phase": "full",
639
+ "loss": 0.7479757890105247,
640
+ "questions": 16,
641
+ "microbatches": 4,
642
+ "elapsed_seconds": 56.6233723340556,
643
+ "batch_question_ids_sha256": "ebcd0c069c1aad6db15af178bfb86b95232c842c724cae991202c127dddf4df4"
644
+ },
645
+ {
646
+ "step": 69,
647
+ "phase": "full",
648
+ "loss": 0.5635014213621616,
649
+ "questions": 16,
650
+ "microbatches": 4,
651
+ "elapsed_seconds": 57.35782457515597,
652
+ "batch_question_ids_sha256": "021f795640c8ebe5b45f63378c3fb38ba2c849f5d70b747f653482f046de3fba"
653
+ },
654
+ {
655
+ "step": 70,
656
+ "phase": "full",
657
+ "loss": 0.6172329038381577,
658
+ "questions": 16,
659
+ "microbatches": 4,
660
+ "elapsed_seconds": 58.09276952967048,
661
+ "batch_question_ids_sha256": "968367da65a754ce0d7eb72abb8cbb64b558c86aa9d8adda10e119723caa2bb9"
662
+ },
663
+ {
664
+ "step": 71,
665
+ "phase": "full",
666
+ "loss": 0.3874517008662224,
667
+ "questions": 16,
668
+ "microbatches": 4,
669
+ "elapsed_seconds": 58.833030235022306,
670
+ "batch_question_ids_sha256": "5f2d81f647af85f4a48f30464c10a61c6a85b7d5d35f687d527ade870b0c1e7e"
671
+ },
672
+ {
673
+ "step": 72,
674
+ "phase": "full",
675
+ "loss": 0.7938513308763504,
676
+ "questions": 16,
677
+ "microbatches": 4,
678
+ "elapsed_seconds": 59.57471597567201,
679
+ "batch_question_ids_sha256": "6780ef3f330f6cd45a377b3f3c501ecc3ef2ae65d08769f207f4a68c152e3dbe"
680
+ },
681
+ {
682
+ "step": 73,
683
+ "phase": "full",
684
+ "loss": 0.3528990224003792,
685
+ "questions": 16,
686
+ "microbatches": 4,
687
+ "elapsed_seconds": 60.321714212186635,
688
+ "batch_question_ids_sha256": "d521c674f97a615228053a262b84518189f9580e2c73b32eee51bcef5ed76d29"
689
+ },
690
+ {
691
+ "step": 74,
692
+ "phase": "full",
693
+ "loss": 0.4417751505970955,
694
+ "questions": 16,
695
+ "microbatches": 4,
696
+ "elapsed_seconds": 61.217427681200206,
697
+ "batch_question_ids_sha256": "6c26d6bf906bf556525d65a2709f2f59a7cee21e2f62c8bad8c7015fcde12f89"
698
+ },
699
+ {
700
+ "step": 75,
701
+ "phase": "full",
702
+ "loss": 0.45803333818912506,
703
+ "questions": 16,
704
+ "microbatches": 4,
705
+ "elapsed_seconds": 61.9478549240157,
706
+ "batch_question_ids_sha256": "afabf643cddd0692428615c4aa81345acbbcdab1c52a5c9a48c815f9e2700755"
707
+ },
708
+ {
709
+ "step": 76,
710
+ "phase": "full",
711
+ "loss": 0.5286638140678406,
712
+ "questions": 16,
713
+ "microbatches": 4,
714
+ "elapsed_seconds": 62.70219527184963,
715
+ "batch_question_ids_sha256": "ba8ce0da2f777532f12642355dcce1c4d29f8e0865010f26b4bf03c56d7723d1"
716
+ },
717
+ {
718
+ "step": 77,
719
+ "phase": "full",
720
+ "loss": 0.4356435313820839,
721
+ "questions": 16,
722
+ "microbatches": 4,
723
+ "elapsed_seconds": 63.456245011650026,
724
+ "batch_question_ids_sha256": "9e0a4687654abac2570a41f8674ab95fcade2ac8e5e238d4661499dff7d58834"
725
+ },
726
+ {
727
+ "step": 78,
728
+ "phase": "full",
729
+ "loss": 0.3878141790628433,
730
+ "questions": 16,
731
+ "microbatches": 4,
732
+ "elapsed_seconds": 64.19278421159834,
733
+ "batch_question_ids_sha256": "e36ae8ba0c5eaffb1a534d69421fb9118e4f45453e90b83cbbf5f29ef284cafc"
734
+ },
735
+ {
736
+ "step": 79,
737
+ "phase": "full",
738
+ "loss": 0.4297530949115753,
739
+ "questions": 16,
740
+ "microbatches": 4,
741
+ "elapsed_seconds": 64.9118534643203,
742
+ "batch_question_ids_sha256": "171d14c1b2acfe9895a9353fe3dc68dcde6a998182a601a2746bf8fb10de20e3"
743
+ },
744
+ {
745
+ "step": 80,
746
+ "phase": "full",
747
+ "loss": 0.3846726156771183,
748
+ "questions": 16,
749
+ "microbatches": 4,
750
+ "elapsed_seconds": 65.67086058389395,
751
+ "batch_question_ids_sha256": "5091eab21841dd9384edd2c719867375591caf9d084ba03c065878eb451a5674"
752
+ },
753
+ {
754
+ "step": 81,
755
+ "phase": "full",
756
+ "loss": 0.6219327598810196,
757
+ "questions": 16,
758
+ "microbatches": 4,
759
+ "elapsed_seconds": 66.4101623352617,
760
+ "batch_question_ids_sha256": "f94cae69521603c0b60eb1cc6bb5370d2ea3d089f5b6bfd9a260b41a766ff974"
761
+ },
762
+ {
763
+ "step": 82,
764
+ "phase": "full",
765
+ "loss": 0.38340673968195915,
766
+ "questions": 16,
767
+ "microbatches": 4,
768
+ "elapsed_seconds": 67.12694064807147,
769
+ "batch_question_ids_sha256": "f804d3278fd4930c24d09af7cc2e1fc70170c35cac45cbed9779d59f2d765e08"
770
+ },
771
+ {
772
+ "step": 83,
773
+ "phase": "full",
774
+ "loss": 0.3208001106977463,
775
+ "questions": 16,
776
+ "microbatches": 4,
777
+ "elapsed_seconds": 67.86835037916899,
778
+ "batch_question_ids_sha256": "19b60c4f93e1ba7368b1f066c22c4ba32b8b66981562b2cbe2485da7433c07aa"
779
+ },
780
+ {
781
+ "step": 84,
782
+ "phase": "full",
783
+ "loss": 0.39219067245721817,
784
+ "questions": 16,
785
+ "microbatches": 4,
786
+ "elapsed_seconds": 68.58466890268028,
787
+ "batch_question_ids_sha256": "ebdf0b28c90afc00f6691d1a09af18f9f66ebaf71377d0ba210fc4892869014c"
788
+ },
789
+ {
790
+ "step": 85,
791
+ "phase": "full",
792
+ "loss": 0.3839174583554268,
793
+ "questions": 16,
794
+ "microbatches": 4,
795
+ "elapsed_seconds": 69.3296839715913,
796
+ "batch_question_ids_sha256": "4f04bf635d321e106f9856ba99a47fd6c2c033ae0450aa32c6dfe595a74ae4ed"
797
+ },
798
+ {
799
+ "step": 86,
800
+ "phase": "full",
801
+ "loss": 0.278085982427001,
802
+ "questions": 16,
803
+ "microbatches": 4,
804
+ "elapsed_seconds": 70.06139855179936,
805
+ "batch_question_ids_sha256": "ba9c22de6939f9c021221fe5a43ff5f50b15a70b21ed4effe8d97caf79b049ca"
806
+ },
807
+ {
808
+ "step": 87,
809
+ "phase": "full",
810
+ "loss": 0.3127078302204609,
811
+ "questions": 16,
812
+ "microbatches": 4,
813
+ "elapsed_seconds": 70.79180381260812,
814
+ "batch_question_ids_sha256": "c9f35e67c5374cdfd4ecee262a8723f0f688d0387f073072324c9d259d9f4323"
815
+ },
816
+ {
817
+ "step": 88,
818
+ "phase": "full",
819
+ "loss": 0.5134549420326948,
820
+ "questions": 16,
821
+ "microbatches": 4,
822
+ "elapsed_seconds": 71.52020991500467,
823
+ "batch_question_ids_sha256": "64ab5437c436a5c934bb05696c9b73f14c9838d3b210eeeb1114438303808400"
824
+ },
825
+ {
826
+ "step": 89,
827
+ "phase": "full",
828
+ "loss": 0.2528883069753647,
829
+ "questions": 16,
830
+ "microbatches": 4,
831
+ "elapsed_seconds": 72.25572444312274,
832
+ "batch_question_ids_sha256": "6ff06537c7c8217ebac7a64762ff5ee340f367a0f7652d25d5cd799c0d15cbac"
833
+ },
834
+ {
835
+ "step": 90,
836
+ "phase": "full",
837
+ "loss": 0.3719199150800705,
838
+ "questions": 16,
839
+ "microbatches": 4,
840
+ "elapsed_seconds": 72.99558722041547,
841
+ "batch_question_ids_sha256": "8fd09d36031bf8ee878da3b38a0694c5e416636fcd74e603f03dd205a9a3ab33"
842
+ },
843
+ {
844
+ "step": 91,
845
+ "phase": "full",
846
+ "loss": 0.3900993410497904,
847
+ "questions": 16,
848
+ "microbatches": 4,
849
+ "elapsed_seconds": 73.72918267734349,
850
+ "batch_question_ids_sha256": "322fb22dffd39fb986e5bfd20ac78a9963d2d5b01978598e3ea9ff285c9b09d3"
851
+ },
852
+ {
853
+ "step": 92,
854
+ "phase": "full",
855
+ "loss": 0.3622107747942209,
856
+ "questions": 16,
857
+ "microbatches": 4,
858
+ "elapsed_seconds": 74.43911021761596,
859
+ "batch_question_ids_sha256": "d0417dd44fb2608169bd7f2ab17d57acd203be7f8242f3c2e89ad317009a2e56"
860
+ },
861
+ {
862
+ "step": 93,
863
+ "phase": "full",
864
+ "loss": 0.37237807689234614,
865
+ "questions": 16,
866
+ "microbatches": 4,
867
+ "elapsed_seconds": 75.17962502967566,
868
+ "batch_question_ids_sha256": "57be67994833e0a92fac3ba6c7b6b1fe3b8383d53284136aad0b9a926f82aa1f"
869
+ },
870
+ {
871
+ "step": 94,
872
+ "phase": "full",
873
+ "loss": 0.26953236758708954,
874
+ "questions": 16,
875
+ "microbatches": 4,
876
+ "elapsed_seconds": 75.93803230766207,
877
+ "batch_question_ids_sha256": "9c3cbfa9806ac2c618d0a1e3ca3ef41c91b3a86c8aa45a36071383124ee71262"
878
+ },
879
+ {
880
+ "step": 95,
881
+ "phase": "full",
882
+ "loss": 0.40398708917200565,
883
+ "questions": 16,
884
+ "microbatches": 4,
885
+ "elapsed_seconds": 76.67349392175674,
886
+ "batch_question_ids_sha256": "100fb8785fb8c78e98988f555871d3d9a4ed2716c9bff93597e8a64de5e2db46"
887
+ },
888
+ {
889
+ "step": 96,
890
+ "phase": "full",
891
+ "loss": 0.34487732127308846,
892
+ "questions": 16,
893
+ "microbatches": 4,
894
+ "elapsed_seconds": 77.40921703819185,
895
+ "batch_question_ids_sha256": "c781c81a212d6bc7153d4ecdfb49ff5f6570c3bde61d9f7babaa983e433d7348"
896
+ },
897
+ {
898
+ "step": 97,
899
+ "phase": "full",
900
+ "loss": 0.3990098088979721,
901
+ "questions": 16,
902
+ "microbatches": 4,
903
