Caesarrr commited on
Commit
ea2ad5a
·
verified ·
1 Parent(s): 6d11a6d

Publish complete Atlantis Asterix H8 1800-episode comparison

Browse files
reports/atlantis-asterix-h8-v1/README.md ADDED
@@ -0,0 +1,22 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ # Atlantis / Asterix H8 final results
2
+
3
+ All training, evaluation and model/profile publications completed. 18/18 formal VLA evaluations, 1800/1800 episodes; six additional teacher evaluations total 600 episodes.
4
+
5
+ VLA protocol: final optimizer step 5000, episode seeds 1000000–1000099, GPU simulator, parallel 32, capacity 1, 60/15 FPS/Hz, maximum 7200 raw frames. Both profile arms use the same fixed model-specific RTX 3090 profile, measured with CPU simulation/rendering and one visible inference GPU.
6
+
7
+ | Game / model | zero→zero | zero→profile | profile→profile |
8
+ |---|---:|---:|---:|
9
+ | atlantis / qwenoft | 39560.0 | 18588.0 | 39229.0 |
10
+ | atlantis / qwengr00t | 39662.0 | 8167.0 | 28488.0 |
11
+ | atlantis / qwenpi_v3 | 39615.0 | 4369.0 | 26447.0 |
12
+ | asterix / qwenoft | 5803.5 | 374.5 | 3653.5 |
13
+ | asterix / qwengr00t | 5714.0 | 182.0 | 1440.0 |
14
+ | asterix / qwenpi_v3 | 5758.5 | 143.5 | 1015.0 |
15
+
16
+ Profile-aware training improves mean return under measured latency in all six comparisons. Atlantis OFT approaches its zero-latency baseline; the other five remain below their zero→zero baselines. Each condition uses one training seed; 100 shared evaluation seeds do not measure training-seed variability. Profiles differ by architecture, so compare matching-model arms.
17
+
18
+ Training: six zero VLAs and six profile VLAs, each 5000 updates; six profile-aware teachers, each 15M CPU/ALE steps, seed 0, constant LR. All data passed binary H8/reset/causality audits. Atlantis uses 4 actions and 56D history; Asterix uses native 9 actions and 136D history.
19
+
20
+ Publication: all twelve VLA bundles and six teachers are readback-verified in `latency-sensitive-bench/benchmark-models`; six CPU-simulator/single-GPU RTX 3090 profiles are verified in `latency-sensitive-bench/profiles`. VLA W&B: `saberrr-zju/starvla_tasks`; teacher W&B: `saberrr-zju/latency-sensitive-bench`.
21
+
22
+ [Evaluation identities and audits](verified_evaluations.json) · [Model inventory](model_inventory.json) · [Teacher verification](teachers.complete.json) · [Profile verification](completed_profiles.json)
reports/atlantis-asterix-h8-v1/completed_profiles.json ADDED
@@ -0,0 +1,343 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ {
2
+ "checked_at": 1790259638.9276395,
3
+ "profiles": [
4
+ {
5
+ "model": "qwengr00t",
6
+ "model_publication": {
7
+ "checkpoint_sha256": "7b3541140bd2d00d39eca98c8788b773e594884a623ae20d7b9e4dd7b45337f5",
8
+ "episodes": 100,
9
+ "mean_return": 5714.0,
10
+ "model": "qwengr00t",
11
+ "path": "zero-latency/asterix/vla/starvla-qwengr00t-h1/asterix-qwengr00t-hist8-zero-s42-v1",
12
+ "repo_id": "latency-sensitive-bench/benchmark-models",
13
+ "revision": "d445334970ab7e110c3cacb5c8b790d3df20e1ec",
14
+ "run_id": "asterix-qwengr00t-hist8-zero-s42-v1",
15
+ "task": "asterix",
16
+ "training_status": "complete",
17
+ "zero_evaluation_status": "complete",
18
+ "zero_mean_return": 5714.0
19
+ },
20
+ "profile": {
21
+ "GPU_CLASS": "1x-rtx3090",
22
+ "INSTANCE_ID": "instance_d907cdaf680040e0",
23
+ "PROFILE_PATH": "results/profiling/profiles/qwengr00t/1x-rtx3090/asterix/instance_d907cdaf680040e0/profile.json",
24
+ "PROFILE_REF": "qwengr00t/1x-rtx3090/asterix/instance_d907cdaf680040e0",
25
+ "RUN_ID": "20260924T122426657466Z"
26
+ },
27
+ "profile_path": "profiles/qwengr00t/1x-rtx3090/asterix/instance_d907cdaf680040e0/profile.json",
28
+ "profile_repo": "latency-sensitive-bench/profiles",
29
+ "profile_revision": "15001405a402e30b7706476bc92aea45a61bb193",
30
+ "profile_sha256": "5e104064d718ac51ce56aefac5318a155fd099f5c714ff8f0bc4985509639fd1",
31
+ "task": "asterix",
32
+ "verified_sessions": [
33
+ {
34
+ "session": 0,
35
+ "episodes": 10
36
+ },
37
+ {
38
+ "session": 1,
39
+ "episodes": 10
40
+ },
41
+ {
42
+ "session": 2,
43
+ "episodes": 10
44
+ },
45
+ {
46
+ "session": 3,
47
+ "episodes": 10
48
+ },
49
+ {
50
+ "session": 4,
51
+ "episodes": 10
52
+ }
53
+ ],
54
+ "total_episodes": 50,
55
+ "mean_latency_ms": 216.59862126426137,
56
+ "p95_latency_ms": 228.23943430036303,
57
+ "simulator": "cpu",
58
+ "visible_inference_gpus": 1,
59
+ "status": "verified"
60
+ },
61
+ {
62
+ "model": "qwenoft",
63
+ "model_publication": {
64
+ "checkpoint_sha256": "9d5be1187e768a3ba6638610ad70ca58fed54a531b77b0b0dea3a33538def572",
65
+ "episodes": 100,
66
+ "mean_return": 5803.5,
67
+ "model": "qwenoft",
68
+ "path": "zero-latency/asterix/vla/starvla-qwenoft-h1/asterix-qwenoft-hist8-zero-s42-v1",
69
+ "repo_id": "latency-sensitive-bench/benchmark-models",
70
+ "revision": "15c9cd4ab12215ccbcef74a8cd4d051d8e3aae3f",
71
+ "run_id": "asterix-qwenoft-hist8-zero-s42-v1",
72
+ "task": "asterix",
73
+ "training_status": "complete",
74
+ "zero_evaluation_status": "complete"
75
+ },
76
+ "profile": {
77
+ "GPU_CLASS": "1x-rtx3090",
78
+ "INSTANCE_ID": "instance_d907cdaf680040e0",
79
+ "PROFILE_PATH": "results/profiling/profiles/qwenoft/1x-rtx3090/asterix/instance_d907cdaf680040e0/profile.json",
80
+ "PROFILE_REF": "qwenoft/1x-rtx3090/asterix/instance_d907cdaf680040e0",
81
+ "RUN_ID": "20260924T110210301668Z"
82
+ },
83
+ "profile_path": "profiles/qwenoft/1x-rtx3090/asterix/instance_d907cdaf680040e0/profile.json",
84
+ "profile_repo": "latency-sensitive-bench/profiles",
85
+ "profile_revision": "fbbba76b21fa705cf4e5d05ec1ecfb82a88b12d8",
86
+ "profile_sha256": "4e2e3ed923c7500248ce4ce1508f124faec78282d90d6b99d070a9e29bdb6146",
87
+ "task": "asterix",
88
+ "verified_sessions": [
89
+ {
90
+ "session": 0,
91
+ "episodes": 10
92
+ },
93
+ {
94
+ "session": 1,
95
+ "episodes": 10
96
+ },
97
+ {
98
+ "session": 2,
99
+ "episodes": 10
100
+ },
101
+ {
102
+ "session": 3,
103
+ "episodes": 10
104
+ },
105
+ {
106
+ "session": 4,
107
+ "episodes": 10
108
+ }
109
+ ],
110
+ "total_episodes": 50,
111
+ "mean_latency_ms": 114.93909660206005,
112
+ "p95_latency_ms": 120.04566190615296,
113
+ "simulator": "cpu",
114
+ "visible_inference_gpus": 1,
115
+ "status": "verified"
116
+ },
117
+ {
118
+ "model": "qwenpi_v3",
119
+ "model_publication": {
120
+ "checkpoint_sha256": "381d1ed73a6f3949cb1880d6f1feb18db13068aeb6db39a7974ae9fb1fba0f56",
121
+ "episodes": 100,
122
+ "mean_return": 5758.5,
123
+ "model": "qwenpi_v3",
124
+ "path": "zero-latency/asterix/vla/starvla-qwenpi_v3-h1/asterix-qwenpi_v3-hist8-zero-s42-v1",
125
+ "repo_id": "latency-sensitive-bench/benchmark-models",
126
+ "revision": "2c66bf8b40bf9dc6fd9b9ea26ccd9fb082f3b883",
127
+ "run_id": "asterix-qwenpi_v3-hist8-zero-s42-v1",
128
+ "task": "asterix",
129
+ "training_status": "complete",
130
+ "zero_evaluation_status": "complete",
131
+ "zero_mean_return": 5758.5
132
+ },
133
+ "profile": {
134
+ "GPU_CLASS": "1x-rtx3090",
135
+ "INSTANCE_ID": "instance_d907cdaf680040e0",
136
+ "PROFILE_PATH": "results/profiling/profiles/qwenpi_v3/1x-rtx3090/asterix/instance_d907cdaf680040e0/profile.json",
137
+ "PROFILE_REF": "qwenpi_v3/1x-rtx3090/asterix/instance_d907cdaf680040e0",
138
+ "RUN_ID": "20260924T124905574960Z"
139
+ },
140
+ "profile_path": "profiles/qwenpi_v3/1x-rtx3090/asterix/instance_d907cdaf680040e0/profile.json",
141
+ "profile_repo": "latency-sensitive-bench/profiles",
142
+ "profile_revision": "36f784f257aabf73721d5edd63abc7b3dee1a272",
143
+ "profile_sha256": "a5acfbb259cc3431bcce9dedef862a1e3895bf81659aa3c1707fa54c277852fc",
144
+ "task": "asterix",
145
+ "verified_sessions": [
146
+ {
147
+ "session": 0,
148
+ "episodes": 10
149
+ },
150
+ {
151
+ "session": 1,
152
+ "episodes": 10
153
+ },
154
+ {
155
+ "session": 2,
156
+ "episodes": 10
157
+ },
158
+ {
159
+ "session": 3,
160
+ "episodes": 10
161
+ },
162
+ {
163
+ "session": 4,
164
+ "episodes": 10
165
+ }
166
+ ],
167
+ "total_episodes": 50,
168
+ "mean_latency_ms": 319.3320188773304,
169
+ "p95_latency_ms": 338.4848217498511,
170
+ "simulator": "cpu",
171
+ "visible_inference_gpus": 1,
172
+ "status": "verified"
173
+ },
174
+ {
175
+ "model": "qwengr00t",
176
+ "model_publication": {
177
+ "checkpoint_sha256": "dc7081c898fd8324d5c9aa892b5221ad7388a83428cc3508c67b30fc71d0761a",
178
+ "episodes": 100,
179
+ "mean_return": 39662.0,
180
+ "model": "qwengr00t",
181
+ "path": "zero-latency/atlantis/vla/starvla-qwengr00t-h1/atlantis-qwengr00t-hist8-zero-s42-v1",
182
+ "repo_id": "latency-sensitive-bench/benchmark-models",
183
+ "revision": "aaf650c36c93cb8ff87d3a5ff6ffd635d24c9004",
184
+ "run_id": "atlantis-qwengr00t-hist8-zero-s42-v1",
185
+ "task": "atlantis",
186
+ "training_status": "complete",
187
+ "zero_evaluation_status": "complete"
188
+ },
189
+ "profile": {
190
+ "GPU_CLASS": "1x-rtx3090",
191
+ "INSTANCE_ID": "instance_d907cdaf680040e0",
192
+ "PROFILE_PATH": "results/profiling/profiles/qwengr00t/1x-rtx3090/atlantis/instance_d907cdaf680040e0/profile.json",
193
+ "PROFILE_REF": "qwengr00t/1x-rtx3090/atlantis/instance_d907cdaf680040e0",
194
+ "RUN_ID": "20260924T112532341532Z"
195
+ },
196
+ "profile_path": "profiles/qwengr00t/1x-rtx3090/atlantis/instance_d907cdaf680040e0/profile.json",
197
+ "profile_repo": "latency-sensitive-bench/profiles",
198
+ "profile_revision": "a00f633aa3227e2be8648b4f1337d7ccd188d18f",
199
+ "profile_sha256": "13e970f85825c898ecb5730ff6adb6c71995a68406227733311a15b3f0c202e9",
200
+ "task": "atlantis",
201
+ "verified_sessions": [
202
+ {
203
+ "session": 0,
204
+ "episodes": 10
205
+ },
206
+ {
207
+ "session": 1,
208
+ "episodes": 10
209
+ },
210
+ {
211
+ "session": 2,
212
+ "episodes": 10
213
+ },
214
+ {
215
+ "session": 3,
216
+ "episodes": 10
217
+ },
218
+ {
219
+ "session": 4,
220
+ "episodes": 10
221
+ }
222
+ ],
223
+ "total_episodes": 50,
224
+ "mean_latency_ms": 212.99372203647695,
225
+ "p95_latency_ms": 221.8632535457611,
226
+ "simulator": "cpu",
227
+ "visible_inference_gpus": 1,
228
+ "status": "verified"
229
+ },
230
+ {
231
+ "model": "qwenoft",
232
+ "model_publication": {
233
+ "checkpoint_sha256": "f91d4079a94f779856220e4a954750624228a3da596b5ecdd70861e888a97877",
234
+ "episodes": 100,
235
+ "mean_return": 39560.0,
236
+ "model": "qwenoft",
237
+ "path": "zero-latency/atlantis/vla/starvla-qwenoft-h1/atlantis-qwenoft-hist8-zero-s42-v1",
238
+ "repo_id": "latency-sensitive-bench/benchmark-models",
239
+ "revision": "15c9cd4ab12215ccbcef74a8cd4d051d8e3aae3f",
240
+ "run_id": "atlantis-qwenoft-hist8-zero-s42-v1",
241
+ "task": "atlantis",
242
+ "training_status": "complete",
243
+ "zero_evaluation_status": "complete"
244
+ },
245
+ "profile": {
246
+ "GPU_CLASS": "1x-rtx3090",
247
+ "INSTANCE_ID": "instance_8dd832253b2c3d69",
248
+ "PROFILE_PATH": "results/profiling/profiles/qwenoft/1x-rtx3090/atlantis/instance_8dd832253b2c3d69/profile.json",
249
+ "PROFILE_REF": "qwenoft/1x-rtx3090/atlantis/instance_8dd832253b2c3d69",
250
+ "RUN_ID": "20260924T110210282159Z"
251
+ },
252
+ "profile_path": "profiles/qwenoft/1x-rtx3090/atlantis/instance_8dd832253b2c3d69/profile.json",
253