+ "elapsed_seconds": 78.14464715123177,
904
+ "batch_question_ids_sha256": "8b19c5c7751e07c22469266bfa40bb1eaac4670fe203a8c3e8a289350956893c"
905
+ },
906
+ {
907
+ "step": 98,
908
+ "phase": "full",
909
+ "loss": 0.4635494127869606,
910
+ "questions": 16,
911
+ "microbatches": 4,
912
+ "elapsed_seconds": 78.87295348849148,
913
+ "batch_question_ids_sha256": "043573f3a457c1f8cfbcefabfef11e430a1dd573e779dd7e19f7dc26a4a1ebad"
914
+ },
915
+ {
916
+ "step": 99,
917
+ "phase": "full",
918
+ "loss": 0.3696139045059681,
919
+ "questions": 16,
920
+ "microbatches": 4,
921
+ "elapsed_seconds": 79.60388481710106,
922
+ "batch_question_ids_sha256": "933b002848b57b1ae58768d4d38da9823ca19226389dff7c41f6ba3e58ff14b6"
923
+ },
924
+ {
925
+ "step": 100,
926
+ "phase": "full",
927
+ "loss": 0.26828806940466166,
928
+ "questions": 16,
929
+ "microbatches": 4,
930
+ "elapsed_seconds": 80.44768039882183,
931
+ "batch_question_ids_sha256": "2a2b48544fe4380e81979729c21f00410faff377b407c4afdee1bef7530a71af",
932
+ "dev": {
933
+ "questions": 372,
934
+ "objective": "observed_outcome",
935
+ "eligible_questions": 372,
936
+ "target_ce": 0.47277030621569244,
937
+ "target_kl": 0.47277030621569244,
938
+ "target_tv": 0.26819442928540077,
939
+ "by_target_kind": {
940
+ "observed_outcome": {
941
+ "n": 372,
942
+ "ce": 0.47277030621569244,
943
+ "kl": 0.47277030621569244,
944
+ "tv": 0.26819442928540077
945
+ }
946
+ },
947
+ "independent_targets": {
948
+ "teacher": {
949
+ "n": 0,
950
+ "ce": 0.0,
951
+ "kl": 0.0,
952
+ "tv": 0.0
953
+ },
954
+ "gold_distribution": {
955
+ "n": 372,
956
+ "ce": 0.4926936688108985,
957
+ "kl": 0.22269107620336712,
958
+ "tv": 0.1714566637835109
959
+ }
960
+ },
961
+ "note": "question-mean distribution recovery; policy targets and reference argmax labels are not observed outcomes"
962
+ }
963
+ }
964
+ ]
variants/events_paired_seed17/backbone_config/config.json ADDED
@@ -0,0 +1,63 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ {
2
+ "architectures": [
3
+ "Qwen3ForCausalLM"
4
+ ],
5
+ "attention_bias": false,
6
+ "attention_dropout": 0.0,
7
+ "bos_token_id": 151643,
8
+ "dtype": "float32",
9
+ "eos_token_id": 151645,
10
+ "head_dim": 128,
11
+ "hidden_act": "silu",
12
+ "hidden_size": 1024,
13
+ "initializer_range": 0.02,
14
+ "intermediate_size": 3072,
15
+ "layer_types": [
16
+ "full_attention",
17
+ "full_attention",
18
+ "full_attention",
19
+ "full_attention",
20
+ "full_attention",
21
+ "full_attention",
22
+ "full_attention",
23
+ "full_attention",
24
+ "full_attention",
25
+ "full_attention",
26
+ "full_attention",
27
+ "full_attention",
28
+ "full_attention",
29
+ "full_attention",
30
+ "full_attention",
31
+ "full_attention",
32
+ "full_attention",
33
+ "full_attention",
34
+ "full_attention",
35
+ "full_attention",
36
+ "full_attention",
37
+ "full_attention",
38
+ "full_attention",
39
+ "full_attention",
40
+ "full_attention",
41
+ "full_attention",
42
+ "full_attention",
43
+ "full_attention"
44
+ ],
45
+ "max_position_embeddings": 40960,
46
+ "max_window_layers": 28,
47
+ "model_type": "qwen3",
48
+ "num_attention_heads": 16,
49
+ "num_hidden_layers": 28,
50
+ "num_key_value_heads": 8,
51
+ "pad_token_id": null,
52
+ "rms_norm_eps": 1e-06,
53
+ "rope_parameters": {
54
+ "rope_theta": 1000000,
55
+ "rope_type": "default"
56
+ },
57
+ "sliding_window": null,
58
+ "tie_word_embeddings": true,
59
+ "transformers_version": "5.17.0",
60
+ "use_cache": false,
61
+ "use_sliding_window": false,
62
+ "vocab_size": 151936
63
+ }
variants/events_paired_seed17/best.safetensors ADDED
@@ -0,0 +1,3 @@
 
 
 
 
1
+ version https://git-lfs.github.com/spec/v1
2
+ oid sha256:bbefcf829b9e36253ef46ab076f7d8fca00ffb33762e8358dbe780b3605bb777
3
+ size 2385039280
variants/events_paired_seed17/config.json ADDED
@@ -0,0 +1,69 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ {
2
+ "input": "data/scaled_games_v4/events",
3
+ "output_dir": "runs/events_paired_seed17",
4
+ "model": "Qwen/Qwen3-0.6B",
5
+ "revision": "main",
6
+ "init_checkpoint": "/home/rwang/openjev_codex_20260917/runs/v3_teacher_coords_multi_seed17",
7
+ "objective": "observed_outcome",
8
+ "loss": "paired_brier_pg",
9
+ "reward_samples": 32,
10
+ "gradient_checkpointing": true,
11
+ "set_head": "attention",
12
+ "steps": 100,
13
+ "head_steps": 0,
14
+ "batch_questions": 16,
15
+ "microbatch_questions": 4,
16
+ "max_microbatch_tokens": 16384,
17
+ "eval_every": 50,
18
+ "max_length": 8192,
19
+ "seed": 17,
20
+ "backbone_lr": 2e-05,
21
+ "head_lr": 0.0002,
22
+ "head_warmup_lr": 0.001,
23
+ "precision": "bf16",
24
+ "disable_native_triton": true,
25
+ "validate_only": false,
26
+ "self_check": false,
27
+ "schema_version": "openjev-decision-pipeline-v1",
28
+ "resolved_model_revision": "c1899de289a04d12100db370d81485cdf75e47ca",
29
+ "initialization": "local DecisionModel warm start, fresh optimizer",
30
+ "data_sha256": {
31
+ "data/scaled_games_v4/events/train.jsonl": "4cebf698e4b71dc37d4d0e985b6577dd565ac370ac82b6f47f375aa6af69f8c9",
32
+ "data/scaled_games_v4/events/dev.jsonl": "0c352ee1064a5fb5d2aacad59cbea57afb897886d3d09e5ca108c866a5f893db",
33
+ "data/scaled_games_v4/events/calibration.jsonl": "4a8a8a2b81c3c46df6bb4c8690e97374dd0165cbef40566bb7f87c1f425d5562",
34
+ "data/scaled_games_v4/events/test.jsonl": "884bec71891668199925809d8708347f56f7bf06e8363e4a436347967b1423e3",
35
+ "data/scaled_games_v4/events/ood.jsonl": "3c00d845524634ca3b34559761cdaf0c52b7b7633d5517d1e7ee7f31697cea48"
36
+ },
37
+ "deps": {
38
+ "torch": "2.14.0",
39
+ "transformers": "5.17.0",
40
+ "safetensors": "0.8.0"
41
+ },
42
+ "gpu": "NVIDIA A100-SXM4-80GB",
43
+ "parameter_storage": "float32",
44
+ "forward_autocast": "bfloat16",
45
+ "parameter_count": 596250498,
46
+ "train_questions": 1124,
47
+ "all_train_questions": 1124,
48
+ "schema_counts": {
49
+ "records": 2356,
50
+ "questions_by_split": {
51
+ "train": 1124,
52
+ "dev": 372,
53
+ "calibration": 368,
54
+ "test": 364,
55
+ "ood": 128
56
+ },
57
+ "eligible_by_split_objective": {
58
+ "calibration/gold_distribution": 368,
59
+ "dev/gold_distribution": 372,
60
+ "ood/gold_distribution": 128,
61
+ "test/gold_distribution": 364,
62
+ "train/gold_distribution": 1124
63
+ }
64
+ },
65
+ "loss_description": "equal question weight; complete candidate normalization; sampled loss is a score-function surrogate",
66
+ "rlcd_candidate": "independent paired categorical proper-reward policy gradient",
67
+ "selection": "minimum dev target CE; held-out test first evaluated after training and checkpoint selection",
68
+ "parallelism": "complete candidate paths batched in one backbone call per microbatch; no prefix sharing"
69
+ }
variants/events_paired_seed17/summary.json ADDED
@@ -0,0 +1,136 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ {
2
+ "best_step": 100,
3
+ "best_dev_target_ce": 0.45392147839038,
4
+ "selected_on": "dev target CE",
5
+ "objective": "observed_outcome",
6
+ "metrics_by_split": {
7
+ "dev": {
8
+ "questions": 372,
9
+ "objective": "observed_outcome",
10
+ "eligible_questions": 372,
11
+ "target_ce": 0.45392147839038,
12
+ "target_kl": 0.45392147839038,
13
+ "target_tv": 0.2908940469690758,
14
+ "by_target_kind": {
15
+ "observed_outcome": {
16
+ "n": 372,
17
+ "ce": 0.45392147839038,
18
+ "kl": 0.45392147839038,
19
+ "tv": 0.2908940469690758
20
+ }
21
+ },
22
+ "independent_targets": {
23
+ "teacher": {
24
+ "n": 0,
25
+ "ce": 0.0,
26
+ "kl": 0.0,
27
+ "tv": 0.0
28
+ },
29
+ "gold_distribution": {
30
+ "n": 372,
31
+ "ce": 0.45846966947460877,
32
+ "kl": 0.18846707686707742,
33
+ "tv": 0.18377175030687837
34
+ }
35
+ },
36
+ "note": "question-mean distribution recovery; policy targets and reference argmax labels are not observed outcomes"
37
+ },
38
+ "calibration": {
39
+ "questions": 368,
40
+ "objective": "observed_outcome",
41
+ "eligible_questions": 368,
42
+ "target_ce": 0.4863960135417897,
43
+ "target_kl": 0.4863960135417897,
44
+ "target_tv": 0.3079795091580387,
45
+ "by_target_kind": {
46
+ "observed_outcome": {
47
+ "n": 368,
48
+ "ce": 0.4863960135417897,
49
+ "kl": 0.4863960135417897,
50
+ "tv": 0.3079795091580387
51
+ }
52
+ },
53
+ "independent_targets": {
54
+ "teacher": {
55
+ "n": 0,
56
+ "ce": 0.0,
57
+ "kl": 0.0,
58
+ "tv": 0.0
59
+ },
60
+ "gold_distribution": {
61
+ "n": 368,
62
+ "ce": 0.46718815035607203,
63
+ "kl": 0.1808918312753622,
64
+ "tv": 0.1788949192262398
65
+ }
66
+ },
67
+ "note": "question-mean distribution recovery; policy targets and reference argmax labels are not observed outcomes"
68
+ },
69
+ "test": {
70
+ "questions": 364,
71
+ "objective": "observed_outcome",
72
+ "eligible_questions": 364,
73
+ "target_ce": 0.42791067276663786,
74
+ "target_kl": 0.42791067276663786,
75
+ "target_tv": 0.2817070426047916,
76
+ "by_target_kind": {
77
+ "observed_outcome": {
78
+ "n": 364,
79
+ "ce": 0.42791067276663786,
80
+ "kl": 0.42791067276663786,
81
+ "tv": 0.2817070426047916
82
+ }
83
+ },
84
+ "independent_targets": {
85
+ "teacher": {
86
+ "n": 0,
87
+ "ce": 0.0,
88
+ "kl": 0.0,
89
+ "tv": 0.0
90
+ },
91
+ "gold_distribution": {
92
+ "n": 364,
93
+ "ce": 0.4476948609498736,
94
+ "kl": 0.18692406062375522,
95
+ "tv": 0.18073862322396844
96
+ }
97
+ },
98
+ "note": "question-mean distribution recovery; policy targets and reference argmax labels are not observed outcomes"
99
+ },
100
+ "ood": {
101
+ "questions": 128,
102
+ "objective": "observed_outcome",
103
+ "eligible_questions": 128,
104
+ "target_ce": 0.3296768191800103,
105
+ "target_kl": 0.3296768191800103,
106
+ "target_tv": 0.24400384461107302,
107
+ "by_target_kind": {
108
+ "observed_outcome": {
109
+ "n": 128,
110
+ "ce": 0.3296768191800103,
111
+ "kl": 0.3296768191800103,
112
+ "tv": 0.24400384461107302
113
+ }
114
+ },
115
+ "independent_targets": {
116
+ "teacher": {
117
+ "n": 0,
118
+ "ce": 0.0,
119
+ "kl": 0.0,
120
+ "tv": 0.0
121
+ },
122
+ "gold_distribution": {
123
+ "n": 128,
124
+ "ce": 0.3458871208971331,
125
+ "kl": 0.1123998992605057,
126
+ "tv": 0.139747922104369
127
+ }
128
+ },
129
+ "note": "question-mean distribution recovery; policy targets and reference argmax labels are not observed outcomes"
130
+ }
131
+ },
132
+ "temperature": 1.0,
133
+ "temperature_fitted": false,
134
+ "training_seconds": 100.81704722624272,
135
+ "max_gpu_allocated_gb": 11.964137984
136
+ }
variants/events_paired_seed17/target_audit.json ADDED
The diff for this file is too large to render. See raw diff
 
variants/events_paired_seed17/tokenizer/chat_template.jinja ADDED
@@ -0,0 +1,89 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ {%- if tools %}
2
+ {{- '<|im_start|>system\n' }}
3
+ {%- if messages[0].role == 'system' %}
4
+ {{- messages[0].content + '\n\n' }}
5
+ {%- endif %}
6
+ {{- "# Tools\n\nYou may call one or more functions to assist with the user query.\n\nYou are provided with function signatures within <tools></tools> XML tags:\n<tools>" }}
7
+ {%- for tool in tools %}
8
+ {{- "\n" }}
9
+ {{- tool | tojson }}
10
+ {%- endfor %}
11
+ {{- "\n</tools>\n\nFor each function call, return a json object with function name and arguments within <tool_call></tool_call> XML tags:\n<tool_call>\n{\"name\": <function-name>, \"arguments\": <args-json-object>}\n</tool_call><|im_end|>\n" }}
12
+ {%- else %}
13
+ {%- if messages[0].role == 'system' %}
14
+ {{- '<|im_start|>system\n' + messages[0].content + '<|im_end|>\n' }}
15
+ {%- endif %}
16
+ {%- endif %}
17
+ {%- set ns = namespace(multi_step_tool=true, last_query_index=messages|length - 1) %}
18
+ {%- for message in messages[::-1] %}
19
+ {%- set index = (messages|length - 1) - loop.index0 %}
20
+ {%- if ns.multi_step_tool and message.role == "user" and message.content is string and not(message.content.startswith('<tool_response>') and message.content.endswith('</tool_response>')) %}
21
+ {%- set ns.multi_step_tool = false %}
22
+ {%- set ns.last_query_index = index %}
23
+ {%- endif %}
24
+ {%- endfor %}
25
+ {%- for message in messages %}
26
+ {%- if message.content is string %}
27
+ {%- set content = message.content %}
28
+ {%- else %}
29
+ {%- set content = '' %}
30
+ {%- endif %}
31
+ {%- if (message.role == "user") or (message.role == "system" and not loop.first) %}
32
+ {{- '<|im_start|>' + message.role + '\n' + content + '<|im_end|>' + '\n' }}
33
+ {%- elif message.role == "assistant" %}
34
+ {%- set reasoning_content = '' %}
35
+ {%- if message.reasoning_content is string %}
36
+ {%- set reasoning_content = message.reasoning_content %}
37
+ {%- else %}
38
+ {%- if '</think>' in content %}
39
+ {%- set reasoning_content = content.split('</think>')[0].rstrip('\n').split('<think>')[-1].lstrip('\n') %}
40
+ {%- set content = content.split('</think>')[-1].lstrip('\n') %}
41
+ {%- endif %}
42
+ {%- endif %}
43
+ {%- if loop.index0 > ns.last_query_index %}
44
+ {%- if loop.last or (not loop.last and reasoning_content) %}
45
+ {{- '<|im_start|>' + message.role + '\n<think>\n' + reasoning_content.strip('\n') + '\n</think>\n\n' + content.lstrip('\n') }}
46
+ {%- else %}
47
+ {{- '<|im_start|>' + message.role + '\n' + content }}
48
+ {%- endif %}
49
+ {%- else %}
50
+ {{- '<|im_start|>' + message.role + '\n' + content }}
51
+ {%- endif %}
52
+ {%- if message.tool_calls %}
53
+ {%- for tool_call in message.tool_calls %}
54
+ {%- if (loop.first and content) or (not loop.first) %}
55
+ {{- '\n' }}
56
+ {%- endif %}
57
+ {%- if tool_call.function %}
58
+ {%- set tool_call = tool_call.function %}
59
+ {%- endif %}
60
+ {{- '<tool_call>\n{"name": "' }}
61
+ {{- tool_call.name }}
62
+ {{- '", "arguments": ' }}
63
+ {%- if tool_call.arguments is string %}
64
+ {{- tool_call.arguments }}
65
+ {%- else %}
66
+ {{- tool_call.arguments | tojson }}
67
+ {%- endif %}
68
+ {{- '}\n</tool_call>' }}
69
+ {%- endfor %}
70
+ {%- endif %}
71
+ {{- '<|im_end|>\n' }}
72
+ {%- elif message.role == "tool" %}
73
+ {%- if loop.first or (messages[loop.index0 - 1].role != "tool") %}
74
+ {{- '<|im_start|>user' }}
75
+ {%- endif %}
76
+ {{- '\n<tool_response>\n' }}
77
+ {{- content }}
78
+ {{- '\n</tool_response>' }}
79
+ {%- if loop.last or (messages[loop.index0 + 1].role != "tool") %}
80
+ {{- '<|im_end|>\n' }}
81
+ {%- endif %}
82
+ {%- endif %}
83
+ {%- endfor %}
84
+ {%- if add_generation_prompt %}
85
+ {{- '<|im_start|>assistant\n' }}
86
+ {%- if enable_thinking is defined and enable_thinking is false %}
87
+ {{- '<think>\n\n</think>\n\n' }}
88
+ {%- endif %}
89
+ {%- endif %}
variants/events_paired_seed17/tokenizer/tokenizer.json ADDED
@@ -0,0 +1,3 @@
 
 
 
 
1
+ version https://git-lfs.github.com/spec/v1
2
+ oid sha256:be75606093db2094d7cd20f3c2f385c212750648bd6ea4fb2bf507a6a4c55506
3
+ size 11422650
variants/events_paired_seed17/tokenizer/tokenizer_config.json ADDED
@@ -0,0 +1,30 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ {
2
+ "add_prefix_space": false,
3
+ "backend": "tokenizers",
4
+ "bos_token": null,
5
+ "clean_up_tokenization_spaces": false,
6
+ "eos_token": "<|im_end|>",
7
+ "errors": "replace",
8
+ "extra_special_tokens": [
9
+ "<|im_start|>",
10
+ "<|im_end|>",
11
+ "<|object_ref_start|>",
12
+ "<|object_ref_end|>",
13
+ "<|box_start|>",
14
+ "<|box_end|>",
15
+ "<|quad_start|>",
16
+ "<|quad_end|>",
17
+ "<|vision_start|>",
18
+ "<|vision_end|>",
19
+ "<|vision_pad|>",
20
+ "<|image_pad|>",
21
+ "<|video_pad|>"
22
+ ],
23
+ "is_local": true,
24
+ "local_files_only": true,
25
+ "model_max_length": 131072,
26
+ "pad_token": "<|endoftext|>",
27
+ "split_special_tokens": false,
28
+ "tokenizer_class": "Qwen2Tokenizer",
29
+ "unk_token": null
30
+ }
variants/events_paired_seed17/train_log.json ADDED
@@ -0,0 +1,964 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ [
2
+ {
3
+ "step": 1,
4
+ "phase": "full",
5
+ "loss": -0.0748950969427824,
6
+ "questions": 16,
7
+ "microbatches": 4,
8
+ "elapsed_seconds": 1.1360169062390924,
9
+ "batch_question_ids_sha256": "9815e3084b0060ca1bf3f53e09caa2e8ba2ecfbfa18cd5a90c1b6ca1dc29fbcb"
10
+ },
11
+ {
12
+ "step": 2,
13
+ "phase": "full",
14
+ "loss": 0.22397566214203835,
15
+ "questions": 16,
16
+ "microbatches": 4,
17
+ "elapsed_seconds": 1.9364449437707663,
18
+ "batch_question_ids_sha256": "0766e00bccd0b9d63535226a2a7454d88d591d6a1adb32befcd64d40174bd87c"
19
+ },
20
+ {
21
+ "step": 3,
22
+ "phase": "full",
23
+ "loss": -0.009121095761656761,
24
+ "questions": 16,
25
+ "microbatches": 4,
26
+ "elapsed_seconds": 2.6859495593234897,
27
+ "batch_question_ids_sha256": "3eb8fb253ed8d59b98fd13fb12ae3aa08473ff98bf07f26c7537d8d4c6999318"
28
+ },
29
+ {
30
+ "step": 4,
31
+ "phase": "full",
32
+ "loss": -0.04071702528744936,
33
+ "questions": 16,
34
+ "microbatches": 4,
35
+ "elapsed_seconds": 3.4203991554677486,
36
+ "batch_question_ids_sha256": "b72b9555459b9660b8cfea9c061b0d69e21b213181cf19973fa9312c1d740e4c"
37
+ },
38
+ {
39
+ "step": 5,
40
+ "phase": "full",
41
+ "loss": -0.030815408565104008,
42
+ "questions": 16,
43