+ "profile_repo": "latency-sensitive-bench/profiles",
254
+ "profile_revision": "7efc9af422864b9a1cb569aa81f1b2b1e216b89e",
255
+ "profile_sha256": "2c8c1f14845e89c6171e1502ff5af6d61d12049e6445d98d63f31e4ffca384f4",
256
+ "task": "atlantis",
257
+ "verified_sessions": [
258
+ {
259
+ "session": 0,
260
+ "episodes": 10
261
+ },
262
+ {
263
+ "session": 1,
264
+ "episodes": 10
265
+ },
266
+ {
267
+ "session": 2,
268
+ "episodes": 10
269
+ },
270
+ {
271
+ "session": 3,
272
+ "episodes": 10
273
+ },
274
+ {
275
+ "session": 4,
276
+ "episodes": 10
277
+ }
278
+ ],
279
+ "total_episodes": 50,
280
+ "mean_latency_ms": 116.31885849907087,
281
+ "p95_latency_ms": 120.63936149701476,
282
+ "simulator": "cpu",
283
+ "visible_inference_gpus": 1,
284
+ "status": "verified"
285
+ },
286
+ {
287
+ "model": "qwenpi_v3",
288
+ "model_publication": {
289
+ "checkpoint_sha256": "2f4a775690c76dc1b0e4da77a9d45f88df289e17914b895b5a65afa7fc8d214d",
290
+ "episodes": 100,
291
+ "mean_return": 39615.0,
292
+ "model": "qwenpi_v3",
293
+ "path": "zero-latency/atlantis/vla/starvla-qwenpi_v3-h1/atlantis-qwenpi_v3-hist8-zero-s42-v1",
294
+ "repo_id": "latency-sensitive-bench/benchmark-models",
295
+ "revision": "d445334970ab7e110c3cacb5c8b790d3df20e1ec",
296
+ "run_id": "atlantis-qwenpi_v3-hist8-zero-s42-v1",
297
+ "task": "atlantis",
298
+ "training_status": "complete",
299
+ "zero_evaluation_status": "complete"
300
+ },
301
+ "profile": {
302
+ "GPU_CLASS": "1x-rtx3090",
303
+ "INSTANCE_ID": "instance_8dd832253b2c3d69",
304
+ "PROFILE_PATH": "results/profiling/profiles/qwenpi_v3/1x-rtx3090/atlantis/instance_8dd832253b2c3d69/profile.json",
305
+ "PROFILE_REF": "qwenpi_v3/1x-rtx3090/atlantis/instance_8dd832253b2c3d69",
306
+ "RUN_ID": "20260924T122701185096Z"
307
+ },
308
+ "profile_path": "profiles/qwenpi_v3/1x-rtx3090/atlantis/instance_8dd832253b2c3d69/profile.json",
309
+ "profile_repo": "latency-sensitive-bench/profiles",
310
+ "profile_revision": "0de3f61f3635e7e706487c6cae62a4897bf636d5",
311
+ "profile_sha256": "72797b4102f37db4cb13f8a64e26b82d9a0757bf5f1ab8c996546e020790831f",
312
+ "task": "atlantis",
313
+ "verified_sessions": [
314
+ {
315
+ "session": 0,
316
+ "episodes": 10
317
+ },
318
+ {
319
+ "session": 1,
320
+ "episodes": 10
321
+ },
322
+ {
323
+ "session": 2,
324
+ "episodes": 10
325
+ },
326
+ {
327
+ "session": 3,
328
+ "episodes": 10
329
+ },
330
+ {
331
+ "session": 4,
332
+ "episodes": 10
333
+ }
334
+ ],
335
+ "total_episodes": 50,
336
+ "mean_latency_ms": 321.8234160766942,
337
+ "p95_latency_ms": 341.5536783073097,
338
+ "simulator": "cpu",
339
+ "visible_inference_gpus": 1,
340
+ "status": "verified"
341
+ }
342
+ ]
343
+ }
reports/atlantis-asterix-h8-v1/model_inventory.json ADDED
@@ -0,0 +1,124 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ {
2
+ "models": [
3
+ {
4
+ "task": "atlantis",
5
+ "model": "qwengr00t",
6
+ "run_id": "atlantis-qwengr00t-hist8-profile-s42-v1",
7
+ "checkpoint_sha256": "408625af9e210710db6b3b20a6053967e225cbf3abedff0073a231dda6a05e22",
8
+ "training_step": 5000,
9
+ "hf_revision": "b3a6fd68fc00812311dc7feeaccc51b3093dc5e7",
10
+ "hf_path": "latency-aware/atlantis/vla/starvla-qwengr00t-h1/atlantis-qwengr00t-hist8-profile-s42-v1",
11
+ "wandb_url": "https://wandb.ai/saberrr-zju/starvla_tasks/runs/0c35d237"
12
+ },
13
+ {
14
+ "task": "atlantis",
15
+ "model": "qwengr00t",
16
+ "run_id": "atlantis-qwengr00t-hist8-zero-s42-v1",
17
+ "checkpoint_sha256": "dc7081c898fd8324d5c9aa892b5221ad7388a83428cc3508c67b30fc71d0761a",
18
+ "training_step": 5000,
19
+ "hf_revision": "a020cd48a05f6851f25b53c5bc63c46a3d3fb5b8",
20
+ "hf_path": "zero-latency/atlantis/vla/starvla-qwengr00t-h1/atlantis-qwengr00t-hist8-zero-s42-v1",
21
+ "wandb_url": "https://wandb.ai/saberrr-zju/starvla_tasks/runs/60bc68a7"
22
+ },
23
+ {
24
+ "task": "atlantis",
25
+ "model": "qwenoft",
26
+ "run_id": "atlantis-qwenoft-hist8-profile-s42-v1",
27
+ "checkpoint_sha256": "3fc3190e4151c33d203f69da461d6c56042236200ea9f61369a08c97cb28b88b",
28
+ "training_step": 5000,
29
+ "hf_revision": "3bbc0bfd7369c818456525bfbdac9580a382c739",
30
+ "hf_path": "latency-aware/atlantis/vla/starvla-qwenoft-h1/atlantis-qwenoft-hist8-profile-s42-v1",
31
+ "wandb_url": "https://wandb.ai/saberrr-zju/starvla_tasks/runs/7ea15469"
32
+ },
33
+ {
34
+ "task": "atlantis",
35
+ "model": "qwenoft",
36
+ "run_id": "atlantis-qwenoft-hist8-zero-s42-v1",
37
+ "checkpoint_sha256": "f91d4079a94f779856220e4a954750624228a3da596b5ecdd70861e888a97877",
38
+ "training_step": 5000,
39
+ "hf_revision": "2c66bf8b40bf9dc6fd9b9ea26ccd9fb082f3b883",
40
+ "hf_path": "zero-latency/atlantis/vla/starvla-qwenoft-h1/atlantis-qwenoft-hist8-zero-s42-v1",
41
+ "wandb_url": "https://wandb.ai/saberrr-zju/starvla_tasks/runs/40be052b"
42
+ },
43
+ {
44
+ "task": "atlantis",
45
+ "model": "qwenpi_v3",
46
+ "run_id": "atlantis-qwenpi_v3-hist8-profile-s42-v1",
47
+ "checkpoint_sha256": "8da09e91d1c6c3908cd1536d6c072e3b6f669b40c82fe9e98705f363691e10ba",
48
+ "training_step": 5000,
49
+ "hf_revision": "9502b332f984a867aa753d425d510775a389e0e3",
50
+ "hf_path": "latency-aware/atlantis/vla/starvla-qwenpi_v3-h1/atlantis-qwenpi_v3-hist8-profile-s42-v1",
51
+ "wandb_url": "https://wandb.ai/saberrr-zju/starvla_tasks/runs/e58d89ce"
52
+ },
53
+ {
54
+ "task": "atlantis",
55
+ "model": "qwenpi_v3",
56
+ "run_id": "atlantis-qwenpi_v3-hist8-zero-s42-v1",
57
+ "checkpoint_sha256": "2f4a775690c76dc1b0e4da77a9d45f88df289e17914b895b5a65afa7fc8d214d",
58
+ "training_step": 5000,
59
+ "hf_revision": "576668278cd0a3be955f127d1ec13b026b3c1b13",
60
+ "hf_path": "zero-latency/atlantis/vla/starvla-qwenpi_v3-h1/atlantis-qwenpi_v3-hist8-zero-s42-v1",
61
+ "wandb_url": "https://wandb.ai/saberrr-zju/starvla_tasks/runs/04d16bbd"
62
+ },
63
+ {
64
+ "task": "asterix",
65
+ "model": "qwengr00t",
66
+ "run_id": "asterix-qwengr00t-hist8-profile-s42-v1",
67
+ "checkpoint_sha256": "a192db0f4f9707271e453150467a4d833c29479f3e6f6f8031039e84dab8afcf",
68
+ "training_step": 5000,
69
+ "hf_revision": "6d11a6d4946f278d9b165d260496cd757b330c3d",
70
+ "hf_path": "latency-aware/asterix/vla/starvla-qwengr00t-h1/asterix-qwengr00t-hist8-profile-s42-v1",
71
+ "wandb_url": "https://wandb.ai/saberrr-zju/starvla_tasks/runs/946c608a"
72
+ },
73
+ {
74
+ "task": "asterix",
75
+ "model": "qwengr00t",
76
+ "run_id": "asterix-qwengr00t-hist8-zero-s42-v1",
77
+ "checkpoint_sha256": "7b3541140bd2d00d39eca98c8788b773e594884a623ae20d7b9e4dd7b45337f5",
78
+ "training_step": 5000,
79
+ "hf_revision": "22e43559d63f898ca1b9bc0418b65a79efcaffd5",
80
+ "hf_path": "zero-latency/asterix/vla/starvla-qwengr00t-h1/asterix-qwengr00t-hist8-zero-s42-v1",
81
+ "wandb_url": "https://wandb.ai/saberrr-zju/starvla_tasks/runs/42b8e7f3"
82
+ },
83
+ {
84
+ "task": "asterix",
85
+ "model": "qwenoft",
86
+ "run_id": "asterix-qwenoft-hist8-profile-s42-v1",
87
+ "checkpoint_sha256": "b96bf1398d3d767b13be1cc76eefdfe70fdab349231ef5a4636495c548fe09ac",
88
+ "training_step": 5000,
89
+ "hf_revision": "a4dd7774b90faf83cfec251a3a23249e3bbac080",
90
+ "hf_path": "latency-aware/asterix/vla/starvla-qwenoft-h1/asterix-qwenoft-hist8-profile-s42-v1",
91
+ "wandb_url": "https://wandb.ai/saberrr-zju/starvla_tasks/runs/ac01280d"
92
+ },
93
+ {
94
+ "task": "asterix",
95
+ "model": "qwenoft",
96
+ "run_id": "asterix-qwenoft-hist8-zero-s42-v1",
97
+ "checkpoint_sha256": "9d5be1187e768a3ba6638610ad70ca58fed54a531b77b0b0dea3a33538def572",
98
+ "training_step": 5000,
99
+ "hf_revision": "d445334970ab7e110c3cacb5c8b790d3df20e1ec",
100
+ "hf_path": "zero-latency/asterix/vla/starvla-qwenoft-h1/asterix-qwenoft-hist8-zero-s42-v1",
101
+ "wandb_url": "https://wandb.ai/saberrr-zju/starvla_tasks/runs/aa718e0f"
102
+ },
103
+ {
104
+ "task": "asterix",
105
+ "model": "qwenpi_v3",
106
+ "run_id": "asterix-qwenpi_v3-hist8-profile-s42-v1",
107
+ "checkpoint_sha256": "19fc503b7ab9158c445f92258e489da35d2a73a87c12e3e7954a7e0f51ea75eb",
108
+ "training_step": 5000,
109
+ "hf_revision": "b53f7bbd530c986912433448d7dc52952349bd99",
110
+ "hf_path": "latency-aware/asterix/vla/starvla-qwenpi_v3-h1/asterix-qwenpi_v3-hist8-profile-s42-v1",
111
+ "wandb_url": "https://wandb.ai/saberrr-zju/starvla_tasks/runs/f9d438a5"
112
+ },
113
+ {
114
+ "task": "asterix",
115
+ "model": "qwenpi_v3",
116
+ "run_id": "asterix-qwenpi_v3-hist8-zero-s42-v1",
117
+ "checkpoint_sha256": "381d1ed73a6f3949cb1880d6f1feb18db13068aeb6db39a7974ae9fb1fba0f56",
118
+ "training_step": 5000,
119
+ "hf_revision": "8336631d8221b15728cd5c32ff242c339499276f",
120
+ "hf_path": "zero-latency/asterix/vla/starvla-qwenpi_v3-h1/asterix-qwenpi_v3-hist8-zero-s42-v1",
121
+ "wandb_url": "https://wandb.ai/saberrr-zju/starvla_tasks/runs/ec20df07"
122
+ }
123
+ ]
124
+ }
reports/atlantis-asterix-h8-v1/profile_data.complete.json ADDED
@@ -0,0 +1,78 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ {
2
+ "status": "complete",
3
+ "checked_at": 1790268505.1894896,
4
+ "datasets": [
5
+ {
6
+ "task": "atlantis",
7
+ "model": "qwenoft",
8
+ "episodes": 200,
9
+ "train_labels": 322419,
10
+ "val_labels": 35989,
11
+ "checked_rows": 358408,
12
+ "teacher_revision": "34f68e5162100362de0710d7745edba677fc0957",
13
+ "profile_revision": "7efc9af422864b9a1cb569aa81f1b2b1e216b89e",
14
+ "evidence": "data_verification/atlantis-qwenoft",
15
+ "status": "verified"
16
+ },
17
+ {
18
+ "task": "atlantis",
19
+ "model": "qwengr00t",
20
+ "episodes": 250,
21
+ "train_labels": 365845,
22
+ "val_labels": 39091,
23
+ "checked_rows": 404936,
24
+ "teacher_revision": "de60cabd0f3b7484773374f83be46a777657f74a",
25
+ "profile_revision": "a00f633aa3227e2be8648b4f1337d7ccd188d18f",
26
+ "evidence": "data_verification/atlantis-qwengr00t",
27
+ "status": "verified"
28
+ },
29
+ {
30
+ "task": "atlantis",
31
+ "model": "qwenpi_v3",
32
+ "episodes": 250,
33
+ "train_labels": 362699,
34
+ "val_labels": 39022,
35
+ "checked_rows": 401721,
36
+ "teacher_revision": "5ed3e54aac1bd957c882b426218ca2e76db5b720",
37
+ "profile_revision": "0de3f61f3635e7e706487c6cae62a4897bf636d5",
38
+ "evidence": "data_verification/atlantis-qwenpi_v3",
39
+ "status": "verified"
40
+ },
41
+ {
42
+ "task": "asterix",
43
+ "model": "qwenoft",
44
+ "episodes": 350,
45
+ "train_labels": 323534,
46
+ "val_labels": 38047,
47
+ "checked_rows": 361581,
48
+ "teacher_revision": "da734589fb8669108a8021ea79f823665e1b11c8",
49
+ "profile_revision": "fbbba76b21fa705cf4e5d05ec1ecfb82a88b12d8",
50
+ "evidence": "data_verification/asterix-qwenoft",
51
+ "status": "verified"
52
+ },
53
+ {
54
+ "task": "asterix",
55
+ "model": "qwengr00t",
56
+ "episodes": 550,
57
+ "train_labels": 345402,
58
+ "val_labels": 36335,
59
+ "checked_rows": 381737,
60
+ "teacher_revision": "b2fe384570e74126afe1b9f75ca523747809bd36",
61
+ "profile_revision": "15001405a402e30b7706476bc92aea45a61bb193",
62
+ "evidence": "data_verification/asterix-qwengr00t",
63
+ "status": "verified"
64
+ },
65
+ {
66
+ "task": "asterix",
67
+ "model": "qwenpi_v3",
68
+ "episodes": 800,
69
+ "train_labels": 337461,
70
+ "val_labels": 37403,
71
+ "checked_rows": 374864,
72
+ "teacher_revision": "e90f33e9f069a205f56b8d2e1e64e08d879cd821",
73
+ "profile_revision": "36f784f257aabf73721d5edd63abc7b3dee1a272",
74