+ "microbatches": 4,
44
+ "elapsed_seconds": 4.176274714060128,
45
+ "batch_question_ids_sha256": "851370879eecc2bb2b872bf9f26a90b0c422dbb910faa7ab99625462dd70126c"
46
+ },
47
+ {
48
+ "step": 6,
49
+ "phase": "full",
50
+ "loss": -0.003552184090949595,
51
+ "questions": 16,
52
+ "microbatches": 4,
53
+ "elapsed_seconds": 4.941503707319498,
54
+ "batch_question_ids_sha256": "a44ac9fef40754316a48915525e55ee3c8c5f613d2a3a206a70c53f6a360d6a1"
55
+ },
56
+ {
57
+ "step": 7,
58
+ "phase": "full",
59
+ "loss": 0.005697006359696388,
60
+ "questions": 16,
61
+ "microbatches": 4,
62
+ "elapsed_seconds": 5.700752309523523,
63
+ "batch_question_ids_sha256": "dd15d5df027d50d8bc83cc139ce58f36adf2613954b88024320afffaf4f275ce"
64
+ },
65
+ {
66
+ "step": 8,
67
+ "phase": "full",
68
+ "loss": 0.057425894298830826,
69
+ "questions": 16,
70
+ "microbatches": 4,
71
+ "elapsed_seconds": 6.455997716635466,
72
+ "batch_question_ids_sha256": "b16e21ce00030d172555d2ed47f3935798c5293130655f2d4bb114fab8d4f82e"
73
+ },
74
+ {
75
+ "step": 9,
76
+ "phase": "full",
77
+ "loss": -0.03418128867633641,
78
+ "questions": 16,
79
+ "microbatches": 4,
80
+ "elapsed_seconds": 7.2420299937948585,
81
+ "batch_question_ids_sha256": "db13a8f636762d4e5c56f60dcecdc4a04f7acccee36ed4846fe242c110813e33"
82
+ },
83
+ {
84
+ "step": 10,
85
+ "phase": "full",
86
+ "loss": 0.04701452422887087,
87
+ "questions": 16,
88
+ "microbatches": 4,
89
+ "elapsed_seconds": 8.022835675626993,
90
+ "batch_question_ids_sha256": "d29a2ab7bae9372a6460e6d41d2320b0a13b55633fcc21e2a06f06a64a2b8f22"
91
+ },
92
+ {
93
+ "step": 11,
94
+ "phase": "full",
95
+ "loss": -0.004810761660337448,
96
+ "questions": 16,
97
+ "microbatches": 4,
98
+ "elapsed_seconds": 8.767681133002043,
99
+ "batch_question_ids_sha256": "e435c49e7cce06b5c387742c1cf2eae1fb3f57349a03ff3229713bf8e9179b7f"
100
+ },
101
+ {
102
+ "step": 12,
103
+ "phase": "full",
104
+ "loss": 0.13739927113056183,
105
+ "questions": 16,
106
+ "microbatches": 4,
107
+ "elapsed_seconds": 9.533075327984989,
108
+ "batch_question_ids_sha256": "52067a44724ec88a3f85cace2ddba01f511bb396cbf27877697e85efd4f0c68e"
109
+ },
110
+ {
111
+ "step": 13,
112
+ "phase": "full",
113
+ "loss": -0.07629979046760127,
114
+ "questions": 16,
115
+ "microbatches": 4,
116
+ "elapsed_seconds": 10.30644958280027,
117
+ "batch_question_ids_sha256": "d01e765c3bf9b2b1edf10d82d3b784f51514a21682d040badd6d9c1c29c3069f"
118
+ },
119
+ {
120
+ "step": 14,
121
+ "phase": "full",
122
+ "loss": 0.06013527140021324,
123
+ "questions": 16,
124
+ "microbatches": 4,
125
+ "elapsed_seconds": 11.06470196787268,
126
+ "batch_question_ids_sha256": "f7d76860a7e7abbd079278179b2a6054aa8b1bf90c6e38db91f414cf906eff6a"
127
+ },
128
+ {
129
+ "step": 15,
130
+ "phase": "full",
131
+ "loss": 0.12951497547328472,
132
+ "questions": 16,
133
+ "microbatches": 4,
134
+ "elapsed_seconds": 11.837062750943005,
135
+ "batch_question_ids_sha256": "1d089d97e3e6836ec1771dba1776d38f6a8ca414f96cc3c25fd0053ed3e27a44"
136
+ },
137
+ {
138
+ "step": 16,
139
+ "phase": "full",
140
+ "loss": -0.11920184083282948,
141
+ "questions": 16,
142
+ "microbatches": 4,
143
+ "elapsed_seconds": 12.592718491330743,
144
+ "batch_question_ids_sha256": "abd204307499e62673daa334cd3a4533448a0ef971b517e7acbf23783905ae91"
145
+ },
146
+ {
147
+ "step": 17,
148
+ "phase": "full",
149
+ "loss": -0.003990098484791815,
150
+ "questions": 16,
151
+ "microbatches": 4,
152
+ "elapsed_seconds": 13.34634671639651,
153
+ "batch_question_ids_sha256": "6d4029fa8044653ff0ccbe80e4371cccd55b6fc7c6462fb81d77f565bde1e2f5"
154
+ },
155
+ {
156
+ "step": 18,
157
+ "phase": "full",
158
+ "loss": 0.0632672868669033,
159
+ "questions": 16,
160
+ "microbatches": 4,
161
+ "elapsed_seconds": 14.103745453059673,
162
+ "batch_question_ids_sha256": "88ddecb0e43b996acf8e321896454ceb6e79095126d68380cfdc7c74ba66d04a"
163
+ },
164
+ {
165
+ "step": 19,
166
+ "phase": "full",
167
+ "loss": 0.042018274776637554,
168
+ "questions": 16,
169
+ "microbatches": 4,
170
+ "elapsed_seconds": 14.861994631588459,
171
+ "batch_question_ids_sha256": "3ba3b72125f7617c5cb4dfc031843bc10773d8aa7747ef6eef76ec813abc7a11"
172
+ },
173
+ {
174
+ "step": 20,
175
+ "phase": "full",
176
+ "loss": 0.05378047423437238,
177
+ "questions": 16,
178
+ "microbatches": 4,
179
+ "elapsed_seconds": 15.614400346763432,
180
+ "batch_question_ids_sha256": "a8fbfc400a3dd61e2468ddde908d8e6914e8742e66df75fa209575bbd49f674f"
181
+ },
182
+ {
183
+ "step": 21,
184
+ "phase": "full",
185
+ "loss": 0.020703650545328856,
186
+ "questions": 16,
187
+ "microbatches": 4,
188
+ "elapsed_seconds": 16.42179041542113,
189
+ "batch_question_ids_sha256": "49a7b1f6bb6891cd06f7c495942bbdbd98cfbd7435f49d77922ab364b9c1c5f5"
190
+ },
191
+ {
192
+ "step": 22,
193
+ "phase": "full",
194
+ "loss": -0.010295026702806354,
195
+ "questions": 16,
196
+ "microbatches": 4,
197
+ "elapsed_seconds": 17.1975815044716,
198
+ "batch_question_ids_sha256": "d93e27a20a5666b8fe2ea5489dac0a58f99a770bfb0d4ee6870d36621ed78dee"
199
+ },
200
+ {
201
+ "step": 23,
202
+ "phase": "full",
203
+ "loss": -0.031696483260020614,
204
+ "questions": 16,
205
+ "microbatches": 4,
206
+ "elapsed_seconds": 18.08850848581642,
207
+ "batch_question_ids_sha256": "88845418c486e4c7d856761b8aca680a7ca52ec59c84219e708745af76a445c2"
208
+ },
209
+ {
210
+ "step": 24,
211
+ "phase": "full",
212
+ "loss": 0.0040627864527778,
213
+ "questions": 16,
214
+ "microbatches": 4,
215
+ "elapsed_seconds": 18.842839125543833,
216
+ "batch_question_ids_sha256": "b7fcb8d686d0a87ea4fd6ded60d595c7ccc4b132fc739e047564d72f5944577f"
217
+ },
218
+ {
219
+ "step": 25,
220
+ "phase": "full",
221
+ "loss": 0.004738317336887121,
222
+ "questions": 16,
223
+ "microbatches": 4,
224
+ "elapsed_seconds": 19.609677392989397,
225
+ "batch_question_ids_sha256": "b2e6af990ea722e6cb276f57a736c23766306ae29b211bf00447d868b434c962"
226
+ },
227
+ {
228
+ "step": 26,
229
+ "phase": "full",
230
+ "loss": 0.06255067884922028,
231
+ "questions": 16,
232
+ "microbatches": 4,
233
+ "elapsed_seconds": 20.376745360903442,
234
+ "batch_question_ids_sha256": "67253956a0835f33db1fa53443fed54476a027338e0ffdd7726dcb48fbef1c9a"
235
+ },
236
+ {
237
+ "step": 27,
238
+ "phase": "full",
239
+ "loss": 0.06562656164169312,
240
+ "questions": 16,
241
+ "microbatches": 4,
242
+ "elapsed_seconds": 21.134697519242764,
243
+ "batch_question_ids_sha256": "b37cd333ba07e45e13875704fb0cc42b2f44fc5317737590bb27d88a7ce61452"
244
+ },
245
+ {
246
+ "step": 28,
247
+ "phase": "full",
248
+ "loss": 0.03742673061788082,
249
+ "questions": 16,
250
+ "microbatches": 4,
251
+ "elapsed_seconds": 21.901406745426357,
252
+ "batch_question_ids_sha256": "d2a5a95e218ea89d7469dfdd8ff320de68f88ff52171675c0ccf7c9c3d95a891"
253
+ },
254
+ {
255
+ "step": 29,
256
+ "phase": "full",
257
+ "loss": 0.006606193259358406,
258
+ "questions": 16,
259
+ "microbatches": 4,
260
+ "elapsed_seconds": 22.663256482221186,
261
+ "batch_question_ids_sha256": "1c6aa40b463139c53b7a93df96c3a1db9c53fac3123be7e8dda4fc8f53bc7996"
262
+ },
263
+ {
264
+ "step": 30,
265
+ "phase": "full",
266
+ "loss": 0.0024183355271816254,
267
+ "questions": 16,
268
+ "microbatches": 4,
269
+ "elapsed_seconds": 23.401013052091002,
270
+ "batch_question_ids_sha256": "be09ca5b88c61e0617249157b2d75545b12e403abaf2ff570be1e6d30856d742"
271
+ },
272
+ {
273
+ "step": 31,
274
+ "phase": "full",
275
+ "loss": -0.010469562374055386,
276
+ "questions": 16,
277
+ "microbatches": 4,
278
+ "elapsed_seconds": 24.170286952517927,
279
+ "batch_question_ids_sha256": "08f4544d1dfa624e78df3609c0059204a16146f8421bf85c14d0b10b5926b448"
280
+ },
281
+ {
282
+ "step": 32,
283
+ "phase": "full",
284
+ "loss": -0.0184428533539176,
285
+ "questions": 16,
286
+ "microbatches": 4,
287
+ "elapsed_seconds": 24.91613855585456,
288
+ "batch_question_ids_sha256": "f8a0c95d4ce5a21d3b7bdd381eb7c9d15a251da10b33f0a8521a3e81eeb1450e"
289
+ },
290
+ {
291
+ "step": 33,
292
+ "phase": "full",
293
+ "loss": -0.026352540589869022,
294
+ "questions": 16,
295
+ "microbatches": 4,
296
+ "elapsed_seconds": 25.68266553618014,
297
+ "batch_question_ids_sha256": "4c396fcb15e0a31d52d00473dc2f3bee3767da23146fb5038b6e63d9ad5016c8"
298
+ },
299
+ {
300
+ "step": 34,
301
+ "phase": "full",
302
+ "loss": 0.005794679290374916,
303
+ "questions": 16,
304
+ "microbatches": 4,
305
+ "elapsed_seconds": 26.436185279861093,
306