+ "evidence": "data_verification/asterix-qwenpi_v3",
75
+ "status": "verified"
76
+ }
77
+ ]
78
+ }
reports/atlantis-asterix-h8-v1/profile_training.complete.json ADDED
@@ -0,0 +1,55 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ {
2
+ "training_status": "complete",
3
+ "checked_at": 1790278366.7720268,
4
+ "runs": [
5
+ {
6
+ "run_id": "atlantis-qwenoft-hist8-profile-s42-v1",
7
+ "training_step": 5000,
8
+ "checkpoint_sha256": "3fc3190e4151c33d203f69da461d6c56042236200ea9f61369a08c97cb28b88b",
9
+ "wandb_url": "https://wandb.ai/saberrr-zju/starvla_tasks/runs/7ea15469",
10
+ "validation_loss": 0.14601874351501465,
11
+ "evidence": "verification/atlantis_qwenoft_profile_training.json"
12
+ },
13
+ {
14
+ "run_id": "atlantis-qwengr00t-hist8-profile-s42-v1",
15
+ "training_step": 5000,
16
+ "checkpoint_sha256": "408625af9e210710db6b3b20a6053967e225cbf3abedff0073a231dda6a05e22",
17
+ "wandb_url": "https://wandb.ai/saberrr-zju/starvla_tasks/runs/0c35d237",
18
+ "validation_loss": 0.022569088265299797,
19
+ "evidence": "verification/atlantis_qwengr00t_profile_training.json"
20
+ },
21
+ {
22
+ "run_id": "atlantis-qwenpi_v3-hist8-profile-s42-v1",
23
+ "training_step": 5000,
24
+ "checkpoint_sha256": "8da09e91d1c6c3908cd1536d6c072e3b6f669b40c82fe9e98705f363691e10ba",
25
+ "wandb_url": "https://wandb.ai/saberrr-zju/starvla_tasks/runs/e58d89ce",
26
+ "validation_loss": 0.04210428148508072,
27
+ "evidence": "verification/atlantis_qwenpi_v3_profile_training.json"
28
+ },
29
+ {
30
+ "run_id": "asterix-qwenoft-hist8-profile-s42-v1",
31
+ "training_step": 5000,
32
+ "checkpoint_sha256": "b96bf1398d3d767b13be1cc76eefdfe70fdab349231ef5a4636495c548fe09ac",
33
+ "wandb_url": "https://wandb.ai/saberrr-zju/starvla_tasks/runs/ac01280d",
34
+ "validation_loss": 0.198922336101532,
35
+ "evidence": "verification/asterix_qwenoft_profile_training.json"
36
+ },
37
+ {
38
+ "run_id": "asterix-qwengr00t-hist8-profile-s42-v1",
39
+ "training_step": 5000,
40
+ "checkpoint_sha256": "a192db0f4f9707271e453150467a4d833c29479f3e6f6f8031039e84dab8afcf",
41
+ "wandb_url": "https://wandb.ai/saberrr-zju/starvla_tasks/runs/946c608a",
42
+ "validation_loss": 0.04735454171895981,
43
+ "evidence": "verification/asterix_qwengr00t_profile_training.json"
44
+ },
45
+ {
46
+ "run_id": "asterix-qwenpi_v3-hist8-profile-s42-v1",
47
+ "training_step": 5000,
48
+ "checkpoint_sha256": "19fc503b7ab9158c445f92258e489da35d2a73a87c12e3e7954a7e0f51ea75eb",
49
+ "wandb_url": "https://wandb.ai/saberrr-zju/starvla_tasks/runs/f9d438a5",
50
+ "validation_loss": 0.04895472526550293,
51
+ "evidence": "verification/asterix_qwenpi_v3_profile_training.json"
52
+ }
53
+ ],
54
+ "scope": "training only; evaluation and publication tracked separately"
55
+ }
reports/atlantis-asterix-h8-v1/teachers.complete.json ADDED
@@ -0,0 +1,116 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ {
2
+ "status": "complete",
3
+ "checked_at": 1790262550.2685425,
4
+ "teachers": [
5
+ {
6
+ "task": "atlantis",
7
+ "model": "qwenoft",
8
+ "episodes": 100,
9
+ "mean_return": 38411.0,
10
+ "best_checkpoint_env_steps": 14508032,
11
+ "final_env_steps": 15007744,
12
+ "hf_revision": "34f68e5162100362de0710d7745edba677fc0957",
13
+ "hf_path": "latency-aware/atlantis/small-policy/sample-factory-qwenoft-hist8-cpu15m-s0-v1",
14
+ "checkpoint_sha256": "40bdf046972b5794c4e48ad7ca919c9a601958a88e113089919ede354a7789f6",
15
+ "profile_sha256": "2c8c1f14845e89c6171e1502ff5af6d61d12049e6445d98d63f31e4ffca384f4",
16
+ "profile_revision": "7efc9af422864b9a1cb569aa81f1b2b1e216b89e",
17
+ "wandb_readback": {
18
+ "eval_corrected/profile_sample/profile_sample/mean_return": 38411
19
+ },
20
+ "evidence": "teacher_verification/atlantis-qwenoft",
21
+ "status": "verified"
22
+ },
23
+ {
24
+ "task": "atlantis",
25
+ "model": "qwengr00t",
26
+ "episodes": 100,
27
+ "mean_return": 29064.0,
28
+ "best_checkpoint_env_steps": 12156928,
29
+ "final_env_steps": 15007744,
30
+ "hf_revision": "de60cabd0f3b7484773374f83be46a777657f74a",
31
+ "hf_path": "latency-aware/atlantis/small-policy/sample-factory-qwengr00t-hist8-cpu15m-s0-v1",
32
+ "checkpoint_sha256": "8289271b5da29b9cfaf30017f86c457e1a1ed0edd9f998a9f5844354015f0238",
33
+ "profile_sha256": "13e970f85825c898ecb5730ff6adb6c71995a68406227733311a15b3f0c202e9",
34
+ "profile_revision": "a00f633aa3227e2be8648b4f1337d7ccd188d18f",
35
+ "wandb_readback": {
36
+ "eval_corrected/profile_sample/profile_sample/mean_return": 29064
37
+ },
38
+ "evidence": "teacher_verification/atlantis-qwengr00t",
39
+ "status": "verified"
40
+ },
41
+ {
42
+ "task": "atlantis",
43
+ "model": "qwenpi_v3",
44
+ "episodes": 100,
45
+ "mean_return": 27780.0,
46
+ "best_checkpoint_env_steps": 14065664,
47
+ "final_env_steps": 15007744,
48
+ "hf_revision": "5ed3e54aac1bd957c882b426218ca2e76db5b720",
49
+ "hf_path": "latency-aware/atlantis/small-policy/sample-factory-qwenpi_v3-hist8-cpu15m-s0-v1",
50
+ "checkpoint_sha256": "47af64b82e2d905e3c0d4c671d5957e66921bb56c3275ee8721c979b597e26f7",
51
+ "profile_sha256": "72797b4102f37db4cb13f8a64e26b82d9a0757bf5f1ab8c996546e020790831f",
52
+ "profile_revision": "0de3f61f3635e7e706487c6cae62a4897bf636d5",
53
+ "wandb_readback": {
54
+ "eval_corrected/profile_sample/profile_sample/mean_return": 27780
55
+ },
56
+ "evidence": "teacher_verification/atlantis-qwenpi_v3",
57
+ "status": "verified"
58
+ },
59
+ {
60
+ "task": "asterix",
61
+ "model": "qwenoft",
62
+ "episodes": 100,
63
+ "mean_return": 3941.0,
64
+ "best_checkpoint_env_steps": 13893632,
65
+ "final_env_steps": 15007744,
66
+ "hf_revision": "da734589fb8669108a8021ea79f823665e1b11c8",
67
+ "hf_path": "latency-aware/asterix/small-policy/sample-factory-qwenoft-hist8-cpu15m-s0-v1",
68
+ "checkpoint_sha256": "93ba21325d92fa47f617c3daaffd9d7604a8c93bf8fe4bc1d4aacc4facd5430c",
69
+ "profile_sha256": "4e2e3ed923c7500248ce4ce1508f124faec78282d90d6b99d070a9e29bdb6146",
70
+ "profile_revision": "fbbba76b21fa705cf4e5d05ec1ecfb82a88b12d8",
71
+ "wandb_readback": {
72
+ "eval_corrected/profile_sample/profile_sample/mean_return": 3941
73
+ },
74
+ "evidence": "teacher_verification/asterix-qwenoft",
75
+ "status": "verified"
76
+ },
77
+ {
78
+ "task": "asterix",
79
+ "model": "qwengr00t",
80
+ "episodes": 100,
81
+ "mean_return": 1852.5,
82
+ "best_checkpoint_env_steps": 13975552,
83
+ "final_env_steps": 15007744,
84
+ "hf_revision": "b2fe384570e74126afe1b9f75ca523747809bd36",
85
+ "hf_path": "latency-aware/asterix/small-policy/sample-factory-qwengr00t-hist8-cpu15m-s0-v1",
86
+ "checkpoint_sha256": "8fc78a650079a3fc14de7667d3a4750edfa4d7481762719888be571559cfa8e0",
87
+ "profile_sha256": "5e104064d718ac51ce56aefac5318a155fd099f5c714ff8f0bc4985509639fd1",
88
+ "profile_revision": "15001405a402e30b7706476bc92aea45a61bb193",
89
+ "wandb_readback": {
90
+ "eval_corrected/profile_sample/profile_sample/mean_return": 1852.5
91
+ },
92
+ "evidence": "teacher_verification/asterix-qwengr00t",
93
+ "status": "verified"
94
+ },
95
+ {
96
+ "task": "asterix",
97
+ "model": "qwenpi_v3",
98
+ "episodes": 100,
99
+ "mean_return": 1141.0,
100
+ "best_checkpoint_env_steps": 14974976,
101
+ "final_env_steps": 15007744,
102
+ "hf_revision": "e90f33e9f069a205f56b8d2e1e64e08d879cd821",
103
+ "hf_path": "latency-aware/asterix/small-policy/sample-factory-qwenpi_v3-hist8-cpu15m-s0-v1",
104
+ "checkpoint_sha256": "a33ccc8e9f4921d176a6c257dab1b3f9b824d5db99b1908b4e73a201453f216f",
105
+ "profile_sha256": "a5acfbb259cc3431bcce9dedef862a1e3895bf81659aa3c1707fa54c277852fc",
106
+ "profile_revision": "36f784f257aabf73721d5edd63abc7b3dee1a272",
107
+ "wandb_readback": {
108
+ "eval_corrected/profile_sample/profile_sample/mean_return": 1141
109
+ },
110
+ "evidence": "teacher_verification/asterix-qwenpi_v3",
111
+ "status": "verified"
112
+ }
113
+ ],
114
+ "total_eval_episodes": 600,
115
+ "publication_verification": "catalog emitted only after every published file was read back at its fixed HF revision"
116
+ }
reports/atlantis-asterix-h8-v1/verified_evaluations.json ADDED
@@ -0,0 +1,2970 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ {
2
+ "checked_at": 1790279163.3969283,
3
+ "evaluations": [
4
+ {
5
+ "task": "atlantis",
6
+ "model": "qwengr00t",
7
+ "run_id": "atlantis-qwengr00t-hist8-profile-s42-v1",
8
+ "condition": "profile_to_profile",
9
+ "episodes": 100,
10
+ "mean_return": 28488.0,
11
+ "std_return": 7787.377992623705,
12
+ "mean_raw_frames": 6222.63,
13
+ "true_episode_end_count": 38,
14
+ "frame_cap_count": 62,
15
+ "action_audit": {
16
+ "eval_protocol_version": "hist8-history01-sf-render-v4",
17
+ "history_encoding": "binary_0_1",
18
+ "episodes": 100,
19
+ "episode_seeds": [
20
+ 1000000,
21
+ 1000001,
22
+ 1000002,
23
+ 1000003,
24
+ 1000004,
25
+ 1000005,
26
+ 1000006,
27
+ 1000007,
28
+ 1000008,
29
+ 1000009,
30
+ 1000010,
31
+ 1000011,
32
+ 1000012,
33
+ 1000013,
34
+ 1000014,
35
+ 1000015,
36
+ 1000016,
37
+ 1000017,
38
+ 1000018,
39
+ 1000019,
40
+ 1000020,
41
+ 1000021,
42
+ 1000022,
43
+ 1000023,
44
+ 1000024,
45
+ 1000025,
46
+ 1000026,
47
+ 1000027,
48
+ 1000028,
49
+ 1000029,
50
+ 1000030,
51
+ 1000031,
52
+ 1000032,
53
+ 1000033,
54
+ 1000034,
55
+ 1000035,
56
+ 1000036,
57
+ 1000037,
58
+ 1000038,
59
+ 1000039,
60
+ 1000040,
61
+ 1000041,
62
+ 1000042,
63
+ 1000043,
64
+ 1000044,
65
+ 1000045,
66
+ 1000046,
67
+ 1000047,
68
+ 1000048,
69
+ 1000049,
70
+ 1000050,
71
+ 1000051,
72
+ 1000052,
73
+ 1000053,
74
+ 1000054,
75
+ 1000055,
76
+ 1000056,
77
+ 1000057,
78
+ 1000058,
79
+ 1000059,
80
+ 1000060,
81
+ 1000061,
82
+ 1000062,
83
+ 1000063,
84
+ 1000064,
85
+ 1000065,
86
+ 1000066,
87
+ 1000067,
88
+ 1000068,
89
+ 1000069,
90
+ 1000070,
91
+ 1000071,
92
+ 1000072,
93
+ 1000073,
94
+ 1000074,
95
+ 1000075,
96
+ 1000076,
97
+ 1000077,
98
+ 1000078,
99
+ 1000079,
100
+ 1000080,
101
+ 1000081,
102
+ 1000082,
103
+ 1000083,
104
+ 1000084,
105
+ 1000085,
106
+ 1000086,
107
+ 1000087,
108
+ 1000088,
109
+ 1000089,
110
+ 1000090,
111
+ 1000091,
112
+ 1000092,
113
+ 1000093,
114
+ 1000094,
115
+ 1000095,
116
+ 1000096,
117
+ 1000097,
118
+ 1000098,
119
+ 1000099
120
+ ],
121
+ "issued_actions": 38910,
122
+ "action_counts": {
123
+ "2": 18234,
124
+ "0": 19455,
125
+ "3": 1221
126
+ },
127
+ "largest_action_fraction": 0.5,
128
+ "raw_score_min": [
129
+ -0.15221619606018066,
130
+ -0.14649659395217896,
131
+ -0.13478434085845947,
132
+ -0.2636457681655884
133
+ ],
134
+ "raw_score_max": [
135
+ 1.0779457092285156,
136
+ 0.11617469787597656,
137
+ 1.1053106784820557,
138
+ 1.0580275058746338
139
+ ],
140
+ "raw_score_mean": [
141
+ 0.5025190477344667,
142
+ -0.0010393734900449576,
143
+ 0.4768200179841657,
144
+ 0.027724201036296428
145
+ ]
146
+ },
147
+ "checkpoint_sha256": "408625af9e210710db6b3b20a6053967e225cbf3abedff0073a231dda6a05e22",
148
+ "training_step": 5000,
149
+ "profile_sha256": "13e970f85825c898ecb5730ff6adb6c71995a68406227733311a15b3f0c202e9",
150
+ "hf_revision": "b3a6fd68fc00812311dc7feeaccc51b3093dc5e7",
151
+ "hf_path": "latency-aware/atlantis/vla/starvla-qwengr00t-h1/atlantis-qwengr00t-hist8-profile-s42-v1",
152
+ "wandb_url": "https://wandb.ai/saberrr-zju/starvla_tasks/runs/0c35d237",
153
+ "wandb_readback": {
154