+ "batch_question_ids_sha256": "c59ffed718f7ded62b1028779c65e8759de67070b26019e8642f75dd2e43a512"
307
+ },
308
+ {
309
+ "step": 35,
310
+ "phase": "full",
311
+ "loss": 0.12788009410724044,
312
+ "questions": 16,
313
+ "microbatches": 4,
314
+ "elapsed_seconds": 27.187837212346494,
315
+ "batch_question_ids_sha256": "0e54ce4db6221fac98c15ce77d50701153708881dda86adbcb2776637215c9bd"
316
+ },
317
+ {
318
+ "step": 36,
319
+ "phase": "full",
320
+ "loss": 0.09676597267389297,
321
+ "questions": 16,
322
+ "microbatches": 4,
323
+ "elapsed_seconds": 27.944907465949655,
324
+ "batch_question_ids_sha256": "24c257b893015e139aaec823c9fa8ed630a7d2df239552f229ebb832a7746723"
325
+ },
326
+ {
327
+ "step": 37,
328
+ "phase": "full",
329
+ "loss": 0.04663453111425042,
330
+ "questions": 16,
331
+ "microbatches": 4,
332
+ "elapsed_seconds": 28.70325310435146,
333
+ "batch_question_ids_sha256": "bc22d73d0ac85948fd9bbd0813992b093f9beb230652c6234447dbbbfb4add35"
334
+ },
335
+ {
336
+ "step": 38,
337
+ "phase": "full",
338
+ "loss": -0.06176069751381874,
339
+ "questions": 16,
340
+ "microbatches": 4,
341
+ "elapsed_seconds": 29.47435860335827,
342
+ "batch_question_ids_sha256": "6e47f7951fb42ce413ed2f4ce54d8832b534a3e9d0f2e35829a1608286227ef4"
343
+ },
344
+ {
345
+ "step": 39,
346
+ "phase": "full",
347
+ "loss": 0.0276058129966259,
348
+ "questions": 16,
349
+ "microbatches": 4,
350
+ "elapsed_seconds": 30.218379349447787,
351
+ "batch_question_ids_sha256": "b5bfa6d3b4f6544cb42bc8b82b7c88ad3430367ce2f47dd0c3182d7f55e26796"
352
+ },
353
+ {
354
+ "step": 40,
355
+ "phase": "full",
356
+ "loss": -0.026078838855028152,
357
+ "questions": 16,
358
+ "microbatches": 4,
359
+ "elapsed_seconds": 30.989923455752432,
360
+ "batch_question_ids_sha256": "e356d8692763396b1f305f967bc42bbea222b3adcfd7355c8ffd489d7c4a8702"
361
+ },
362
+ {
363
+ "step": 41,
364
+ "phase": "full",
365
+ "loss": 0.04557723645120859,
366
+ "questions": 16,
367
+ "microbatches": 4,
368
+ "elapsed_seconds": 31.751174583099782,
369
+ "batch_question_ids_sha256": "eddf9bafc965ad35c8a9189e6a7b19eb336e5f2736ab490e5d0f7547f885b883"
370
+ },
371
+ {
372
+ "step": 42,
373
+ "phase": "full",
374
+ "loss": -0.04995053354650736,
375
+ "questions": 16,
376
+ "microbatches": 4,
377
+ "elapsed_seconds": 32.51822272129357,
378
+ "batch_question_ids_sha256": "93f507828c7fa8715902d87edef912b05fe2920c25f400745b5d4b1835ff1fcd"
379
+ },
380
+ {
381
+ "step": 43,
382
+ "phase": "full",
383
+ "loss": 0.09411330241709948,
384
+ "questions": 16,
385
+ "microbatches": 4,
386
+ "elapsed_seconds": 33.269474141299725,
387
+ "batch_question_ids_sha256": "50f72e83de9b1b9daf154f013aa536b1454d8d8c13b503ee6efa9007c721faa6"
388
+ },
389
+ {
390
+ "step": 44,
391
+ "phase": "full",
392
+ "loss": 0.05266818264499307,
393
+ "questions": 16,
394
+ "microbatches": 4,
395
+ "elapsed_seconds": 34.04855671431869,
396
+ "batch_question_ids_sha256": "71067cb1ca0d666b5a16e53227ae1840e468c0c39e879ef8004ac73520b396f9"
397
+ },
398
+ {
399
+ "step": 45,
400
+ "phase": "full",
401
+ "loss": -0.12074423860758543,
402
+ "questions": 16,
403
+ "microbatches": 4,
404
+ "elapsed_seconds": 34.81311255134642,
405
+ "batch_question_ids_sha256": "b9270e713e4fca02ecc7205c7ac1e320d3c87286eb7237ed0b02a4686b626bc4"
406
+ },
407
+ {
408
+ "step": 46,
409
+ "phase": "full",
410
+ "loss": 0.001318799564614892,
411
+ "questions": 16,
412
+ "microbatches": 4,
413
+ "elapsed_seconds": 35.60654359217733,
414
+ "batch_question_ids_sha256": "fbecc4d2e806b5b42f4f82a7e99465e1fceb5b7b61dfc3cd167654deeef897d0"
415
+ },
416
+ {
417
+ "step": 47,
418
+ "phase": "full",
419
+ "loss": -0.028849713504314423,
420
+ "questions": 16,
421
+ "microbatches": 4,
422
+ "elapsed_seconds": 36.41548177506775,
423
+ "batch_question_ids_sha256": "9fb1e3f8e3eb222e0e377eeedad5b0df90d11389465803b67a9b5090d2e79ca0"
424
+ },
425
+ {
426
+ "step": 48,
427
+ "phase": "full",
428
+ "loss": -0.015318945050239563,
429
+ "questions": 16,
430
+ "microbatches": 4,
431
+ "elapsed_seconds": 37.20380338933319,
432
+ "batch_question_ids_sha256": "8ec776449fb8174bd55b829feff0711ca000a9e09d903e741da6f91e2d21a136"
433
+ },
434
+ {
435
+ "step": 49,
436
+ "phase": "full",
437
+ "loss": -0.04532638657838106,
438
+ "questions": 16,
439
+ "microbatches": 4,
440
+ "elapsed_seconds": 38.08001133892685,
441
+ "batch_question_ids_sha256": "a20669c6cee5e157b8b54729bb88921f4ff6df89b4dda467ec25abb10f0e78da"
442
+ },
443
+ {
444
+ "step": 50,
445
+ "phase": "full",
446
+ "loss": -0.04311649454757571,
447
+ "questions": 16,
448
+ "microbatches": 4,
449
+ "elapsed_seconds": 38.827290212735534,
450
+ "batch_question_ids_sha256": "cab2a2f34731e8a959300614046a8b8fc4c926971dd5909a1a6e3de0ad3acb26",
451
+ "dev": {
452
+ "questions": 372,
453
+ "objective": "observed_outcome",
454
+ "eligible_questions": 372,
455
+ "target_ce": 0.46014975992706547,
456
+ "target_kl": 0.46014975992706547,
457
+ "target_tv": 0.307618331572389,
458
+ "by_target_kind": {
459
+ "observed_outcome": {
460
+ "n": 372,
461
+ "ce": 0.46014975992706547,
462
+ "kl": 0.46014975992706547,
463
+ "tv": 0.307618331572389
464
+ }
465
+ },
466
+ "independent_targets": {
467
+ "teacher": {
468
+ "n": 0,
469
+ "ce": 0.0,
470
+ "kl": 0.0,
471
+ "tv": 0.0
472
+ },
473
+ "gold_distribution": {
474
+ "n": 372,
475
+ "ce": 0.4655643873714251,
476
+ "kl": 0.1955617947638946,
477
+ "tv": 0.19696661073944272
478
+ }
479
+ },
480
+ "note": "question-mean distribution recovery; policy targets and reference argmax labels are not observed outcomes"
481
+ }
482
+ },
483
+ {
484
+ "step": 51,
485
+ "phase": "full",
486
+ "loss": 0.04189913161098957,
487
+ "questions": 16,
488
+ "microbatches": 4,
489
+ "elapsed_seconds": 45.038797207176685,
490
+ "batch_question_ids_sha256": "46758e1c6996a3f1a932a11e8372539949f8ac8f4c0415b36f65c713e195096f"
491
+ },
492
+ {
493
+ "step": 52,
494
+ "phase": "full",
495
+ "loss": -0.07959715370088816,
496
+ "questions": 16,
497
+ "microbatches": 4,
498
+ "elapsed_seconds": 45.79904679581523,
499
+ "batch_question_ids_sha256": "d87fcd81d136c657c13e49aa033db193bfbb121e4a721fa86c16aad4c0a7c39d"
500
+ },
501
+ {
502
+ "step": 53,
503
+ "phase": "full",
504
+ "loss": -0.051664914935827255,
505
+ "questions": 16,
506
+ "microbatches": 4,
507
+ "elapsed_seconds": 46.56836490239948,
508
+ "batch_question_ids_sha256": "4fae217aef8823502570d85ef34334cdfe2994ba567eef62990a60ac5e3ede4a"
509
+ },
510
+ {
511
+ "step": 54,
512
+ "phase": "full",
513
+ "loss": 0.08312265900895,
514
+ "questions": 16,
515
+ "microbatches": 4,
516
+ "elapsed_seconds": 47.331904822960496,
517
+ "batch_question_ids_sha256": "edbcfec53ad977392c69172be6901bdb4eee1725bf8fd83582f64ac6320e77cd"
518
+ },
519
+ {
520
+ "step": 55,
521
+ "phase": "full",
522
+ "loss": 0.011041590012609959,
523
+ "questions": 16,
524
+ "microbatches": 4,
525
+ "elapsed_seconds": 48.086882687173784,
526
+ "batch_question_ids_sha256": "7af8e2cece43ae333d4687839255ee144e5968fc9aa063efa2ff6dd96c42d1b2"
527
+ },
528
+ {
529
+ "step": 56,
530
+ "phase": "full",
531
+ "loss": 0.13069539703428745,
532
+ "questions": 16,
533
+ "microbatches": 4,
534
+ "elapsed_seconds": 48.844940822571516,
535
+ "batch_question_ids_sha256": "6d579036d49f401569b6e29499a444725fe7a64de0907bc07bd389fd352e615b"
536
+ },
537
+ {
538
+ "step": 57,
539
+ "phase": "full",
540
+ "loss": -0.058048512786626816,
541
+ "questions": 16,
542
+ "microbatches": 4,
543
+ "elapsed_seconds": 49.59209310915321,
544
+ "batch_question_ids_sha256": "848946e51c61b43a984fc49047cb8c9f0fae9f28c6447d8b4ee2c74dafe198bc"
545
+ },
546
+ {
547
+ "step": 58,
548
+ "phase": "full",
549
+ "loss": 0.11841028090566397,
550
+ "questions": 16,
551
+ "microbatches": 4,
552
+ "elapsed_seconds": 50.3503486905247,
553
+ "batch_question_ids_sha256": "7a207cb4e3ff6157802e3b560ec0b63a488a8ab0eabaa5bcdea45eeeb107d3da"
554
+ },
555
+ {
556
+ "step": 59,
557
+ "phase": "full",
558
+ "loss": 0.07705274596810341,
559
+ "questions": 16,
560
+ "microbatches": 4,
561
+ "elapsed_seconds": 51.090083728544414,
562
+ "batch_question_ids_sha256": "6ed68210f517203266a035e4dc739a5e91ad1773fb37b2d954aabbc97ce44cc5"
563
+ },
564
+ {
565
+ "step": 60,
566
+ "phase": "full",
567
+ "loss": -0.01653895527124405,
568
+ "questions": 16,
569
+ "microbatches": 4,
570
+ "elapsed_seconds": 51.86966119892895,
571
+ "batch_question_ids_sha256": "cbce464330f46f2f323539fb9d7962cede14e7074be588012e1dfa684881e6d4"
572
+ },
573
+ {
574
+ "step": 61,
575
+ "phase": "full",