+ "eval_atari_hist8_v1/profile_to_profile/profile_sample/profile_sample/mean_return": 28488
155
+ },
156
+ "status": "verified"
157
+ },
158
+ {
159
+ "task": "atlantis",
160
+ "model": "qwengr00t",
161
+ "run_id": "atlantis-qwengr00t-hist8-zero-s42-v1",
162
+ "condition": "zero_to_zero",
163
+ "episodes": 100,
164
+ "mean_return": 39662.0,
165
+ "std_return": 2849.483461962887,
166
+ "mean_raw_frames": 7200.0,
167
+ "true_episode_end_count": 0,
168
+ "frame_cap_count": 100,
169
+ "action_audit": {
170
+ "eval_protocol_version": "hist8-history01-sf-render-v4",
171
+ "history_encoding": "binary_0_1",
172
+ "episodes": 100,
173
+ "episode_seeds": [
174
+ 1000000,
175
+ 1000001,
176
+ 1000002,
177
+ 1000003,
178
+ 1000004,
179
+ 1000005,
180
+ 1000006,
181
+ 1000007,
182
+ 1000008,
183
+ 1000009,
184
+ 1000010,
185
+ 1000011,
186
+ 1000012,
187
+ 1000013,
188
+ 1000014,
189
+ 1000015,
190
+ 1000016,
191
+ 1000017,
192
+ 1000018,
193
+ 1000019,
194
+ 1000020,
195
+ 1000021,
196
+ 1000022,
197
+ 1000023,
198
+ 1000024,
199
+ 1000025,
200
+ 1000026,
201
+ 1000027,
202
+ 1000028,
203
+ 1000029,
204
+ 1000030,
205
+ 1000031,
206
+ 1000032,
207
+ 1000033,
208
+ 1000034,
209
+ 1000035,
210
+ 1000036,
211
+ 1000037,
212
+ 1000038,
213
+ 1000039,
214
+ 1000040,
215
+ 1000041,
216
+ 1000042,
217
+ 1000043,
218
+ 1000044,
219
+ 1000045,
220
+ 1000046,
221
+ 1000047,
222
+ 1000048,
223
+ 1000049,
224
+ 1000050,
225
+ 1000051,
226
+ 1000052,
227
+ 1000053,
228
+ 1000054,
229
+ 1000055,
230
+ 1000056,
231
+ 1000057,
232
+ 1000058,
233
+ 1000059,
234
+ 1000060,
235
+ 1000061,
236
+ 1000062,
237
+ 1000063,
238
+ 1000064,
239
+ 1000065,
240
+ 1000066,
241
+ 1000067,
242
+ 1000068,
243
+ 1000069,
244
+ 1000070,
245
+ 1000071,
246
+ 1000072,
247
+ 1000073,
248
+ 1000074,
249
+ 1000075,
250
+ 1000076,
251
+ 1000077,
252
+ 1000078,
253
+ 1000079,
254
+ 1000080,
255
+ 1000081,
256
+ 1000082,
257
+ 1000083,
258
+ 1000084,
259
+ 1000085,
260
+ 1000086,
261
+ 1000087,
262
+ 1000088,
263
+ 1000089,
264
+ 1000090,
265
+ 1000091,
266
+ 1000092,
267
+ 1000093,
268
+ 1000094,
269
+ 1000095,
270
+ 1000096,
271
+ 1000097,
272
+ 1000098,
273
+ 1000099
274
+ ],
275
+ "issued_actions": 180000,
276
+ "action_counts": {
277
+ "0": 58729,
278
+ "2": 51845,
279
+ "3": 32048,
280
+ "1": 37378
281
+ },
282
+ "largest_action_fraction": 0.32627222222222224,
283
+ "raw_score_min": [
284
+ -0.09850892424583435,
285
+ -0.1825193166732788,
286
+ -0.18227225542068481,
287
+ -0.49613189697265625
288
+ ],
289
+ "raw_score_max": [
290
+ 1.1174542903900146,
291
+ 1.10590398311615,
292
+ 1.1816869974136353,
293
+ 1.0849536657333374
294
+ ],
295
+ "raw_score_mean": [
296
+ 0.32900199618344067,
297
+ 0.21200553999988964,
298
+ 0.2875584516548888,
299
+ 0.1787945480706416
300
+ ]
301
+ },
302
+ "checkpoint_sha256": "dc7081c898fd8324d5c9aa892b5221ad7388a83428cc3508c67b30fc71d0761a",
303
+ "training_step": 5000,
304
+ "profile_sha256": null,
305
+ "hf_revision": "092da3f6bbb57f817376ff9f19ea75580152d163",
306
+ "hf_path": "zero-latency/atlantis/vla/starvla-qwengr00t-h1/atlantis-qwengr00t-hist8-zero-s42-v1",
307
+ "wandb_url": "https://wandb.ai/saberrr-zju/starvla_tasks/runs/60bc68a7",
308
+ "wandb_readback": {
309
+ "eval_atari_hist8_v1/zero_to_zero/fixed/latency_0/mean_return": 39662
310
+ },
311
+ "status": "verified"
312
+ },
313
+ {
314
+ "task": "atlantis",
315
+ "model": "qwengr00t",
316
+ "run_id": "atlantis-qwengr00t-hist8-zero-s42-v1",
317
+ "condition": "zero_to_profile",
318
+ "episodes": 100,
319
+ "mean_return": 8167.0,
320
+ "std_return": 3419.4167631337364,
321
+ "mean_raw_frames": 3416.06,
322
+ "true_episode_end_count": 100,
323
+ "frame_cap_count": 0,
324
+ "action_audit": {
325
+ "eval_protocol_version": "hist8-history01-sf-render-v4",
326
+ "history_encoding": "binary_0_1",
327
+ "episodes": 100,
328
+ "episode_seeds": [
329
+ 1000000,
330
+ 1000001,
331
+ 1000002,
332
+ 1000003,
333
+ 1000004,
334
+ 1000005,
335
+ 1000006,
336
+ 1000007,
337
+ 1000008,
338
+ 1000009,
339
+ 1000010,
340
+ 1000011,
341
+ 1000012,
342
+ 1000013,
343
+ 1000014,
344
+ 1000015,
345
+ 1000016,
346
+ 1000017,
347
+ 1000018,
348
+ 1000019,
349
+ 1000020,
350
+ 1000021,
351
+ 1000022,
352
+ 1000023,
353
+ 1000024,
354
+ 1000025,
355
+ 1000026,
356
+ 1000027,
357
+ 1000028,
358
+ 1000029,
359
+ 1000030,
360
+ 1000031,
361
+ 1000032,
362
+ 1000033,
363
+ 1000034,
364
+ 1000035,
365
+ 1000036,
366
+ 1000037,
367
+ 1000038,
368
+ 1000039,
369
+ 1000040,
370
+ 1000041,
371
+ 1000042,
372
+ 1000043,
373
+ 1000044,
374
+ 1000045,
375
+ 1000046,
376
+ 1000047,
377
+ 1000048,
378
+ 1000049,
379
+ 1000050,
380
+ 1000051,
381
+ 1000052,
382
+ 1000053,
383
+ 1000054,
384
+ 1000055,
385
+ 1000056,
386
+ 1000057,
387
+ 1000058,
388
+ 1000059,
389
+ 1000060,
390
+ 1000061,
391
+ 1000062,
392
+ 1000063,
393
+ 1000064,
394
+ 1000065,
395
+ 1000066,
396
+ 1000067,
397
+ 1000068,
398
+ 1000069,
399
+ 1000070,
400
+ 1000071,
401
+ 1000072,
402
+ 1000073,
403
+ 1000074,
404
+ 1000075,
405
+ 1000076,
406
+ 1000077,
407
+ 1000078,
408
+ 1000079,
409
+ 1000080,
410
+ 1000081,
411
+ 1000082,
412
+ 1000083,
413
+ 1000084,
414
+ 1000085,
415
+ 1000086,
416
+ 1000087,
417
+ 1000088,
418
+ 1000089,
419
+ 1000090,
420
+ 1000091,
421
+ 1000092,
422
+ 1000093,
423
+ 1000094,
424
+ 1000095,
425
+ 1000096,
426
+ 1000097,
427
+ 1000098,
428
+ 1000099
429
+ ],
430
+ "issued_actions": 21398,
431
+ "action_counts": {
432
+ "0": 8381,
433
+ "3": 4882,
434
+ "2": 1906,
435
+ "1": 6229
436
+ },
437
+ "largest_action_fraction": 0.39167211888961584,
438
+ "raw_score_min": [
439
+ -0.3323414921760559,
440
+ -0.7480053305625916,
441
+ -0.6249756813049316,
442
+ -1.3646661043167114
443
+ ],
444
+ "raw_score_max": [
445
+ 1.4730366468429565,
446
+ 1.3714518547058105,
447
+ 1.3062940835952759,
448
+ 1.2629424333572388
449
+ ],
450
+ "raw_score_mean": [
451
+ 0.40206359148115434,
452
+ 0.29438784433748055,
453
+ 0.08243494440433187,
454
+ 0.2276015812253758
455
+ ]
456
+ },
457
+ "checkpoint_sha256": "dc7081c898fd8324d5c9aa892b5221ad7388a83428cc3508c67b30fc71d0761a",
458
+ "training_step": 5000,
459
+ "profile_sha256": "13e970f85825c898ecb5730ff6adb6c71995a68406227733311a15b3f0c202e9",
460
+ "hf_revision": "a020cd48a05f6851f25b53c5bc63c46a3d3fb5b8",
461
+ "hf_path": "zero-latency/atlantis/vla/starvla-qwengr00t-h1/atlantis-qwengr00t-hist8-zero-s42-v1",
462
+ "wandb_url": "https://wandb.ai/saberrr-zju/starvla_tasks/runs/60bc68a7",
463
+ "wandb_readback": {
464
+ "eval_atari_hist8_v1/zero_to_profile/profile_sample/profile_sample/mean_return": 8167
465
+ },
466
+ "status": "verified"
467
+ },
468
+ {
469
+ "task": "atlantis",
470
+ "model": "qwenoft",
471
+ "run_id": "atlantis-qwenoft-hist8-profile-s42-v1",
472
+ "condition": "profile_to_profile",
473
+ "episodes": 100,
474
+ "mean_return": 39229.0,
475
+ "std_return": 4808.66499145033,
476
+ "mean_raw_frames": 7188.68,
477
+ "true_episode_end_count": 3,
478
+ "frame_cap_count": 97,
479
+ "action_audit": {
480
+ "eval_protocol_version": "hist8-history01-sf-render-v4",
481
+ "history_encoding": "binary_0_1",
482
+ "episodes": 100,
483
+ "episode_seeds": [
484
+ 1000000,
485
+ 1000001,
486
+ 1000002,
487
+ 1000003,
488
+ 1000004,
489
+ 1000005,
490
+ 1000006,
491
+ 1000007,
492
+ 1000008,
493
+ 1000009,
494
+ 1000010,
495
+ 1000011,
496
+ 1000012,
497
+ 1000013,
498
+ 1000014,
499
+ 1000015,
500
+ 1000016,
501
+ 1000017,
502
+ 1000018,
503
+ 1000019,
504
+ 1000020,
505
+ 1000021,
506
+ 1000022,
507
+ 1000023,
508
+ 1000024,
509
+ 1000025,
510
+ 1000026,
511
+ 1000027,
512
+ 1000028,
513
+ 1000029,
514
+ 1000030,
515
+ 1000031,
516
+ 1000032,
517
+ 1000033,
518
+ 1000034,
519
+ 1000035,
520
+ 1000036,
521
+ 1000037,
522
+ 1000038,
523
+ 1000039,
524
+ 1000040,
525
+ 1000041,
526
+ 1000042,
527
+ 1000043,
528
+ 1000044,
529
+ 1000045,
530
+ 1000046,
531
+ 1000047,
532
+ 1000048,
533
+ 1000049,
534
+ 1000050,
535
+ 1000051,
536
+ 1000052,
537
+ 1000053,
538
+ 1000054,
539
+ 1000055,
540
+ 1000056,
541
+ 1000057,
542
+ 1000058,
543
+ 1000059,
544
+ 1000060,
545
+ 1000061,
546
+ 1000062,
547
+ 1000063,
548
+ 1000064,
549
+ 1000065,
550
+ 1000066,
551
+ 1000067,
552
+ 1000068,
553
+ 1000069,
554
+ 1000070,
555
+ 1000071,
556
+ 1000072,
557
+ 1000073,
558
+ 1000074,
559
+ 1000075,
560
+ 1000076,
561
+ 1000077,
562
+ 1000078,
563
+ 1000079,
564
+ 1000080,
565
+ 1000081,
566
+ 1000082,
567
+ 1000083,
568
+ 1000084,
569
+ 1000085,
570
+ 1000086,
571
+ 1000087,
572
+ 1000088,
573
+ 1000089,
574
+ 1000090,
575
+ 1000091,
576
+ 1000092,
577
+ 1000093,
578
+ 1000094,
579
+ 1000095,
580
+ 1000096,
581
+ 1000097,
582
+ 1000098,
583
+ 1000099
584
+ ],
585
+ "issued_actions": 89771,
586
+ "action_counts": {
587
+ "2": 30582,
588
+ "0": 40406,
589
+ "1": 11958,
590
+ "3": 6825
591
+ },
592
+ "largest_action_fraction": 0.45010081206625746,
593
+ "raw_score_min": [
594
+ -9.246599197387695,
595
+ -12.155576705932617,
596
+ -11.755697250366211,
597
+ -13.084254264831543
598
+ ],
599
+ "raw_score_max": [
600
+ 16.499719619750977,
601
+ 14.29150390625,
602
+ 17.30875015258789,
603
+ 15.202037811279297
604
+ ],
605
+ "raw_score_mean": [
606
+ 4.346093761626635,
607
+ -2.074545676786429,
608
+ 1.2331405716554533,
609
+ -2.7278721423409684
610
+ ]
611
+ },
612
+ "checkpoint_sha256": "3fc3190e4151c33d203f69da461d6c56042236200ea9f61369a08c97cb28b88b",
613
+ "training_step": 5000,
614
+ "profile_sha256": "2c8c1f14845e89c6171e1502ff5af6d61d12049e6445d98d63f31e4ffca384f4",
615
+ "hf_revision": "3bbc0bfd7369c818456525bfbdac9580a382c739",
616
+ "hf_path": "latency-aware/atlantis/vla/starvla-qwenoft-h1/atlantis-qwenoft-hist8-profile-s42-v1",
617
+ "wandb_url": "https://wandb.ai/saberrr-zju/starvla_tasks/runs/7ea15469",
618
+ "wandb_readback": {
619
+ "eval_atari_hist8_v1/profile_to_profile/profile_sample/profile_sample/mean_return": 39229
620
+ },
621
+ "status": "verified"
622
+ },
623
+ {
624
+ "task": "atlantis",
625
+ "model": "qwenoft",
626
+ "run_id": "atlantis-qwenoft-hist8-zero-s42-v1",
627
+ "condition": "zero_to_zero",
628
+ "episodes": 100,
629
+ "mean_return": 39560.0,
630
+ "std_return": 2747.2531736263404,
631
+ "mean_raw_frames": 7200.0,
632
+ "true_episode_end_count": 0,
633
+ "frame_cap_count": 100,
634
+ "action_audit": {
635
+ "eval_protocol_version": "hist8-history01-sf-render-v4",
636
+ "history_encoding": "binary_0_1",
637
+ "episodes": 100,
638
+ "episode_seeds": [
639
+ 1000000,
640
+ 1000001,
641
+ 1000002,
642
+ 1000003,
643
+ 1000004,
644
+ 1000005,
645
+ 1000006,
646
+ 1000007,
647
+ 1000008,
648
+ 1000009,
649
+ 1000010,
650
+ 1000011,
651
+ 1000012,
652
+ 1000013,
653
+ 1000014,
654
+ 1000015,
655
+ 1000016,
656
+ 1000017,
657
+ 1000018,
658
+ 1000019,
659
+ 1000020,
660
+ 1000021,
661
+ 1000022,
662
+ 1000023,
663
+ 1000024,
664
+ 1000025,
665
+ 1000026,
666
+ 1000027,
667
+ 1000028,
668
+ 1000029,
669
+ 1000030,
670
+ 1000031,
671
+ 1000032,
672
+ 1000033,
673
+ 1000034,
674
+ 1000035,
675
+ 1000036,
676
+ 1000037,
677
+ 1000038,
678
+ 1000039,
679
+ 1000040,
680
+ 1000041,
681
+ 1000042,
682
+ 1000043,
683
+ 1000044,
684
+ 1000045,
685
+ 1000046,
686
+ 1000047,
687
+ 1000048,
688
+ 1000049,
689
+ 1000050,
690
+ 1000051,
691
+ 1000052,
692
+ 1000053,
693
+ 1000054,
694
+ 1000055,
695
+ 1000056,
696
+ 1000057,
697
+ 1000058,
698
+ 1000059,
699
+ 1000060,
700
+ 1000061,
701
+ 1000062,
702
+ 1000063,
703
+ 1000064,
704
+ 1000065,
705
+ 1000066,
706
+ 1000067,
707
+ 1000068,
708
+ 1000069,
709
+ 1000070,
710
+ 1000071,
711
+ 1000072,
712
+ 1000073,
713
+ 1000074,
714
+ 1000075,
715
+ 1000076,
716
+ 1000077,
717
+ 1000078,
718
+ 1000079,
719
+ 1000080,
720
+ 1000081,
721
+ 1000082,
722
+ 1000083,
723
+ 1000084,
724
+ 1000085,
725
+ 1000086,
726
+ 1000087,
727
+ 1000088,
728
+ 1000089,
729
+ 1000090,
730
+ 1000091,
731
+ 1000092,
732
+ 1000093,
733
+ 1000094,
734
+ 1000095,
735
+ 1000096,
736
+ 1000097,
737
+ 1000098,
738
+ 1000099
739
+ ],
740
+ "issued_actions": 180000,
741
+ "action_counts": {
742
+ "0": 58770,
743
+ "2": 51927,
744
+ "3": 32305,
745
+ "1": 36998
746
+ },
747
+ "largest_action_fraction": 0.3265,
748
+ "raw_score_min": [
749
+ -11.834962844848633,
750
+ -13.451245307922363,
751
+ -15.489179611206055,
752
+ -12.49017333984375
753
+ ],
754
+ "raw_score_max": [
755
+ 20.337247848510742,
756
+ 21.288619995117188,
757
+ 20.40045928955078,
758
+ 20.8565731048584
759
+ ],
760
+ "raw_score_mean": [
761
+ 3.5760303405304956,
762
+ -1.520519410827311,
763
+ 1.385377855150236,
764
+ -1.9132211819803964
765
+ ]
766
+ },
767
+ "checkpoint_sha256": "f91d4079a94f779856220e4a954750624228a3da596b5ecdd70861e888a97877",
768
+ "training_step": 5000,
769
+ "profile_sha256": null,
770
+ "hf_revision": "6491a03f7da933756ee76f8804d1e0b9b1190fc6",
771
+ "hf_path": "zero-latency/atlantis/vla/starvla-qwenoft-h1/atlantis-qwenoft-hist8-zero-s42-v1",
772
+ "wandb_url": "https://wandb.ai/saberrr-zju/starvla_tasks/runs/40be052b",
773
+ "wandb_readback": {
774
+ "eval_atari_hist8_v1/zero_to_zero/fixed/latency_0/mean_return": 39560
775
+ },
776
+ "status": "verified"
777
+ },
778
+ {
779
+ "task": "atlantis",
780
+ "model": "qwenoft",
781
+ "run_id": "atlantis-qwenoft-hist8-zero-s42-v1",
782
+ "condition": "zero_to_profile",
783
+ "episodes": 100,
784
+ "mean_return": 18588.0,
785
+ "std_return": 6544.543987169771,
786
+ "mean_raw_frames": 5017.61,
787
+ "true_episode_end_count": 93,
788
+ "frame_cap_count": 7,
789
+ "action_audit": {
790
+ "eval_protocol_version": "hist8-history01-sf-render-v4",
791
+ "history_encoding": "binary_0_1",
792
+ "episodes": 100,
793
+ "episode_seeds": [
794
+ 1000000,
795
+ 1000001,
796
+ 1000002,
797
+ 1000003,
798
+ 1000004,
799
+ 1000005,
800
+ 1000006,
801
+ 1000007,
802
+ 1000008,
803
+ 1000009,
804
+ 1000010,
805
+ 1000011,
806
+ 1000012,
807
+ 1000013,
808
+ 1000014,
809
+ 1000015,
810
+ 1000016,
811
+ 1000017,
812
+ 1000018,
813
+ 1000019,
814
+ 1000020,
815
+ 1000021,
816
+ 1000022,
817
+ 1000023,
818
+ 1000024,
819
+ 1000025,
820
+ 1000026,
821
+ 1000027,
822
+ 1000028,
823
+ 1000029,
824
+ 1000030,
825
+ 1000031,
826
+ 1000032,
827
+ 1000033,
828
+ 1000034,
829
+ 1000035,
830
+ 1000036,
831
+ 1000037,
832
+ 1000038,
833
+ 1000039,
834
+ 1000040,
835
+ 1000041,
836
+ 1000042,
837
+ 1000043,
838
+ 1000044,
839
+ 1000045,
840
+ 1000046,
841
+ 1000047,
842
+ 1000048,
843
+ 1000049,
844
+ 1000050,
845
+ 1000051,
846
+ 1000052,
847
+ 1000053,
848
+ 1000054,
849
+ 1000055,
850
+ 1000056,
851
+ 1000057,
852
+ 1000058,
853
+ 1000059,
854
+ 1000060,
855
+ 1000061,
856
+ 1000062,
857
+ 1000063,
858
+ 1000064,
859
+ 1000065,
860
+ 1000066,
861
+ 1000067,
862
+ 1000068,
863
+ 1000069,
864
+ 1000070,
865
+ 1000071,
866
+ 1000072,
867
+ 1000073,
868
+ 1000074,
869
+ 1000075,
870
+ 1000076,
871
+ 1000077,
872
+ 1000078,
873
+ 1000079,
874
+ 1000080,
875
+ 1000081,
876
+ 1000082,
877
+ 1000083,
878
+ 1000084,
879
+ 1000085,
880
+ 1000086,
881
+ 1000087,
882
+ 1000088,
883
+ 1000089,
884
+ 1000090,
885
+ 1000091,
886
+ 1000092,
887
+ 1000093,
888
+ 1000094,
889
+ 1000095,
890
+ 1000096,
891
+ 1000097,
892
+ 1000098,
893
+ 1000099
894
+ ],
895
+ "issued_actions": 62677,
896
+ "action_counts": {
897
+ "0": 15821,
898
+ "2": 29659,
899
+ "3": 12666,
900
+ "1": 4531
901
+ },
902
+ "largest_action_fraction": 0.47320388659316814,
903
+ "raw_score_min": [
904
+ -16.4777889251709,
905
+ -17.36240005493164,
906
+ -17.736751556396484,
907
+ -16.52406120300293
908
+ ],
909
+ "raw_score_max": [
910
+ 20.059650421142578,
911
+ 20.963884353637695,
912
+ 20.347089767456055,
913
+ 20.745391845703125
914
+ ],
915
+ "raw_score_mean": [
916
+ 1.5759939295484995,
917
+ -6.00103662508784,
918
+ 6.868458961622308,
919
+ -0.553409212165336
920
+ ]
921
+ },
922
+ "checkpoint_sha256": "f91d4079a94f779856220e4a954750624228a3da596b5ecdd70861e888a97877",
923
+ "training_step": 5000,
924
+ "profile_sha256": "2c8c1f14845e89c6171e1502ff5af6d61d12049e6445d98d63f31e4ffca384f4",
925
+ "hf_revision": "2c66bf8b40bf9dc6fd9b9ea26ccd9fb082f3b883",
926
+ "hf_path": "zero-latency/atlantis/vla/starvla-qwenoft-h1/atlantis-qwenoft-hist8-zero-s42-v1",
927
+ "wandb_url": "https://wandb.ai/saberrr-zju/starvla_tasks/runs/40be052b",
928
+ "wandb_readback": {
929
+ "eval_atari_hist8_v1/zero_to_profile/profile_sample/profile_sample/mean_return": 18588
930
+ },
931
+ "status": "verified"
932
+ },
933
+ {
934
+ "task": "atlantis",
935
+ "model": "qwenpi_v3",
936
+ "run_id": "atlantis-qwenpi_v3-hist8-profile-s42-v1",
937
+ "condition": "profile_to_profile",
938
+ "episodes": 100,
939
+ "mean_return": 26447.0,
940
+ "std_return": 6626.363331420938,
941
+ "mean_raw_frames": 6163.37,
942
+ "true_episode_end_count": 63,
943
+ "frame_cap_count": 37,
944
+ "action_audit": {
945
+ "eval_protocol_version": "hist8-history01-sf-render-v4",
946
+ "history_encoding": "binary_0_1",
947
+ "episodes": 100,
948
+ "episode_seeds": [
949
+ 1000000,
950
+ 1000001,
951
+ 1000002,
952
+ 1000003,
953
+ 1000004,
954
+ 1000005,
955
+ 1000006,
956
+ 1000007,
957
+ 1000008,
958
+ 1000009,
959
+ 1000010,
960
+ 1000011,
961
+ 1000012,
962
+ 1000013,
963
+ 1000014,
964
+ 1000015,
965
+ 1000016,
966
+ 1000017,
967
+ 1000018,
968
+ 1000019,
969
+ 1000020,
970
+ 1000021,
971
+ 1000022,
972
+ 1000023,
973
+ 1000024,
974
+ 1000025,
975
+ 1000026,
976
+ 1000027,
977
+ 1000028,
978
+ 1000029,
979
+ 1000030,
980
+ 1000031,
981
+ 1000032,
982
+ 1000033,
983
+ 1000034,
984
+ 1000035,
985
+ 1000036,
986
+ 1000037,
987
+ 1000038,
988
+ 1000039,
989
+ 1000040,
990
+ 1000041,
991
+ 1000042,
992
+ 1000043,
993
+ 1000044,
994
+ 1000045,
995
+ 1000046,
996
+ 1000047,
997
+ 1000048,
998
+ 1000049,
999
+ 1000050,
1000
+ 1000051,
1001
+ 1000052,
1002
+ 1000053,
1003
+ 1000054,
1004
+ 1000055,
1005
+ 1000056,
1006
+ 1000057,
1007
+ 1000058,
1008
+ 1000059,
1009
+ 1000060,
1010
+ 1000061,
1011
+ 1000062,
1012
+ 1000063,
1013
+ 1000064,
1014
+ 1000065,
1015
+ 1000066,
1016
+ 1000067,
1017
+ 1000068,
1018
+ 1000069,
1019
+ 1000070,
1020
+ 1000071,
1021
+ 1000072,
1022
+ 1000073,
1023
+ 1000074,
1024
+ 1000075,
1025
+ 1000076,
1026
+ 1000077,
1027
+ 1000078,
1028
+ 1000079,
1029
+ 1000080,
1030
+ 1000081,
1031
+ 1000082,
1032
+ 1000083,
1033
+ 1000084,
1034
+ 1000085,
1035
+ 1000086,
1036
+ 1000087,
1037
+ 1000088,
1038
+ 1000089,
1039
+ 1000090,
1040
+ 1000091,
1041
+ 1000092,
1042
+ 1000093,
1043
+ 1000094,
1044
+ 1000095,
1045
+ 1000096,
1046
+ 1000097,
1047
+ 1000098,
1048
+ 1000099
1049
+ ],
1050
+ "issued_actions": 29900,
1051
+ "action_counts": {
1052
+ "0": 14936,
1053
+ "2": 9763,
1054
+ "3": 5191,
1055
+ "1": 10
1056
+ },
1057
+ "largest_action_fraction": 0.49953177257525083,
1058
+ "raw_score_min": [
1059
+ -0.1228821873664856,
1060
+ -0.17654967308044434,
1061
+ -0.14361578226089478,
1062
+ -0.09556090831756592
1063
+ ],
1064
+ "raw_score_max": [
1065
+ 1.120415449142456,
1066
+ 1.0631968975067139,
1067
+ 1.083156704902649,
1068
+ 1.2033812999725342
1069
+ ],
1070
+ "raw_score_mean": [
1071
+ 0.5102518546798447,
1072
+ -0.00530607589835108,
1073
+ 0.32707725962838874,
1074
+ 0.1694935479282984
1075
+ ]
1076
+ },
1077
+ "checkpoint_sha256": "8da09e91d1c6c3908cd1536d6c072e3b6f669b40c82fe9e98705f363691e10ba",
1078
+ "training_step": 5000,
1079
+ "profile_sha256": "72797b4102f37db4cb13f8a64e26b82d9a0757bf5f1ab8c996546e020790831f",
1080
+ "hf_revision": "9502b332f984a867aa753d425d510775a389e0e3",
1081
+ "hf_path": "latency-aware/atlantis/vla/starvla-qwenpi_v3-h1/atlantis-qwenpi_v3-hist8-profile-s42-v1",
1082
+ "wandb_url": "https://wandb.ai/saberrr-zju/starvla_tasks/runs/e58d89ce",
1083
+ "wandb_readback": {
1084
+ "eval_atari_hist8_v1/profile_to_profile/profile_sample/profile_sample/mean_return": 26447
1085
+ },
1086
+ "status": "verified"
1087
+ },
1088
+ {
1089
+ "task": "atlantis",
1090
+ "model": "qwenpi_v3",
1091
+ "run_id": "atlantis-qwenpi_v3-hist8-zero-s42-v1",
1092
+ "condition": "zero_to_zero",
1093
+ "episodes": 100,
1094
+ "mean_return": 39615.0,
1095
+ "std_return": 2879.179570641609,
1096
+ "mean_raw_frames": 7200.0,
1097
+ "true_episode_end_count": 0,
1098
+ "frame_cap_count": 100,
1099
+ "action_audit": {
1100
+ "eval_protocol_version": "hist8-history01-sf-render-v4",
1101
+ "history_encoding": "binary_0_1",
1102
+ "episodes": 100,
1103
+ "episode_seeds": [
1104
+ 1000000,
1105
+ 1000001,
1106
+ 1000002,
1107
+ 1000003,
1108
+ 1000004,
1109
+ 1000005,
1110
+ 1000006,
1111
+ 1000007,
1112
+ 1000008,
1113
+ 1000009,
1114
+ 1000010,
1115
+ 1000011,
1116
+ 1000012,
1117
+ 1000013,
1118
+ 1000014,
1119
+ 1000015,
1120
+ 1000016,
1121
+ 1000017,
1122
+ 1000018,
1123
+ 1000019,
1124
+ 1000020,
1125
+ 1000021,
1126
+ 1000022,
1127
+ 1000023,
1128
+ 1000024,
1129
+ 1000025,
1130
+ 1000026,
1131
+ 1000027,
1132
+ 1000028,
1133
+ 1000029,
1134
+ 1000030,
1135
+ 1000031,
1136
+ 1000032,
1137
+ 1000033,
1138
+ 1000034,
1139
+ 1000035,
1140
+ 1000036,
1141
+ 1000037,
1142
+ 1000038,
1143
+ 1000039,
1144
+ 1000040,
1145
+ 1000041,
1146
+ 1000042,
1147
+ 1000043,
1148
+ 1000044,
1149
+ 1000045,
1150
+ 1000046,
1151
+ 1000047,
1152
+ 1000048,
1153
+ 1000049,
1154
+ 1000050,
1155
+ 1000051,
1156
+ 1000052,
1157
+ 1000053,
1158
+ 1000054,
1159
+ 1000055,
1160
+ 1000056,
1161
+ 1000057,
1162
+ 1000058,
1163
+ 1000059,
1164
+ 1000060,
1165
+ 1000061,
1166
+ 1000062,
1167
+ 1000063,
1168
+ 1000064,
1169
+ 1000065,
1170
+ 1000066,
1171
+ 1000067,
1172
+ 1000068,
1173
+ 1000069,
1174
+ 1000070,
1175
+ 1000071,
1176
+ 1000072,
1177
+ 1000073,
1178
+ 1000074,
1179
+ 1000075,
1180
+ 1000076,
1181
+ 1000077,
1182
+ 1000078,
1183
+ 1000079,
1184
+ 1000080,
1185
+ 1000081,
1186
+ 1000082,
1187
+ 1000083,
1188
+ 1000084,
1189
+ 1000085,
1190
+ 1000086,
1191
+ 1000087,
1192
+ 1000088,
1193
+ 1000089,
1194
+ 1000090,
1195
+ 1000091,
1196
+ 1000092,
1197
+ 1000093,
1198
+ 1000094,
1199
+ 1000095,
1200
+ 1000096,
1201
+ 1000097,
1202
+ 1000098,
1203
+ 1000099
1204
+ ],
1205
+ "issued_actions": 180000,
1206
+ "action_counts": {
1207
+ "0": 58547,
1208
+ "2": 51587,
1209
+ "3": 32275,
1210
+ "1": 37591
1211
+ },
1212
+ "largest_action_fraction": 0.3252611111111111,
1213
+ "raw_score_min": [
1214
+ -0.10931379348039627,
1215
+ -0.09727180004119873,
1216
+ -0.11251753568649292,
1217
+ -0.15724802017211914
1218
+ ],
1219
+ "raw_score_max": [
1220
+ 1.1030423641204834,
1221
+ 1.1520373821258545,
1222
+ 1.1230825185775757,
1223
+ 1.1029820442199707