576
+ "loss": 0.09852208383381367,
577
+ "questions": 16,
578
+ "microbatches": 4,
579
+ "elapsed_seconds": 52.62419832125306,
580
+ "batch_question_ids_sha256": "7212e07d655e85ec9b791cd6cde65ec3e9ebfa5224d1854c76d9a8e98cc07be3"
581
+ },
582
+ {
583
+ "step": 62,
584
+ "phase": "full",
585
+ "loss": -0.06630048155784607,
586
+ "questions": 16,
587
+ "microbatches": 4,
588
+ "elapsed_seconds": 53.38278101198375,
589
+ "batch_question_ids_sha256": "768e8960dc65d1f0cc4b13cc2c90dbfe8db2c685cceb3a8ed4f834e59a13a2f5"
590
+ },
591
+ {
592
+ "step": 63,
593
+ "phase": "full",
594
+ "loss": 0.00335497222840786,
595
+ "questions": 16,
596
+ "microbatches": 4,
597
+ "elapsed_seconds": 54.137751570902765,
598
+ "batch_question_ids_sha256": "ce0e1bf7240822fc1f373178c20604e9dba0409aa63f2b51415a5f4db012ea78"
599
+ },
600
+ {
601
+ "step": 64,
602
+ "phase": "full",
603
+ "loss": -0.022465532645583153,
604
+ "questions": 16,
605
+ "microbatches": 4,
606
+ "elapsed_seconds": 54.904716577380896,
607
+ "batch_question_ids_sha256": "6b8c351645894c06829c8dffb3045f3a3081c0fcb0b24c924b9576410c78d47c"
608
+ },
609
+ {
610
+ "step": 65,
611
+ "phase": "full",
612
+ "loss": -0.06888989172875881,
613
+ "questions": 16,
614
+ "microbatches": 4,
615
+ "elapsed_seconds": 55.66192784719169,
616
+ "batch_question_ids_sha256": "06de39fb8b5de5ce866344dcb10447ed159378857789cc67daf23061f43e3e7f"
617
+ },
618
+ {
619
+ "step": 66,
620
+ "phase": "full",
621
+ "loss": -0.02314391452819109,
622
+ "questions": 16,
623
+ "microbatches": 4,
624
+ "elapsed_seconds": 56.42401862516999,
625
+ "batch_question_ids_sha256": "1728e00f1c3d6bfc533f2891740148ba51b40dcb766709fed2a21437282f6ee1"
626
+ },
627
+ {
628
+ "step": 67,
629
+ "phase": "full",
630
+ "loss": -0.08001549076288939,
631
+ "questions": 16,
632
+ "microbatches": 4,
633
+ "elapsed_seconds": 57.179139980114996,
634
+ "batch_question_ids_sha256": "26b40bee228ef411ff536ea5c58c9a2fe4790fcc250d1ec97719f5c2b6051983"
635
+ },
636
+ {
637
+ "step": 68,
638
+ "phase": "full",
639
+ "loss": 0.04526185989379883,
640
+ "questions": 16,
641
+ "microbatches": 4,
642
+ "elapsed_seconds": 57.95245848502964,
643
+ "batch_question_ids_sha256": "ebcd0c069c1aad6db15af178bfb86b95232c842c724cae991202c127dddf4df4"
644
+ },
645
+ {
646
+ "step": 69,
647
+ "phase": "full",
648
+ "loss": 0.08819740545004606,
649
+ "questions": 16,
650
+ "microbatches": 4,
651
+ "elapsed_seconds": 58.71377477888018,
652
+ "batch_question_ids_sha256": "021f795640c8ebe5b45f63378c3fb38ba2c849f5d70b747f653482f046de3fba"
653
+ },
654
+ {
655
+ "step": 70,
656
+ "phase": "full",
657
+ "loss": 0.15833008475601673,
658
+ "questions": 16,
659
+ "microbatches": 4,
660
+ "elapsed_seconds": 59.480925949290395,
661
+ "batch_question_ids_sha256": "968367da65a754ce0d7eb72abb8cbb64b558c86aa9d8adda10e119723caa2bb9"
662
+ },
663
+ {
664
+ "step": 71,
665
+ "phase": "full",
666
+ "loss": -0.04195588454604149,
667
+ "questions": 16,
668
+ "microbatches": 4,
669
+ "elapsed_seconds": 60.24882358778268,
670
+ "batch_question_ids_sha256": "5f2d81f647af85f4a48f30464c10a61c6a85b7d5d35f687d527ade870b0c1e7e"
671
+ },
672
+ {
673
+ "step": 72,
674
+ "phase": "full",
675
+ "loss": 0.22126729413866997,
676
+ "questions": 16,
677
+ "microbatches": 4,
678
+ "elapsed_seconds": 61.01563328132033,
679
+ "batch_question_ids_sha256": "6780ef3f330f6cd45a377b3f3c501ecc3ef2ae65d08769f207f4a68c152e3dbe"
680
+ },
681
+ {
682
+ "step": 73,
683
+ "phase": "full",
684
+ "loss": -0.02661465760320425,
685
+ "questions": 16,
686
+ "microbatches": 4,
687
+ "elapsed_seconds": 61.79356468375772,
688
+ "batch_question_ids_sha256": "d521c674f97a615228053a262b84518189f9580e2c73b32eee51bcef5ed76d29"
689
+ },
690
+ {
691
+ "step": 74,
692
+ "phase": "full",
693
+ "loss": -0.1296182470396161,
694
+ "questions": 16,
695
+ "microbatches": 4,
696
+ "elapsed_seconds": 62.70843842625618,
697
+ "batch_question_ids_sha256": "6c26d6bf906bf556525d65a2709f2f59a7cee21e2f62c8bad8c7015fcde12f89"
698
+ },
699
+ {
700
+ "step": 75,
701
+ "phase": "full",
702
+ "loss": -0.01979048363864422,
703
+ "questions": 16,
704
+ "microbatches": 4,
705
+ "elapsed_seconds": 63.46672537922859,
706
+ "batch_question_ids_sha256": "afabf643cddd0692428615c4aa81345acbbcdab1c52a5c9a48c815f9e2700755"
707
+ },
708
+ {
709
+ "step": 76,
710
+ "phase": "full",
711
+ "loss": 0.0012223836965858936,
712
+ "questions": 16,
713
+ "microbatches": 4,
714
+ "elapsed_seconds": 64.2400893131271,
715
+ "batch_question_ids_sha256": "ba8ce0da2f777532f12642355dcce1c4d29f8e0865010f26b4bf03c56d7723d1"
716
+ },
717
+ {
718
+ "step": 77,
719
+ "phase": "full",
720
+ "loss": 0.12047487869858742,
721
+ "questions": 16,
722
+ "microbatches": 4,
723
+ "elapsed_seconds": 65.01171296648681,
724
+ "batch_question_ids_sha256": "9e0a4687654abac2570a41f8674ab95fcade2ac8e5e238d4661499dff7d58834"
725
+ },
726
+ {
727
+ "step": 78,
728
+ "phase": "full",
729
+ "loss": 0.01269853487610817,
730
+ "questions": 16,
731
+ "microbatches": 4,
732
+ "elapsed_seconds": 65.77457793802023,
733
+ "batch_question_ids_sha256": "e36ae8ba0c5eaffb1a534d69421fb9118e4f45453e90b83cbbf5f29ef284cafc"
734
+ },
735
+ {
736
+ "step": 79,
737
+ "phase": "full",
738
+ "loss": 0.15956797171384096,
739
+ "questions": 16,
740
+ "microbatches": 4,
741
+ "elapsed_seconds": 66.52514539007097,
742
+ "batch_question_ids_sha256": "171d14c1b2acfe9895a9353fe3dc68dcde6a998182a601a2746bf8fb10de20e3"
743
+ },
744
+ {
745
+ "step": 80,
746
+ "phase": "full",
747
+ "loss": 0.021434137597680092,
748
+ "questions": 16,
749
+ "microbatches": 4,
750
+ "elapsed_seconds": 67.30584577657282,
751
+ "batch_question_ids_sha256": "5091eab21841dd9384edd2c719867375591caf9d084ba03c065878eb451a5674"
752
+ },
753
+ {
754
+ "step": 81,
755
+ "phase": "full",
756
+ "loss": 0.08233899343758821,
757
+ "questions": 16,
758
+ "microbatches": 4,
759
+ "elapsed_seconds": 68.06916774995625,
760
+ "batch_question_ids_sha256": "f94cae69521603c0b60eb1cc6bb5370d2ea3d089f5b6bfd9a260b41a766ff974"
761
+ },
762
+ {
763
+ "step": 82,
764
+ "phase": "full",
765
+ "loss": 0.06948617822490633,
766
+ "questions": 16,
767
+ "microbatches": 4,
768
+ "elapsed_seconds": 68.82234204933047,
769
+ "batch_question_ids_sha256": "f804d3278fd4930c24d09af7cc2e1fc70170c35cac45cbed9779d59f2d765e08"
770
+ },
771
+ {
772
+ "step": 83,
773
+ "phase": "full",
774
+ "loss": 0.005867251195013523,
775
+ "questions": 16,
776
+ "microbatches": 4,
777
+ "elapsed_seconds": 69.58695920649916,
778
+ "batch_question_ids_sha256": "19b60c4f93e1ba7368b1f066c22c4ba32b8b66981562b2cbe2485da7433c07aa"
779
+ },
780
+ {
781
+ "step": 84,
782
+ "phase": "full",
783
+ "loss": 0.08299335604533553,
784
+ "questions": 16,
785
+ "microbatches": 4,
786
+ "elapsed_seconds": 70.33971774112433,
787
+ "batch_question_ids_sha256": "ebdf0b28c90afc00f6691d1a09af18f9f66ebaf71377d0ba210fc4892869014c"
788
+ },
789
+ {
790
+ "step": 85,
791
+ "phase": "full",
792
+ "loss": 0.011595956981182098,
793
+ "questions": 16,
794
+ "microbatches": 4,
795
+ "elapsed_seconds": 71.10882101673633,
796
+ "batch_question_ids_sha256": "4f04bf635d321e106f9856ba99a47fd6c2c033ae0450aa32c6dfe595a74ae4ed"
797
+ },
798
+ {
799
+ "step": 86,
800
+ "phase": "full",
801
+ "loss": -0.03602656163275242,
802
+ "questions": 16,
803
+ "microbatches": 4,
804
+ "elapsed_seconds": 71.87254969682544,
805
+ "batch_question_ids_sha256": "ba9c22de6939f9c021221fe5a43ff5f50b15a70b21ed4effe8d97caf79b049ca"
806
+ },
807
+ {
808
+ "step": 87,
809
+ "phase": "full",
810
+ "loss": -0.013593184761703014,
811
+ "questions": 16,
812
+ "microbatches": 4,
813
+ "elapsed_seconds": 72.63154661748558,
814
+ "batch_question_ids_sha256": "c9f35e67c5374cdfd4ecee262a8723f0f688d0387f073072324c9d259d9f4323"
815
+ },
816
+ {
817
+ "step": 88,
818
+ "phase": "full",
819
+ "loss": 0.10619061812758446,
820
+ "questions": 16,
821
+ "microbatches": 4,
822
+ "elapsed_seconds": 73.39264491479844,
823
+ "batch_question_ids_sha256": "64ab5437c436a5c934bb05696c9b73f14c9838d3b210eeeb1114438303808400"
824
+ },
825
+ {
826
+ "step": 89,
827
+ "phase": "full",
828
+ "loss": -0.054122308269143105,
829
+ "questions": 16,
830
+ "microbatches": 4,
831
+ "elapsed_seconds": 74.15655770525336,
832
+ "batch_question_ids_sha256": "6ff06537c7c8217ebac7a64762ff5ee340f367a0f7652d25d5cd799c0d15cbac"
833
+ },
834
+ {
835
+ "step": 90,
836
+ "phase": "full",
837
+ "loss": 0.011591964401304722,
838
+ "questions": 16,
839
+ "microbatches": 4,
840
+ "elapsed_seconds": 74.92870382964611,
841
+ "batch_question_ids_sha256": "8fd09d36031bf8ee878da3b38a0694c5e416636fcd74e603f03dd205a9a3ab33"
842
+ },
843
+ {
844
+ "step": 91,
845
+ "phase": "full",
846
+ "loss": -0.060869204346090555,
847
+ "questions": 16,
848
+ "microbatches": 4,
849
+ "elapsed_seconds": 75.69129009637982,
850
+ "batch_question_ids_sha256": "322fb22dffd39fb986e5bfd20ac78a9963d2d5b01978598e3ea9ff285c9b09d3"
851
+ },
852
+ {
853
+ "step": 92,
854
+ "phase": "full",
855
+ "loss": -0.02303236722946167,
856
+ "questions": 16,
857
+ "microbatches": 4,
858
+ "elapsed_seconds": 76.43906010035425,
859
+ "batch_question_ids_sha256": "d0417dd44fb2608169bd7f2ab17d57acd203be7f8242f3c2e89ad317009a2e56"
860
+ },
861
+ {
862
+ "step": 93,
863
+ "phase": "full",
864
+ "loss": 0.017207864671945572,
865
+ "questions": 16,
866
+ "microbatches": 4,
867
+ "elapsed_seconds": 77.20492906402797,
868
+ "batch_question_ids_sha256": "57be67994833e0a92fac3ba6c7b6b1fe3b8383d53284136aad0b9a926f82aa1f"
869
+ },
870
+ {
871
+ "step": 94,
872
+ "phase": "full",
873
+ "loss": -0.04622429329901934,
874
+ "questions": 16,
875
+ "microbatches": 4,
876
+ "elapsed_seconds": 77.99955605249852,
877
+ "batch_question_ids_sha256": "9c3cbfa9806ac2c618d0a1e3ca3ef41c91b3a86c8aa45a36071383124ee71262"
878
+ },
879
+ {
880
+ "step": 95,
881
+ "phase": "full",
882
+ "loss": -0.1110177468508482,
883
+ "questions": 16,
884
+ "microbatches": 4,
885
+ "elapsed_seconds": 78.75541434716433,
886
+ "batch_question_ids_sha256": "100fb8785fb8c78e98988f555871d3d9a4ed2716c9bff93597e8a64de5e2db46"
887
+ },
888
+ {
889
+ "step": 96,
890
+ "phase": "full",
891
+ "loss": -0.015255212783813477,
892
+ "questions": 16,
893
+ "microbatches": 4,
894
+ "elapsed_seconds": 79.51458817347884,
895
+ "batch_question_ids_sha256": "c781c81a212d6bc7153d4ecdfb49ff5f6570c3bde61d9f7babaa983e433d7348"
896
+ },
897
+ {
898
+ "step": 97,
899
+ "phase": "full",
900
+ "loss": -0.06440366432070732,
901
+ "questions": 16,
902
+ "microbatches": 4,
903
+ "elapsed_seconds": 80.28054185304791,
904
+ "batch_question_ids_sha256": "8b19c5c7751e07c22469266bfa40bb1eaac4670fe203a8c3e8a289350956893c"
905
+ },
906
+ {
907
+ "step": 98,
908
+ "phase": "full",
909
+ "loss": -0.042902698274701834,
910
+ "questions": 16,
911
+ "microbatches": 4,
912
+ "elapsed_seconds": 81.04737775586545,
913
+ "batch_question_ids_sha256": "043573f3a457c1f8cfbcefabfef11e430a1dd573e779dd7e19f7dc26a4a1ebad"
914
+ },
915
+ {
916
+ "step": 99,
917
+ "phase": "full",
918
+ "loss": -0.03668666258454323,
919
+ "questions": 16,
920
+ "microbatches": 4,
921
+ "elapsed_seconds": 81.81181224621832,
922
+ "batch_question_ids_sha256": "933b002848b57b1ae58768d4d38da9823ca19226389dff7c41f6ba3e58ff14b6"
923
+ },
924
+ {
925
+ "step": 100,
926
+ "phase": "full",
927
+ "loss": -0.017029337119311094,
928
+ "questions": 16,
929
+ "microbatches": 4,
930
+ "elapsed_seconds": 82.66445402894169,
931
+ "batch_question_ids_sha256": "2a2b48544fe4380e81979729c21f00410faff377b407c4afdee1bef7530a71af",
932
+ "dev": {
933
+ "questions": 372,
934
+ "objective": "observed_outcome",
935
+ "eligible_questions": 372,
936
+ "target_ce": 0.45392147839038,
937
+ "target_kl": 0.45392147839038,
938
+ "target_tv": 0.2908940469690758,
939
+ "by_target_kind": {
940
+ "observed_outcome": {
941
+ "n": 372,
942
+ "ce": 0.45392147839038,
943
+ "kl": 0.45392147839038,
944
+ "tv": 0.2908940469690758
945
+ }
946
+ },
947
+ "independent_targets": {
948
+ "teacher": {
949
+ "n": 0,
950
+ "ce": 0.0,
951
+ "kl": 0.0,
952
+ "tv": 0.0
953
+ },
954
+ "gold_distribution": {
955
+ "n": 372,
956
+ "ce": 0.45846966947460877,
957
+ "kl": 0.18846707686707742,
958
+ "tv": 0.18377175030687837
959
+ }
960
+ },
961
+ "note": "question-mean distribution recovery; policy targets and reference argmax labels are not observed outcomes"
962
+ }
963
+ }
964
+ ]
variants/games_api_seed17/backbone_config/config.json ADDED
@@ -0,0 +1,63 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ {
2
+ "architectures": [
3
+ "Qwen3ForCausalLM"
4
+ ],
5
+ "attention_bias": false,
6
+ "attention_dropout": 0.0,
7
+ "bos_token_id": 151643,
8
+ "dtype": "float32",
9
+ "eos_token_id": 151645,
10
+ "head_dim": 128,
11
+ "hidden_act": "silu",
12
+ "hidden_size": 1024,
13
+ "initializer_range": 0.02,
14
+ "intermediate_size": 3072,
15
+ "layer_types": [
16
+ "full_attention",
17
+ "full_attention",
18
+ "full_attention",
19
+ "full_attention",
20
+ "full_attention",
21
+ "full_attention",
22
+ "full_attention",
23
+ "full_attention",
24
+ "full_attention",
25
+ "full_attention",
26
+ "full_attention",
27
+ "full_attention",
28
+ "full_attention",
29
+ "full_attention",
30
+ "full_attention",
31
+ "full_attention",
32
+ "full_attention",
33
+ "full_attention",
34
+ "full_attention",
35
+ "full_attention",
36
+ "full_attention",
37
+ "full_attention",
38
+ "full_attention",
39
+ "full_attention",
40
+ "full_attention",
41
+ "full_attention",
42
+ "full_attention",
43
+ "full_attention"
44
+ ],
45
+ "max_position_embeddings": 40960,
46
+ "max_window_layers": 28,
47
+ "model_type": "qwen3",
48
+ "num_attention_heads": 16,
49
+ "num_hidden_layers": 28,
50
+ "num_key_value_heads": 8,
51
+ "pad_token_id": null,
52
+ "rms_norm_eps": 1e-06,
53
+ "rope_parameters": {
54
+ "rope_theta": 1000000,
55
+ "rope_type": "default"
56
+ },
57
+ "sliding_window": null,
58
+ "tie_word_embeddings": true,
59
+ "transformers_version": "5.17.0",
60
+ "use_cache": false,
61
+ "use_sliding_window": false,
62
+ "vocab_size": 151936
63
+ }
variants/games_api_seed17/best.safetensors ADDED
@@ -0,0 +1,3 @@
 
 
 
 
1
+ version https://git-lfs.github.com/spec/v1
2
+ oid sha256:45829fc896391847220b67dcf6fc0702427d2f184be384649bcf9343c1552477
3
+ size 2385039280
variants/games_api_seed17/config.json ADDED
@@ -0,0 +1,74 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ {
2
+ "input": "data/scaled_games_labeled_v4",
3
+ "output_dir": "runs/games_api_seed17",
4
+ "model": "Qwen/Qwen3-0.6B",
5
+ "revision": "main",
6
+ "init_checkpoint": "/home/rwang/openjev_codex_20260917/runs/v3_teacher_coords_multi_seed17",
7
+ "objective": "teacher",
8
+ "loss": "ce",
9
+ "reward_samples": 32,
10
+ "gradient_checkpointing": true,
11
+ "set_head": "attention",
12
+ "steps": 150,
13
+ "head_steps": 0,
14
+ "batch_questions": 16,
15
+ "microbatch_questions": 4,
16
+ "max_microbatch_tokens": 16384,
17
+ "eval_every": 50,
18
+ "max_length": 8192,
19
+ "seed": 17,
20
+ "backbone_lr": 2e-05,
21
+ "head_lr": 0.0002,
22
+ "head_warmup_lr": 0.001,
23
+ "precision": "bf16",
24
+ "disable_native_triton": true,
25
+ "validate_only": false,
26
+ "self_check": false,
27
+ "schema_version": "openjev-decision-pipeline-v1",
28
+ "resolved_model_revision": "c1899de289a04d12100db370d81485cdf75e47ca",
29
+ "initialization": "local DecisionModel warm start, fresh optimizer",
30
+ "data_sha256": {
31
+ "data/scaled_games_labeled_v4/train.jsonl": "e21d05f1c34253687f4e7b78974503e7c05a7e145e4cbfc306246938d9f6317c",
32
+ "data/scaled_games_labeled_v4/dev.jsonl": "fd7780fd6bf94263f5653cea8b3575925447981602dba61d0c7ff96b8433ea88",
33
+ "data/scaled_games_labeled_v4/calibration.jsonl": "9632b050ee1069c4393e275fc878c2681f9c36a501c4c6584df19b3fb88d76fd",
34
+ "data/scaled_games_labeled_v4/test.jsonl": "82b62b39f56bcd703e6e7c911aafb20203ff7b48ebf4f4f4c3444694e4823fa0",
35
+ "data/scaled_games_labeled_v4/ood.jsonl": "568c11c100126b3848fb29c9cd3d4b8dc22be42fe2df43a3bb3cf8e7b0128ae1"
36
+ },
37
+ "deps": {
38
+ "torch": "2.14.0",
39
+ "transformers": "5.17.0",
40
+ "safetensors": "0.8.0"
41
+ },
42
+ "gpu": "NVIDIA A100-SXM4-80GB",
43
+ "parameter_storage": "float32",
44
+ "forward_autocast": "bfloat16",
45
+ "parameter_count": 596250498,
46
+ "train_questions": 1520,
47
+ "all_train_questions": 1531,
48
+ "schema_counts": {
49
+ "records": 589,
50
+ "questions_by_split": {
51
+ "train": 1531,
52
+ "dev": 504,
53
+ "calibration": 500,
54
+ "test": 487,
55
+ "ood": 173
56
+ },
57
+ "eligible_by_split_objective": {
58
+ "calibration/gold_distribution": 500,
59
+ "calibration/teacher": 500,
60
+ "dev/gold_distribution": 504,
61
+ "dev/teacher": 501,
62
+ "ood/gold_distribution": 173,
63
+ "ood/teacher": 172,
64
+ "test/gold_distribution": 487,
65
+ "test/teacher": 487,
66
+ "train/gold_distribution": 1531,
67
+ "train/teacher": 1520
68
+ }
69
+ },
70
+ "loss_description": "per-question proper loss; complete candidate normalization",