1224
+ ],
1225
+ "raw_score_mean": [
1226
+ 0.3310999968887113,
1227
+ 0.2096834806125168,
1228
+ 0.2904289227617982,
1229
+ 0.17845669775938278
1230
+ ]
1231
+ },
1232
+ "checkpoint_sha256": "2f4a775690c76dc1b0e4da77a9d45f88df289e17914b895b5a65afa7fc8d214d",
1233
+ "training_step": 5000,
1234
+ "profile_sha256": null,
1235
+ "hf_revision": "aaf650c36c93cb8ff87d3a5ff6ffd635d24c9004",
1236
+ "hf_path": "zero-latency/atlantis/vla/starvla-qwenpi_v3-h1/atlantis-qwenpi_v3-hist8-zero-s42-v1",
1237
+ "wandb_url": "https://wandb.ai/saberrr-zju/starvla_tasks/runs/04d16bbd",
1238
+ "wandb_readback": {
1239
+ "eval_atari_hist8_v1/zero_to_zero/fixed/latency_0/mean_return": 39615
1240
+ },
1241
+ "status": "verified"
1242
+ },
1243
+ {
1244
+ "task": "atlantis",
1245
+ "model": "qwenpi_v3",
1246
+ "run_id": "atlantis-qwenpi_v3-hist8-zero-s42-v1",
1247
+ "condition": "zero_to_profile",
1248
+ "episodes": 100,
1249
+ "mean_return": 4369.0,
1250
+ "std_return": 2132.355270586963,
1251
+ "mean_raw_frames": 2677.98,
1252
+ "true_episode_end_count": 100,
1253
+ "frame_cap_count": 0,
1254
+ "action_audit": {
1255
+ "eval_protocol_version": "hist8-history01-sf-render-v4",
1256
+ "history_encoding": "binary_0_1",
1257
+ "episodes": 100,
1258
+ "episode_seeds": [
1259
+ 1000000,
1260
+ 1000001,
1261
+ 1000002,
1262
+ 1000003,
1263
+ 1000004,
1264
+ 1000005,
1265
+ 1000006,
1266
+ 1000007,
1267
+ 1000008,
1268
+ 1000009,
1269
+ 1000010,
1270
+ 1000011,
1271
+ 1000012,
1272
+ 1000013,
1273
+ 1000014,
1274
+ 1000015,
1275
+ 1000016,
1276
+ 1000017,
1277
+ 1000018,
1278
+ 1000019,
1279
+ 1000020,
1280
+ 1000021,
1281
+ 1000022,
1282
+ 1000023,
1283
+ 1000024,
1284
+ 1000025,
1285
+ 1000026,
1286
+ 1000027,
1287
+ 1000028,
1288
+ 1000029,
1289
+ 1000030,
1290
+ 1000031,
1291
+ 1000032,
1292
+ 1000033,
1293
+ 1000034,
1294
+ 1000035,
1295
+ 1000036,
1296
+ 1000037,
1297
+ 1000038,
1298
+ 1000039,
1299
+ 1000040,
1300
+ 1000041,
1301
+ 1000042,
1302
+ 1000043,
1303
+ 1000044,
1304
+ 1000045,
1305
+ 1000046,
1306
+ 1000047,
1307
+ 1000048,
1308
+ 1000049,
1309
+ 1000050,
1310
+ 1000051,
1311
+ 1000052,
1312
+ 1000053,
1313
+ 1000054,
1314
+ 1000055,
1315
+ 1000056,
1316
+ 1000057,
1317
+ 1000058,
1318
+ 1000059,
1319
+ 1000060,
1320
+ 1000061,
1321
+ 1000062,
1322
+ 1000063,
1323
+ 1000064,
1324
+ 1000065,
1325
+ 1000066,
1326
+ 1000067,
1327
+ 1000068,
1328
+ 1000069,
1329
+ 1000070,
1330
+ 1000071,
1331
+ 1000072,
1332
+ 1000073,
1333
+ 1000074,
1334
+ 1000075,
1335
+ 1000076,
1336
+ 1000077,
1337
+ 1000078,
1338
+ 1000079,
1339
+ 1000080,
1340
+ 1000081,
1341
+ 1000082,
1342
+ 1000083,
1343
+ 1000084,
1344
+ 1000085,
1345
+ 1000086,
1346
+ 1000087,
1347
+ 1000088,
1348
+ 1000089,
1349
+ 1000090,
1350
+ 1000091,
1351
+ 1000092,
1352
+ 1000093,
1353
+ 1000094,
1354
+ 1000095,
1355
+ 1000096,
1356
+ 1000097,
1357
+ 1000098,
1358
+ 1000099
1359
+ ],
1360
+ "issued_actions": 13023,
1361
+ "action_counts": {
1362
+ "0": 8034,
1363
+ "2": 333,
1364
+ "3": 2950,
1365
+ "1": 1706
1366
+ },
1367
+ "largest_action_fraction": 0.6169085464178761,
1368
+ "raw_score_min": [
1369
+ -0.08140607178211212,
1370
+ -0.08260931819677353,
1371
+ -0.09388399124145508,
1372
+ -0.07841962575912476
1373
+ ],
1374
+ "raw_score_max": [
1375
+ 1.0775066614151,
1376
+ 1.0632413625717163,
1377
+ 1.0434043407440186,
1378
+ 1.0697572231292725
1379
+ ],
1380
+ "raw_score_mean": [
1381
+ 0.6259168968651454,
1382
+ 0.13565744175130678,
1383
+ 0.02333082916106862,
1384
+ 0.22537558132693616
1385
+ ]
1386
+ },
1387
+ "checkpoint_sha256": "2f4a775690c76dc1b0e4da77a9d45f88df289e17914b895b5a65afa7fc8d214d",
1388
+ "training_step": 5000,
1389
+ "profile_sha256": "72797b4102f37db4cb13f8a64e26b82d9a0757bf5f1ab8c996546e020790831f",
1390
+ "hf_revision": "576668278cd0a3be955f127d1ec13b026b3c1b13",
1391
+ "hf_path": "zero-latency/atlantis/vla/starvla-qwenpi_v3-h1/atlantis-qwenpi_v3-hist8-zero-s42-v1",
1392
+ "wandb_url": "https://wandb.ai/saberrr-zju/starvla_tasks/runs/04d16bbd",
1393
+ "wandb_readback": {
1394
+ "eval_atari_hist8_v1/zero_to_profile/profile_sample/profile_sample/mean_return": 4369
1395
+ },
1396
+ "status": "verified"
1397
+ },
1398
+ {
1399
+ "task": "asterix",
1400
+ "model": "qwengr00t",
1401
+ "run_id": "asterix-qwengr00t-hist8-profile-s42-v1",
1402
+ "condition": "profile_to_profile",
1403
+ "episodes": 100,
1404
+ "mean_return": 1440.0,
1405
+ "std_return": 686.6221668428715,
1406
+ "mean_raw_frames": 2332.47,
1407
+ "true_episode_end_count": 100,
1408
+ "frame_cap_count": 0,
1409
+ "action_audit": {
1410
+ "eval_protocol_version": "hist8-history01-sf-render-v4",
1411
+ "history_encoding": "binary_0_1",
1412
+ "episodes": 100,
1413
+ "episode_seeds": [
1414
+ 1000000,
1415
+ 1000001,
1416
+ 1000002,
1417
+ 1000003,
1418
+ 1000004,
1419
+ 1000005,
1420
+ 1000006,
1421
+ 1000007,
1422
+ 1000008,
1423
+ 1000009,
1424
+ 1000010,
1425
+ 1000011,
1426
+ 1000012,
1427
+ 1000013,
1428
+ 1000014,
1429
+ 1000015,
1430
+ 1000016,
1431
+ 1000017,
1432
+ 1000018,
1433
+ 1000019,
1434
+ 1000020,
1435
+ 1000021,
1436
+ 1000022,
1437
+ 1000023,
1438
+ 1000024,
1439
+ 1000025,
1440
+ 1000026,
1441
+ 1000027,
1442
+ 1000028,
1443
+ 1000029,
1444
+ 1000030,
1445
+ 1000031,
1446
+ 1000032,
1447
+ 1000033,
1448
+ 1000034,
1449
+ 1000035,
1450
+ 1000036,
1451
+ 1000037,
1452
+ 1000038,
1453
+ 1000039,
1454
+ 1000040,
1455
+ 1000041,
1456
+ 1000042,
1457
+ 1000043,
1458
+ 1000044,
1459
+ 1000045,
1460
+ 1000046,
1461
+ 1000047,
1462
+ 1000048,
1463
+ 1000049,
1464
+ 1000050,
1465
+ 1000051,
1466
+ 1000052,
1467
+ 1000053,
1468
+ 1000054,
1469
+ 1000055,
1470
+ 1000056,
1471
+ 1000057,
1472
+ 1000058,
1473
+ 1000059,
1474
+ 1000060,
1475
+ 1000061,
1476
+ 1000062,
1477
+ 1000063,
1478
+ 1000064,
1479
+ 1000065,
1480
+ 1000066,
1481
+ 1000067,
1482
+ 1000068,
1483
+ 1000069,
1484
+ 1000070,
1485
+ 1000071,
1486
+ 1000072,
1487
+ 1000073,
1488
+ 1000074,
1489
+ 1000075,
1490
+ 1000076,
1491
+ 1000077,
1492
+ 1000078,
1493
+ 1000079,
1494
+ 1000080,
1495
+ 1000081,
1496
+ 1000082,
1497
+ 1000083,
1498
+ 1000084,
1499
+ 1000085,
1500
+ 1000086,
1501
+ 1000087,
1502
+ 1000088,
1503
+ 1000089,
1504
+ 1000090,
1505
+ 1000091,
1506
+ 1000092,
1507
+ 1000093,
1508
+ 1000094,
1509
+ 1000095,
1510
+ 1000096,
1511
+ 1000097,
1512
+ 1000098,
1513
+ 1000099
1514
+ ],
1515
+ "issued_actions": 14615,
1516
+ "action_counts": {
1517
+ "7": 1807,
1518
+ "6": 1173,
1519
+ "5": 9122,
1520
+ "2": 772,
1521
+ "8": 1720,
1522
+ "3": 11,
1523
+ "4": 5,
1524
+ "0": 2,
1525
+ "1": 3
1526
+ },
1527
+ "largest_action_fraction": 0.6241532671912419,
1528
+ "raw_score_min": [
1529
+ -0.1035124659538269,
1530
+ -0.12095272541046143,
1531
+ -0.06751096248626709,
1532
+ -0.0891040563583374,
1533
+ -0.09031903743743896,
1534
+ -0.13926947116851807,
1535
+ -0.07776820659637451,
1536
+ -0.06621575355529785,
1537
+ -0.1278035044670105
1538
+ ],
1539
+ "raw_score_max": [
1540
+ 1.0027379989624023,
1541
+ 1.0079847574234009,
1542
+ 1.0860569477081299,
1543
+ 0.9871726632118225,
1544
+ 0.972202718257904,
1545
+ 1.0391428470611572,
1546
+ 1.1037511825561523,
1547
+ 1.3245031833648682,
1548
+ 1.0730587244033813
1549
+ ],
1550
+ "raw_score_mean": [
1551
+ -0.00413584264603533,
1552
+ 0.0007604830563476222,
1553
+ 0.05498602419388756,
1554
+ 0.0050188393725717605,
1555
+ 0.0030477551167803424,
1556
+ 0.6137112245914245,
1557
+ 0.08402741889386803,
1558
+ 0.12560979057526522,
1559
+ 0.11776857277662967
1560
+ ]
1561
+ },
1562
+ "checkpoint_sha256": "a192db0f4f9707271e453150467a4d833c29479f3e6f6f8031039e84dab8afcf",
1563
+ "training_step": 5000,
1564
+ "profile_sha256": "5e104064d718ac51ce56aefac5318a155fd099f5c714ff8f0bc4985509639fd1",
1565
+ "hf_revision": "6d11a6d4946f278d9b165d260496cd757b330c3d",
1566
+ "hf_path": "latency-aware/asterix/vla/starvla-qwengr00t-h1/asterix-qwengr00t-hist8-profile-s42-v1",
1567
+ "wandb_url": "https://wandb.ai/saberrr-zju/starvla_tasks/runs/946c608a",
1568
+ "wandb_readback": {
1569
+ "eval_atari_hist8_v1/profile_to_profile/profile_sample/profile_sample/mean_return": 1440
1570
+ },
1571
+ "status": "verified"
1572
+ },
1573
+ {
1574
+ "task": "asterix",
1575
+ "model": "qwengr00t",
1576
+ "run_id": "asterix-qwengr00t-hist8-zero-s42-v1",
1577
+ "condition": "zero_to_zero",
1578
+ "episodes": 100,
1579
+ "mean_return": 5714.0,
1580
+ "std_return": 2270.2321467198017,
1581
+ "mean_raw_frames": 5003.95,
1582
+ "true_episode_end_count": 100,
1583
+ "frame_cap_count": 0,
1584
+ "action_audit": {
1585
+ "eval_protocol_version": "hist8-history01-sf-render-v4",
1586
+ "history_encoding": "binary_0_1",
1587
+ "episodes": 100,
1588
+ "episode_seeds": [
1589
+ 1000000,
1590
+ 1000001,
1591
+ 1000002,
1592
+ 1000003,
1593
+ 1000004,
1594
+ 1000005,
1595
+ 1000006,
1596
+ 1000007,
1597
+ 1000008,
1598
+ 1000009,
1599
+ 1000010,
1600
+ 1000011,
1601
+ 1000012,
1602
+ 1000013,
1603
+ 1000014,
1604
+ 1000015,
1605
+ 1000016,
1606
+ 1000017,
1607
+ 1000018,
1608
+ 1000019,
1609
+ 1000020,
1610
+ 1000021,
1611
+ 1000022,
1612
+ 1000023,
1613
+ 1000024,
1614
+ 1000025,
1615
+ 1000026,
1616
+ 1000027,
1617
+ 1000028,
1618
+ 1000029,
1619
+ 1000030,
1620
+ 1000031,
1621
+ 1000032,
1622
+ 1000033,
1623
+ 1000034,
1624
+ 1000035,
1625
+ 1000036,
1626
+ 1000037,
1627
+ 1000038,
1628
+ 1000039,
1629
+ 1000040,
1630
+ 1000041,
1631
+ 1000042,
1632
+ 1000043,
1633
+ 1000044,
1634
+ 1000045,
1635
+ 1000046,
1636
+ 1000047,
1637
+ 1000048,
1638
+ 1000049,
1639
+ 1000050,
1640
+ 1000051,
1641
+ 1000052,
1642
+ 1000053,
1643
+ 1000054,
1644
+ 1000055,
1645
+ 1000056,
1646
+ 1000057,
1647
+ 1000058,
1648
+ 1000059,
1649
+ 1000060,
1650
+ 1000061,
1651
+ 1000062,
1652
+ 1000063,
1653
+ 1000064,
1654
+ 1000065,
1655
+ 1000066,
1656
+ 1000067,
1657
+ 1000068,
1658
+ 1000069,
1659
+ 1000070,
1660
+ 1000071,
1661
+ 1000072,
1662
+ 1000073,
1663
+ 1000074,
1664
+ 1000075,
1665
+ 1000076,
1666
+ 1000077,
1667
+ 1000078,
1668
+ 1000079,
1669
+ 1000080,
1670
+ 1000081,
1671
+ 1000082,
1672
+ 1000083,
1673
+ 1000084,
1674
+ 1000085,
1675
+ 1000086,
1676
+ 1000087,
1677
+ 1000088,
1678
+ 1000089,
1679
+ 1000090,
1680
+ 1000091,
1681
+ 1000092,
1682
+ 1000093,
1683
+ 1000094,
1684
+ 1000095,
1685
+ 1000096,
1686
+ 1000097,
1687
+ 1000098,
1688
+ 1000099
1689
+ ],
1690
+ "issued_actions": 125149,
1691
+ "action_counts": {
1692
+ "6": 41757,
1693
+ "3": 9366,
1694
+ "7": 10131,
1695
+ "4": 2184,
1696
+ "5": 4369,
1697
+ "1": 3331,
1698
+ "8": 29004,
1699
+ "0": 24289,
1700
+ "2": 718
1701
+ },
1702
+ "largest_action_fraction": 0.3336582793310374,
1703
+ "raw_score_min": [
1704
+ -0.2088167667388916,
1705
+ -0.19422024488449097,
1706
+ -0.08199787139892578,
1707
+ -0.4511176347732544,
1708
+ -0.3104654550552368,
1709
+ -0.09569358825683594,
1710
+ -0.2294241189956665,
1711
+ -0.18609082698822021,
1712
+ -0.6733157634735107
1713
+ ],
1714
+ "raw_score_max": [
1715
+ 1.1861666440963745,
1716
+ 1.1156387329101562,
1717
+ 1.1269505023956299,
1718
+ 1.05015230178833,
1719
+ 1.0874724388122559,
1720
+ 1.1275349855422974,
1721
+ 1.0971074104309082,
1722
+ 1.0751999616622925,
1723
+ 1.110088586807251
1724
+ ],
1725
+ "raw_score_mean": [
1726
+ 0.19831424065407252,
1727