71
+ "rlcd_candidate": null,
72
+ "selection": "minimum dev target CE; held-out test first evaluated after training and checkpoint selection",
73
+ "parallelism": "complete candidate paths batched in one backbone call per microbatch; no prefix sharing"
74
+ }
variants/games_api_seed17/summary.json ADDED
@@ -0,0 +1,136 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ {
2
+ "best_step": 150,
3
+ "best_dev_target_ce": 0.7130707709730839,
4
+ "selected_on": "dev target CE",
5
+ "objective": "teacher",
6
+ "metrics_by_split": {
7
+ "dev": {
8
+ "questions": 504,
9
+ "objective": "teacher",
10
+ "eligible_questions": 501,
11
+ "target_ce": 0.7130707709730839,
12
+ "target_kl": 0.10754575952110457,
13
+ "target_tv": 0.1622813057690467,
14
+ "by_target_kind": {
15
+ "teacher_rounded_proxy": {
16
+ "n": 501,
17
+ "ce": 0.7130707709730839,
18
+ "kl": 0.10754575952110457,
19
+ "tv": 0.1622813057690467
20
+ }
21
+ },
22
+ "independent_targets": {
23
+ "teacher": {
24
+ "n": 501,
25
+ "ce": 0.7130707709730839,
26
+ "kl": 0.10754575952110457,
27
+ "tv": 0.1622813057690467
28
+ },
29
+ "gold_distribution": {
30
+ "n": 504,
31
+ "ce": 0.701963093111963,
32
+ "kl": 0.6602369393657108,
33
+ "tv": 0.4357824048845307
34
+ }
35
+ },
36
+ "note": "question-mean distribution recovery; policy targets and reference argmax labels are not observed outcomes"
37
+ },
38
+ "calibration": {
39
+ "questions": 500,
40
+ "objective": "teacher",
41
+ "eligible_questions": 500,
42
+ "target_ce": 0.7210497784051056,
43
+ "target_kl": 0.11017916320286496,
44
+ "target_tv": 0.1606939957216772,
45
+ "by_target_kind": {
46
+ "teacher_rounded_proxy": {
47
+ "n": 500,
48
+ "ce": 0.7210497784051056,
49
+ "kl": 0.11017916320286496,
50
+ "tv": 0.1606939957216772
51
+ }
52
+ },
53
+ "independent_targets": {
54
+ "teacher": {
55
+ "n": 500,
56
+ "ce": 0.7210497784051056,
57
+ "kl": 0.11017916320286496,
58
+ "tv": 0.1606939957216772
59
+ },
60
+ "gold_distribution": {
61
+ "n": 500,
62
+ "ce": 0.690937608859615,
63
+ "kl": 0.6533763270403429,
64
+ "tv": 0.43010127156540956
65
+ }
66
+ },
67
+ "note": "question-mean distribution recovery; policy targets and reference argmax labels are not observed outcomes"
68
+ },
69
+ "test": {
70
+ "questions": 487,
71
+ "objective": "teacher",
72
+ "eligible_questions": 487,
73
+ "target_ce": 0.7178738686092758,
74
+ "target_kl": 0.11519962202078171,
75
+ "target_tv": 0.16197697068307357,
76
+ "by_target_kind": {
77
+ "teacher_rounded_proxy": {
78
+ "n": 487,
79
+ "ce": 0.7178738686092758,
80
+ "kl": 0.11519962202078171,
81
+ "tv": 0.16197697068307357
82
+ }
83
+ },
84
+ "independent_targets": {
85
+ "teacher": {
86
+ "n": 487,
87
+ "ce": 0.7178738686092758,
88
+ "kl": 0.11519962202078171,
89
+ "tv": 0.16197697068307357
90
+ },
91
+ "gold_distribution": {
92
+ "n": 487,
93
+ "ce": 0.6580319982288982,
94
+ "kl": 0.6173470163730734,
95
+ "tv": 0.4194799845862883
96
+ }
97
+ },
98
+ "note": "question-mean distribution recovery; policy targets and reference argmax labels are not observed outcomes"
99
+ },
100
+ "ood": {
101
+ "questions": 173,
102
+ "objective": "teacher",
103
+ "eligible_questions": 172,
104
+ "target_ce": 0.7490552275394107,
105
+ "target_kl": 0.10398421448117022,
106
+ "target_tv": 0.14334074401012564,
107
+ "by_target_kind": {
108
+ "teacher_rounded_proxy": {
109
+ "n": 172,
110
+ "ce": 0.7490552275394107,
111
+ "kl": 0.10398421448117022,
112
+ "tv": 0.14334074401012564
113
+ }
114
+ },
115
+ "independent_targets": {
116
+ "teacher": {
117
+ "n": 172,
118
+ "ce": 0.7490552275394107,
119
+ "kl": 0.10398421448117022,
120
+ "tv": 0.14334074401012564
121
+ },
122
+ "gold_distribution": {
123
+ "n": 173,
124
+ "ce": 0.7033095694567321,
125
+ "kl": 0.660899529493104,
126
+ "tv": 0.43063586291702377
127
+ }
128
+ },
129
+ "note": "question-mean distribution recovery; policy targets and reference argmax labels are not observed outcomes"
130
+ }
131
+ },
132
+ "temperature": 1.0,
133
+ "temperature_fitted": false,
134
+ "training_seconds": 182.26331992074847,
135
+ "max_gpu_allocated_gb": 12.375407104
136
+ }
variants/games_api_seed17/target_audit.json ADDED
The diff for this file is too large to render. See raw diff
 
variants/games_api_seed17/tokenizer/chat_template.jinja ADDED
@@ -0,0 +1,89 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ {%- if tools %}
2
+ {{- '<|im_start|>system\n' }}
3
+ {%- if messages[0].role == 'system' %}
4
+ {{- messages[0].content + '\n\n' }}
5
+ {%- endif %}
6
+ {{- "# Tools\n\nYou may call one or more functions to assist with the user query.\n\nYou are provided with function signatures within <tools></tools> XML tags:\n<tools>" }}
7
+ {%- for tool in tools %}
8
+ {{- "\n" }}
9
+ {{- tool | tojson }}
10
+ {%- endfor %}
11
+ {{- "\n</tools>\n\nFor each function call, return a json object with function name and arguments within <tool_call></tool_call> XML tags:\n<tool_call>\n{\"name\": <function-name>, \"arguments\": <args-json-object>}\n</tool_call><|im_end|>\n" }}
12
+ {%- else %}
13
+ {%- if messages[0].role == 'system' %}
14
+ {{- '<|im_start|>system\n' + messages[0].content + '<|im_end|>\n' }}
15
+ {%- endif %}
16
+ {%- endif %}
17
+ {%- set ns = namespace(multi_step_tool=true, last_query_index=messages|length - 1) %}
18
+ {%- for message in messages[::-1] %}
19
+ {%- set index = (messages|length - 1) - loop.index0 %}
20
+ {%- if ns.multi_step_tool and message.role == "user" and message.content is string and not(message.content.startswith('<tool_response>') and message.content.endswith('</tool_response>')) %}
21
+ {%- set ns.multi_step_tool = false %}
22
+ {%- set ns.last_query_index = index %}
23
+ {%- endif %}
24
+ {%- endfor %}
25
+ {%- for message in messages %}
26
+ {%- if message.content is string %}
27
+ {%- set content = message.content %}
28
+ {%- else %}
29
+ {%- set content = '' %}
30
+ {%- endif %}
31
+ {%- if (message.role == "user") or (message.role == "system" and not loop.first) %}
32
+ {{- '<|im_start|>' + message.role + '\n' + content + '<|im_end|>' + '\n' }}
33
+ {%- elif message.role == "assistant" %}
34
+ {%- set reasoning_content = '' %}
35
+ {%- if message.reasoning_content is string %}
36
+ {%- set reasoning_content = message.reasoning_content %}
37
+ {%- else %}
38
+ {%- if '</think>' in content %}
39
+ {%- set reasoning_content = content.split('</think>')[0].rstrip('\n').split('<think>')[-1].lstrip('\n') %}
40
+ {%- set content = content.split('</think>')[-1].lstrip('\n') %}
41
+ {%- endif %}
42
+ {%- endif %}
43
+ {%- if loop.index0 > ns.last_query_index %}
44
+ {%- if loop.last or (not loop.last and reasoning_content) %}
45
+ {{- '<|im_start|>' + message.role + '\n<think>\n' + reasoning_content.strip('\n') + '\n</think>\n\n' + content.lstrip('\n') }}
46
+ {%- else %}
47
+ {{- '<|im_start|>' + message.role + '\n' + content }}
48
+ {%- endif %}
49
+ {%- else %}
50
+ {{- '<|im_start|>' + message.role + '\n' + content }}
51
+ {%- endif %}
52
+ {%- if message.tool_calls %}
53
+ {%- for tool_call in message.tool_calls %}
54
+ {%- if (loop.first and content) or (not loop.first) %}
55
+ {{- '\n' }}
56
+ {%- endif %}
57
+ {%- if tool_call.function %}
58
+ {%- set tool_call = tool_call.function %}
59
+ {%- endif %}
60
+ {{- '<tool_call>\n{"name": "' }}
61
+ {{- tool_call.name }}
62
+ {{- '", "arguments": ' }}
63
+ {%- if tool_call.arguments is string %}
64
+ {{- tool_call.arguments }}
65
+ {%- else %}
66
+ {{- tool_call.arguments | tojson }}
67
+ {%- endif %}
68
+ {{- '}\n</tool_call>' }}
69
+ {%- endfor %}
70
+ {%- endif %}
71
+ {{- '<|im_end|>\n' }}
72
+ {%- elif message.role == "tool" %}
73
+ {%- if loop.first or (messages[loop.index0 - 1].role != "tool") %}
74
+ {{- '<|im_start|>user' }}
75
+ {%- endif %}
76
+ {{- '\n<tool_response>\n' }}
77
+ {{- content }}
78
+ {{- '\n</tool_response>' }}
79
+ {%- if loop.last or (messages[loop.index0 + 1].role != "tool") %}
80
+ {{- '<|im_end|>\n' }}
81
+ {%- endif %}
82
+ {%- endif %}
83
+ {%- endfor %}
84
+ {%- if add_generation_prompt %}
85
+ {{- '<|im_start|>assistant\n' }}
86
+ {%- if enable_thinking is defined and enable_thinking is false %}
87
+ {{- '<think>\n\n</think>\n\n' }}
88
+ {%- endif %}
89
+ {%- endif %}
variants/games_api_seed17/tokenizer/tokenizer.json ADDED
@@ -0,0 +1,3 @@
 
 
 
 
1
+ version https://git-lfs.github.com/spec/v1
2
+ oid sha256:be75606093db2094d7cd20f3c2f385c212750648bd6ea4fb2bf507a6a4c55506
3
+ size 11422650