+ 0.02671670876697688,
1728
+ 0.0030370978606102527,
1729
+ 0.07622627619930838,
1730
+ 0.011578259722811624,
1731
+ 0.03080988847397458,
1732
+ 0.3282091461042871,
1733
+ 0.08261715072165003,
1734
+ 0.23912502209323958
1735
+ ]
1736
+ },
1737
+ "checkpoint_sha256": "7b3541140bd2d00d39eca98c8788b773e594884a623ae20d7b9e4dd7b45337f5",
1738
+ "training_step": 5000,
1739
+ "profile_sha256": null,
1740
+ "hf_revision": "550ecfbbe35ee7d170d450ff051f297de2097ff8",
1741
+ "hf_path": "zero-latency/asterix/vla/starvla-qwengr00t-h1/asterix-qwengr00t-hist8-zero-s42-v1",
1742
+ "wandb_url": "https://wandb.ai/saberrr-zju/starvla_tasks/runs/42b8e7f3",
1743
+ "wandb_readback": {
1744
+ "eval_atari_hist8_v1/zero_to_zero/fixed/latency_0/mean_return": 5714
1745
+ },
1746
+ "status": "verified"
1747
+ },
1748
+ {
1749
+ "task": "asterix",
1750
+ "model": "qwengr00t",
1751
+ "run_id": "asterix-qwengr00t-hist8-zero-s42-v1",
1752
+ "condition": "zero_to_profile",
1753
+ "episodes": 100,
1754
+ "mean_return": 182.0,
1755
+ "std_return": 125.80143083447024,
1756
+ "mean_raw_frames": 1133.4,
1757
+ "true_episode_end_count": 100,
1758
+ "frame_cap_count": 0,
1759
+ "action_audit": {
1760
+ "eval_protocol_version": "hist8-history01-sf-render-v4",
1761
+ "history_encoding": "binary_0_1",
1762
+ "episodes": 100,
1763
+ "episode_seeds": [
1764
+ 1000000,
1765
+ 1000001,
1766
+ 1000002,
1767
+ 1000003,
1768
+ 1000004,
1769
+ 1000005,
1770
+ 1000006,
1771
+ 1000007,
1772
+ 1000008,
1773
+ 1000009,
1774
+ 1000010,
1775
+ 1000011,
1776
+ 1000012,
1777
+ 1000013,
1778
+ 1000014,
1779
+ 1000015,
1780
+ 1000016,
1781
+ 1000017,
1782
+ 1000018,
1783
+ 1000019,
1784
+ 1000020,
1785
+ 1000021,
1786
+ 1000022,
1787
+ 1000023,
1788
+ 1000024,
1789
+ 1000025,
1790
+ 1000026,
1791
+ 1000027,
1792
+ 1000028,
1793
+ 1000029,
1794
+ 1000030,
1795
+ 1000031,
1796
+ 1000032,
1797
+ 1000033,
1798
+ 1000034,
1799
+ 1000035,
1800
+ 1000036,
1801
+ 1000037,
1802
+ 1000038,
1803
+ 1000039,
1804
+ 1000040,
1805
+ 1000041,
1806
+ 1000042,
1807
+ 1000043,
1808
+ 1000044,
1809
+ 1000045,
1810
+ 1000046,
1811
+ 1000047,
1812
+ 1000048,
1813
+ 1000049,
1814
+ 1000050,
1815
+ 1000051,
1816
+ 1000052,
1817
+ 1000053,
1818
+ 1000054,
1819
+ 1000055,
1820
+ 1000056,
1821
+ 1000057,
1822
+ 1000058,
1823
+ 1000059,
1824
+ 1000060,
1825
+ 1000061,
1826
+ 1000062,
1827
+ 1000063,
1828
+ 1000064,
1829
+ 1000065,
1830
+ 1000066,
1831
+ 1000067,
1832
+ 1000068,
1833
+ 1000069,
1834
+ 1000070,
1835
+ 1000071,
1836
+ 1000072,
1837
+ 1000073,
1838
+ 1000074,
1839
+ 1000075,
1840
+ 1000076,
1841
+ 1000077,
1842
+ 1000078,
1843
+ 1000079,
1844
+ 1000080,
1845
+ 1000081,
1846
+ 1000082,
1847
+ 1000083,
1848
+ 1000084,
1849
+ 1000085,
1850
+ 1000086,
1851
+ 1000087,
1852
+ 1000088,
1853
+ 1000089,
1854
+ 1000090,
1855
+ 1000091,
1856
+ 1000092,
1857
+ 1000093,
1858
+ 1000094,
1859
+ 1000095,
1860
+ 1000096,
1861
+ 1000097,
1862
+ 1000098,
1863
+ 1000099
1864
+ ],
1865
+ "issued_actions": 7130,
1866
+ "action_counts": {
1867
+ "6": 2679,
1868
+ "3": 926,
1869
+ "1": 491,
1870
+ "2": 178,
1871
+ "5": 354,
1872
+ "0": 678,
1873
+ "7": 832,
1874
+ "4": 113,
1875
+ "8": 879
1876
+ },
1877
+ "largest_action_fraction": 0.37573632538569424,
1878
+ "raw_score_min": [
1879
+ -0.11250734329223633,
1880
+ -0.08865809440612793,
1881
+ -0.09899623692035675,
1882
+ -0.17915654182434082,
1883
+ -0.10699084401130676,
1884
+ -0.10701999068260193,
1885
+ -0.14106741547584534,
1886
+ -0.05952806770801544,
1887
+ -0.17616045475006104
1888
+ ],
1889
+ "raw_score_max": [
1890
+ 1.0584603548049927,
1891
+ 1.0680583715438843,
1892
+ 1.0592747926712036,
1893
+ 1.0604641437530518,
1894
+ 1.0169434547424316,
1895
+ 1.0631811618804932,
1896
+ 1.0627062320709229,
1897
+ 1.1027902364730835,
1898
+ 1.0733144283294678
1899
+ ],
1900
+ "raw_score_mean": [
1901
+ 0.08782884661199036,
1902
+ 0.06818614123342755,
1903
+ 0.020853550182093786,
1904
+ 0.12822516662689765,
1905
+ 0.020668613626249395,
1906
+ 0.05096994273499736,
1907
+ 0.3689880494977286,
1908
+ 0.12228238381834657,
1909
+ 0.12859012346830986
1910
+ ]
1911
+ },
1912
+ "checkpoint_sha256": "7b3541140bd2d00d39eca98c8788b773e594884a623ae20d7b9e4dd7b45337f5",
1913
+ "training_step": 5000,
1914
+ "profile_sha256": "5e104064d718ac51ce56aefac5318a155fd099f5c714ff8f0bc4985509639fd1",
1915
+ "hf_revision": "22e43559d63f898ca1b9bc0418b65a79efcaffd5",
1916
+ "hf_path": "zero-latency/asterix/vla/starvla-qwengr00t-h1/asterix-qwengr00t-hist8-zero-s42-v1",
1917
+ "wandb_url": "https://wandb.ai/saberrr-zju/starvla_tasks/runs/42b8e7f3",
1918
+ "wandb_readback": {
1919
+ "eval_atari_hist8_v1/zero_to_profile/profile_sample/profile_sample/mean_return": 182
1920
+ },
1921
+ "status": "verified"
1922
+ },
1923
+ {
1924
+ "task": "asterix",
1925
+ "model": "qwenoft",
1926
+ "run_id": "asterix-qwenoft-hist8-profile-s42-v1",
1927
+ "condition": "profile_to_profile",
1928
+ "episodes": 100,
1929
+ "mean_return": 3653.5,
1930
+ "std_return": 1620.821628063989,
1931
+ "mean_raw_frames": 3952.55,
1932
+ "true_episode_end_count": 99,
1933
+ "frame_cap_count": 1,
1934
+ "action_audit": {
1935
+ "eval_protocol_version": "hist8-history01-sf-render-v4",
1936
+ "history_encoding": "binary_0_1",
1937
+ "episodes": 100,
1938
+ "episode_seeds": [
1939
+ 1000000,
1940
+ 1000001,
1941
+ 1000002,
1942
+ 1000003,
1943
+ 1000004,
1944
+ 1000005,
1945
+ 1000006,
1946
+ 1000007,
1947
+ 1000008,
1948
+ 1000009,
1949
+ 1000010,
1950
+ 1000011,
1951
+ 1000012,
1952
+ 1000013,
1953
+ 1000014,
1954
+ 1000015,
1955
+ 1000016,
1956
+ 1000017,
1957
+ 1000018,
1958
+ 1000019,
1959
+ 1000020,
1960
+ 1000021,
1961
+ 1000022,
1962
+ 1000023,
1963
+ 1000024,
1964
+ 1000025,
1965
+ 1000026,
1966
+ 1000027,
1967
+ 1000028,
1968
+ 1000029,
1969
+ 1000030,
1970
+ 1000031,
1971
+ 1000032,
1972
+ 1000033,
1973
+ 1000034,
1974
+ 1000035,
1975
+ 1000036,
1976
+ 1000037,
1977
+ 1000038,
1978
+ 1000039,
1979
+ 1000040,
1980
+ 1000041,
1981
+ 1000042,
1982
+ 1000043,
1983
+ 1000044,
1984
+ 1000045,
1985
+ 1000046,
1986
+ 1000047,
1987
+ 1000048,
1988
+ 1000049,
1989
+ 1000050,
1990
+ 1000051,
1991
+ 1000052,
1992
+ 1000053,
1993
+ 1000054,
1994
+ 1000055,
1995
+ 1000056,
1996
+ 1000057,
1997
+ 1000058,
1998
+ 1000059,
1999
+ 1000060,
2000
+ 1000061,
2001
+ 1000062,
2002
+ 1000063,
2003
+ 1000064,
2004
+ 1000065,
2005
+ 1000066,
2006
+ 1000067,
2007
+ 1000068,
2008
+ 1000069,
2009
+ 1000070,
2010
+ 1000071,
2011
+ 1000072,
2012
+ 1000073,
2013
+ 1000074,
2014
+ 1000075,
2015
+ 1000076,
2016
+ 1000077,
2017
+ 1000078,
2018
+ 1000079,
2019
+ 1000080,
2020
+ 1000081,
2021
+ 1000082,
2022
+ 1000083,
2023
+ 1000084,
2024
+ 1000085,
2025
+ 1000086,
2026
+ 1000087,
2027
+ 1000088,
2028
+ 1000089,
2029
+ 1000090,
2030
+ 1000091,
2031
+ 1000092,
2032
+ 1000093,
2033
+ 1000094,
2034
+ 1000095,
2035
+ 1000096,
2036
+ 1000097,
2037
+ 1000098,
2038
+ 1000099
2039
+ ],
2040
+ "issued_actions": 49405,
2041
+ "action_counts": {
2042
+ "2": 3308,
2043
+ "8": 4664,
2044
+ "1": 4419,
2045
+ "5": 7523,
2046
+ "0": 21887,
2047
+ "7": 5753,
2048
+ "6": 1626,
2049
+ "4": 211,
2050
+ "3": 14
2051
+ },
2052
+ "largest_action_fraction": 0.4430118409067908,
2053
+ "raw_score_min": [
2054
+ -13.119811058044434,
2055
+ -10.853740692138672,
2056
+ -11.006418228149414,
2057
+ -12.597626686096191,
2058
+ -9.706649780273438,
2059
+ -8.856327056884766,
2060
+ -9.544286727905273,
2061
+ -7.748889446258545,
2062
+ -13.909894943237305
2063
+ ],
2064
+ "raw_score_max": [
2065
+ 21.067642211914062,
2066
+ 21.027864456176758,
2067
+ 20.085676193237305,
2068
+ 17.18351173400879,
2069
+ 15.032642364501953,
2070
+ 21.076366424560547,
2071
+ 18.5609130859375,
2072
+ 19.459716796875,
2073
+ 21.263235092163086
2074
+ ],
2075
+ "raw_score_mean": [
2076
+ 6.1028937123558284,
2077
+ 0.0709031725423172,
2078
+ 0.07752298349673893,
2079
+ -5.34277075081747,
2080
+ -2.9248927316361186,
2081
+ 2.2591020215939106,
2082
+ -0.04072495360855251,
2083
+ 2.744448609118694,
2084
+ -3.116156193460523
2085
+ ]
2086
+ },
2087
+ "checkpoint_sha256": "b96bf1398d3d767b13be1cc76eefdfe70fdab349231ef5a4636495c548fe09ac",
2088
+ "training_step": 5000,
2089
+ "profile_sha256": "4e2e3ed923c7500248ce4ce1508f124faec78282d90d6b99d070a9e29bdb6146",
2090
+ "hf_revision": "a4dd7774b90faf83cfec251a3a23249e3bbac080",
2091
+ "hf_path": "latency-aware/asterix/vla/starvla-qwenoft-h1/asterix-qwenoft-hist8-profile-s42-v1",
2092
+ "wandb_url": "https://wandb.ai/saberrr-zju/starvla_tasks/runs/ac01280d",
2093
+ "wandb_readback": {
2094
+ "eval_atari_hist8_v1/profile_to_profile/profile_sample/profile_sample/mean_return": 3653.5
2095
+ },
2096
+ "status": "verified"
2097
+ },
2098
+ {
2099
+ "task": "asterix",
2100
+ "model": "qwenoft",
2101
+ "run_id": "asterix-qwenoft-hist8-zero-s42-v1",
2102
+ "condition": "zero_to_zero",
2103
+ "episodes": 100,
2104
+ "mean_return": 5803.5,
2105
+ "std_return": 2263.15327585208,
2106
+ "mean_raw_frames": 5061.14,
2107
+ "true_episode_end_count": 100,
2108
+ "frame_cap_count": 0,
2109
+ "action_audit": {
2110
+ "eval_protocol_version": "hist8-history01-sf-render-v4",
2111
+ "history_encoding": "binary_0_1",
2112
+ "episodes": 100,
2113
+ "episode_seeds": [
2114
+ 1000000,
2115
+ 1000001,
2116
+ 1000002,
2117
+ 1000003,
2118
+ 1000004,
2119
+ 1000005,
2120
+ 1000006,
2121
+ 1000007,
2122
+ 1000008,
2123
+ 1000009,
2124
+ 1000010,
2125
+ 1000011,
2126
+ 1000012,
2127
+ 1000013,
2128
+ 1000014,
2129
+ 1000015,
2130
+ 1000016,
2131
+ 1000017,
2132
+ 1000018,
2133
+ 1000019,
2134
+ 1000020,
2135
+ 1000021,
2136
+ 1000022,
2137
+ 1000023,
2138
+ 1000024,
2139
+ 1000025,
2140
+ 1000026,
2141
+ 1000027,
2142
+ 1000028,
2143
+ 1000029,
2144
+ 1000030,
2145
+ 1000031,
2146
+ 1000032,
2147
+ 1000033,
2148
+ 1000034,
2149
+ 1000035,
2150
+ 1000036,
2151
+ 1000037,
2152
+ 1000038,
2153
+ 1000039,
2154
+ 1000040,
2155
+ 1000041,
2156
+ 1000042,
2157
+ 1000043,
2158
+ 1000044,
2159
+ 1000045,
2160
+ 1000046,
2161
+ 1000047,
2162
+ 1000048,
2163
+ 1000049,
2164
+ 1000050,
2165
+ 1000051,
2166
+ 1000052,
2167
+ 1000053,
2168
+ 1000054,
2169
+ 1000055,
2170
+ 1000056,
2171
+ 1000057,
2172
+ 1000058,
2173
+ 1000059,
2174
+ 1000060,
2175
+ 1000061,
2176
+ 1000062,
2177
+ 1000063,
2178
+ 1000064,
2179
+ 1000065,
2180
+ 1000066,
2181
+ 1000067,
2182
+ 1000068,
2183
+ 1000069,
2184
+ 1000070,
2185
+ 1000071,
2186
+ 1000072,
2187
+ 1000073,
2188
+ 1000074,
2189
+ 1000075,
2190
+ 1000076,
2191
+ 1000077,
2192
+ 1000078,
2193
+ 1000079,
2194
+ 1000080,
2195
+ 1000081,
2196
+ 1000082,
2197
+ 1000083,
2198
+ 1000084,
2199
+ 1000085,
2200
+ 1000086,
2201
+ 1000087,
2202
+ 1000088,
2203
+ 1000089,
2204
+ 1000090,
2205
+ 1000091,
2206
+ 1000092,
2207
+ 1000093,
2208
+ 1000094,
2209
+ 1000095,
2210
+ 1000096,
2211
+ 1000097,
2212
+ 1000098,
2213
+ 1000099
2214
+ ],
2215
+ "issued_actions": 126579,
2216
+ "action_counts": {
2217
+ "6": 40546,
2218
+ "3": 9448,
2219
+ "7": 10356,
2220
+ "4": 2260,
2221
+ "5": 4503,
2222
+ "1": 3315,
2223
+ "8": 29750,
2224
+ "0": 25704,
2225
+ "2": 697
2226
+ },
2227
+ "largest_action_fraction": 0.320321696331935,
2228
+ "raw_score_min": [
2229
+ -11.378705978393555,
2230
+ -11.544252395629883,
2231
+ -11.610601425170898,
2232
+ -9.41488265991211,
2233
+ -11.919272422790527,
2234
+ -10.242284774780273,
2235
+ -10.542601585388184,
2236
+ -14.898965835571289,
2237
+ -9.239874839782715
2238
+ ],
2239
+ "raw_score_max": [
2240
+ 25.130844116210938,
2241
+ 24.600074768066406,
2242
+ 27.344585418701172,
2243
+ 24.893482208251953,
2244
+ 29.90880584716797,
2245
+ 24.530048370361328,
2246
+ 20.722490310668945,
2247
+ 25.345619201660156,
2248
+ 22.091463088989258
2249
+ ],
2250
+ "raw_score_mean": [
2251
+ 0.7696608880708099,
2252
+ -4.3064215144101,
2253
+ -4.015774457300812,
2254
+ -1.7123453344700852,
2255
+ -4.07732841685631,
2256
+ -2.8271864798453827,
2257
+ 4.978411157524741,
2258
+ -2.3697603980762096,
2259
+ 2.684130390830407
2260
+ ]
2261
+ },
2262
+ "checkpoint_sha256": "9d5be1187e768a3ba6638610ad70ca58fed54a531b77b0b0dea3a33538def572",
2263
+ "training_step": 5000,
2264
+ "profile_sha256": null,
2265
+ "hf_revision": "6c2cbeb0830a5ce6c8ad7a0c2226c72dce420aca",
2266
+ "hf_path": "zero-latency/asterix/vla/starvla-qwenoft-h1/asterix-qwenoft-hist8-zero-s42-v1",
2267
+ "wandb_url": "https://wandb.ai/saberrr-zju/starvla_tasks/runs/aa718e0f",
2268
+ "wandb_readback": {
2269
+ "eval_atari_hist8_v1/zero_to_zero/fixed/latency_0/mean_return": 5803.5
2270
+ },
2271
+ "status": "verified"
2272
+ },
2273
+ {
2274
+ "task": "asterix",
2275
+ "model": "qwenoft",
2276
+ "run_id": "asterix-qwenoft-hist8-zero-s42-v1",
2277
+ "condition": "zero_to_profile",
2278
+ "episodes": 100,
2279
+ "mean_return": 374.5,
2280
+ "std_return": 188.21463811297994,
2281
+ "mean_raw_frames": 1248.94,
2282
+ "true_episode_end_count": 100,
2283
+ "frame_cap_count": 0,
2284
+ "action_audit": {
2285
+ "eval_protocol_version": "hist8-history01-sf-render-v4",
2286
+ "history_encoding": "binary_0_1",
2287
+ "episodes": 100,
2288
+ "episode_seeds": [
2289
+ 1000000,
2290
+ 1000001,
2291
+ 1000002,
2292
+ 1000003,
2293
+ 1000004,
2294
+ 1000005,
2295
+ 1000006,
2296
+ 1000007,
2297
+ 1000008,
2298
+ 1000009,
2299
+ 1000010,
2300
+ 1000011,
2301
+ 1000012,
2302
+ 1000013,
2303
+ 1000014,
2304
+ 1000015,
2305
+ 1000016,
2306
+ 1000017,
2307
+ 1000018,
2308
+ 1000019,
2309
+ 1000020,
2310
+ 1000021,
2311
+ 1000022,
2312
+ 1000023,
2313
+ 1000024,
2314
+ 1000025,
2315
+ 1000026,
2316
+ 1000027,
2317
+ 1000028,
2318
+ 1000029,
2319
+ 1000030,
2320
+ 1000031,
2321
+ 1000032,
2322
+ 1000033,
2323
+ 1000034,
2324
+ 1000035,
2325
+ 1000036,
2326
+ 1000037,
2327
+ 1000038,
2328
+ 1000039,
2329
+ 1000040,
2330
+ 1000041,
2331
+ 1000042,
2332
+ 1000043,
2333
+ 1000044,
2334
+ 1000045,
2335
+ 1000046,
2336
+ 1000047,
2337
+ 1000048,
2338
+ 1000049,
2339
+ 1000050,
2340
+ 1000051,
2341
+ 1000052,
2342
+ 1000053,
2343
+ 1000054,
2344
+ 1000055,
2345
+ 1000056,
2346
+ 1000057,
2347
+ 1000058,
2348
+ 1000059,
2349
+ 1000060,
2350
+ 1000061,
2351
+ 1000062,
2352
+ 1000063,
2353
+ 1000064,
2354
+ 1000065,
2355
+ 1000066,
2356
+ 1000067,
2357
+ 1000068,
2358
+ 1000069,
2359
+ 1000070,
2360
+ 1000071,
2361
+ 1000072,
2362
+ 1000073,
2363
+ 1000074,
2364
+ 1000075,
2365
+ 1000076,
2366
+ 1000077,
2367
+ 1000078,
2368
+ 1000079,
2369
+ 1000080,
2370
+ 1000081,
2371
+ 1000082,
2372
+ 1000083,
2373
+ 1000084,
2374
+ 1000085,
2375
+ 1000086,
2376
+ 1000087,
2377
+ 1000088,
2378
+ 1000089,
2379
+ 1000090,
2380
+ 1000091,
2381
+ 1000092,
2382
+ 1000093,
2383
+ 1000094,
2384
+ 1000095,
2385
+ 1000096,
2386
+ 1000097,
2387
+ 1000098,
2388
+ 1000099
2389
+ ],
2390
+ "issued_actions": 15645,
2391
+ "action_counts": {
2392
+ "6": 4435,
2393
+ "3": 936,
2394
+ "4": 505,
2395
+ "7": 1122,
2396
+ "1": 860,
2397
+ "5": 142,
2398
+ "8": 5723,
2399
+ "0": 1834,
2400
+ "2": 88
2401
+ },
2402
+ "largest_action_fraction": 0.3658037711728987,
2403
+ "raw_score_min": [
2404
+ -13.79120922088623,
2405
+ -13.78160572052002,
2406
+ -12.283455848693848,
2407
+ -11.349635124206543,
2408
+ -12.58639907836914,
2409
+ -12.535384178161621,
2410
+ -11.73652458190918,
2411
+ -13.926600456237793,
2412
+ -10.862908363342285
2413
+ ],
2414
+ "raw_score_max": [
2415
+ 24.001331329345703,
2416
+ 23.698068618774414,
2417
+ 25.557151794433594,
2418
+ 24.67331314086914,
2419
+ 27.569021224975586,
2420
+ 19.34308624267578,
2421
+ 19.891719818115234,
2422
+ 24.10886001586914,
2423
+ 21.663095474243164
2424
+ ],
2425
+ "raw_score_mean": [
2426
+ -0.891834141335698,
2427
+ -3.7993504673135416,
2428
+ -4.722685660356129,
2429
+ -2.288652651788433,
2430
+ -3.795079822646863,
2431
+ -4.579512725203524,
2432
+ 4.512334494623982,
2433
+ -2.4837288779462248,
2434
+ 5.881643920127016
2435
+ ]
2436
+ },
2437
+ "checkpoint_sha256": "9d5be1187e768a3ba6638610ad70ca58fed54a531b77b0b0dea3a33538def572",
2438
+ "training_step": 5000,
2439
+ "profile_sha256": "4e2e3ed923c7500248ce4ce1508f124faec78282d90d6b99d070a9e29bdb6146",
2440
+ "hf_revision": "d445334970ab7e110c3cacb5c8b790d3df20e1ec",
2441
+ "hf_path": "zero-latency/asterix/vla/starvla-qwenoft-h1/asterix-qwenoft-hist8-zero-s42-v1",
2442
+ "wandb_url": "https://wandb.ai/saberrr-zju/starvla_tasks/runs/aa718e0f",
2443
+ "wandb_readback": {
2444
+ "eval_atari_hist8_v1/zero_to_profile/profile_sample/profile_sample/mean_return": 374.5
2445
+ },
2446
+ "status": "verified"
2447
+ },
2448
+ {
2449
+ "task": "asterix",
2450
+ "model": "qwenpi_v3",
2451
+ "run_id": "asterix-qwenpi_v3-hist8-profile-s42-v1",
2452
+ "condition": "profile_to_profile",
2453
+ "episodes": 100,
2454
+ "mean_return": 1015.0,
2455
+ "std_return": 478.4610746967824,
2456
+ "mean_raw_frames": 1732.77,
2457
+ "true_episode_end_count": 100,
2458
+ "frame_cap_count": 0,
2459
+ "action_audit": {
2460
+ "eval_protocol_version": "hist8-history01-sf-render-v4",
2461
+ "history_encoding": "binary_0_1",
2462
+ "episodes": 100,
2463
+ "episode_seeds": [
2464
+ 1000000,
2465
+ 1000001,
2466
+ 1000002,
2467
+ 1000003,
2468
+ 1000004,
2469
+ 1000005,
2470
+ 1000006,
2471
+ 1000007,
2472
+ 1000008,
2473
+ 1000009,
2474
+ 1000010,
2475
+ 1000011,
2476
+ 1000012,
2477
+ 1000013,
2478
+ 1000014,
2479
+ 1000015,
2480
+ 1000016,
2481
+ 1000017,
2482
+ 1000018,
2483
+ 1000019,
2484
+ 1000020,
2485
+ 1000021,
2486
+ 1000022,
2487
+ 1000023,
2488
+ 1000024,
2489
+ 1000025,
2490
+ 1000026,
2491
+ 1000027,
2492
+ 1000028,
2493
+ 1000029,
2494
+ 1000030,
2495
+ 1000031,
2496
+ 1000032,
2497
+ 1000033,
2498
+ 1000034,
2499
+ 1000035,
2500
+ 1000036,
2501
+ 1000037,
2502
+ 1000038,
2503
+ 1000039,
2504
+ 1000040,
2505
+ 1000041,
2506
+ 1000042,
2507
+ 1000043,
2508
+ 1000044,
2509
+ 1000045,
2510
+ 1000046,
2511
+ 1000047,
2512
+ 1000048,
2513
+ 1000049,
2514
+ 1000050,
2515
+ 1000051,
2516
+ 1000052,
2517
+ 1000053,
2518
+ 1000054,
2519
+ 1000055,
2520
+ 1000056,
2521
+ 1000057,
2522
+ 1000058,
2523
+ 1000059,
2524
+ 1000060,
2525
+ 1000061,
2526
+ 1000062,
2527
+ 1000063,
2528
+ 1000064,
2529
+ 1000065,
2530
+ 1000066,
2531
+ 1000067,
2532
+ 1000068,
2533
+ 1000069,
2534
+ 1000070,
2535
+ 1000071,
2536
+ 1000072,
2537
+ 1000073,
2538
+ 1000074,
2539
+ 1000075,
2540
+ 1000076,
2541
+ 1000077,
2542
+ 1000078,
2543
+ 1000079,
2544
+ 1000080,
2545
+ 1000081,
2546
+ 1000082,
2547
+ 1000083,
2548
+ 1000084,
2549
+ 1000085,
2550
+ 1000086,
2551
+ 1000087,
2552
+ 1000088,
2553
+ 1000089,
2554
+ 1000090,
2555
+ 1000091,
2556
+ 1000092,
2557
+ 1000093,
2558
+ 1000094,
2559
+ 1000095,
2560
+ 1000096,
2561
+ 1000097,
2562
+ 1000098,
2563
+ 1000099
2564
+ ],
2565
+ "issued_actions": 8392,
2566
+ "action_counts": {
2567
+ "6": 1516,
2568
+ "7": 5378,
2569
+ "4": 53,
2570
+ "5": 756,
2571
+ "2": 682,
2572
+ "1": 7
2573
+ },
2574
+ "largest_action_fraction": 0.6408484270734033,
2575
+ "raw_score_min": [
2576
+ -0.10672998428344727,
2577
+ -0.07665398716926575,
2578
+ -0.08862924575805664,
2579
+ -0.0866539478302002,
2580
+ -0.07203245162963867,
2581
+ -0.09199362993240356,
2582
+ -0.14521658420562744,
2583
+ -0.10043677687644958,
2584
+ -0.09294778108596802
2585
+ ],
2586
+ "raw_score_max": [
2587
+ 0.21051234006881714,
2588
+ 1.0479683876037598,
2589
+ 1.0772048234939575,
2590
+ 0.34700167179107666,
2591
+ 1.0611048936843872,
2592
+ 1.088529348373413,
2593
+ 1.12296724319458,
2594
+ 1.1893627643585205,
2595
+ 0.273490309715271
2596
+ ],
2597
+ "raw_score_mean": [
2598
+ -0.01854851725022952,
2599
+ 0.00018712080527031287,
2600
+ 0.08459583689335816,
2601
+ -0.00537758165290375,
2602
+ 0.0002778814795098087,
2603
+ 0.08627376778564416,
2604
+ 0.17766348345399288,
2605
+ 0.6637925541142021,
2606
+ -0.0027456874388469394
2607
+ ]
2608
+ },
2609
+ "checkpoint_sha256": "19fc503b7ab9158c445f92258e489da35d2a73a87c12e3e7954a7e0f51ea75eb",
2610
+ "training_step": 5000,
2611
+ "profile_sha256": "a5acfbb259cc3431bcce9dedef862a1e3895bf81659aa3c1707fa54c277852fc",
2612
+ "hf_revision": "b53f7bbd530c986912433448d7dc52952349bd99",
2613
+ "hf_path": "latency-aware/asterix/vla/starvla-qwenpi_v3-h1/asterix-qwenpi_v3-hist8-profile-s42-v1",
2614
+ "wandb_url": "https://wandb.ai/saberrr-zju/starvla_tasks/runs/f9d438a5",
2615
+ "wandb_readback": {
2616
+ "eval_atari_hist8_v1/profile_to_profile/profile_sample/profile_sample/mean_return": 1015
2617
+ },
2618
+ "status": "verified"
2619
+ },
2620
+ {
2621
+ "task": "asterix",
2622
+ "model": "qwenpi_v3",
2623
+ "run_id": "asterix-qwenpi_v3-hist8-zero-s42-v1",
2624
+ "condition": "zero_to_zero",
2625
+ "episodes": 100,
2626
+ "mean_return": 5758.5,
2627
+ "std_return": 2132.5812411254115,
2628
+ "mean_raw_frames": 5096.1,
2629
+ "true_episode_end_count": 97,
2630
+ "frame_cap_count": 3,
2631
+ "action_audit": {
2632
+ "eval_protocol_version": "hist8-history01-sf-render-v4",
2633
+ "history_encoding": "binary_0_1",
2634
+ "episodes": 100,
2635
+ "episode_seeds": [
2636
+ 1000000,
2637
+ 1000001,
2638
+ 1000002,
2639
+ 1000003,
2640
+ 1000004,
2641
+ 1000005,
2642
+ 1000006,
2643
+ 1000007,
2644
+ 1000008,
2645
+ 1000009,
2646
+ 1000010,
2647
+ 1000011,
2648
+ 1000012,
2649
+ 1000013,
2650
+ 1000014,
2651
+ 1000015,
2652
+ 1000016,
2653
+ 1000017,
2654
+ 1000018,
2655
+ 1000019,
2656
+ 1000020,
2657
+ 1000021,
2658
+ 1000022,
2659
+ 1000023,
2660
+ 1000024,
2661
+ 1000025,
2662
+ 1000026,
2663
+ 1000027,
2664
+ 1000028,
2665
+ 1000029,
2666
+ 1000030,
2667
+ 1000031,
2668
+ 1000032,
2669
+ 1000033,
2670
+ 1000034,
2671
+ 1000035,
2672
+ 1000036,
2673
+ 1000037,
2674
+ 1000038,
2675
+ 1000039,
2676
+ 1000040,
2677
+ 1000041,
2678
+ 1000042,
2679
+ 1000043,
2680
+ 1000044,
2681
+ 1000045,
2682
+ 1000046,
2683
+ 1000047,
2684
+ 1000048,
2685
+ 1000049,
2686
+ 1000050,
2687
+ 1000051,
2688
+ 1000052,
2689
+ 1000053,
2690
+ 1000054,
2691
+ 1000055,
2692
+ 1000056,
2693
+ 1000057,
2694
+ 1000058,
2695
+ 1000059,
2696
+ 1000060,
2697
+ 1000061,
2698
+ 1000062,
2699
+ 1000063,
2700
+ 1000064,
2701
+ 1000065,
2702
+ 1000066,
2703
+ 1000067,
2704
+ 1000068,
2705
+ 1000069,
2706
+ 1000070,
2707
+ 1000071,
2708
+ 1000072,
2709
+ 1000073,
2710
+ 1000074,
2711
+ 1000075,
2712
+ 1000076,
2713
+ 1000077,
2714
+ 1000078,
2715
+ 1000079,
2716
+ 1000080,
2717
+ 1000081,
2718
+ 1000082,
2719
+ 1000083,
2720
+ 1000084,
2721
+ 1000085,
2722
+ 1000086,
2723
+ 1000087,
2724
+ 1000088,
2725
+ 1000089,
2726
+ 1000090,
2727
+ 1000091,
2728
+ 1000092,
2729
+ 1000093,
2730
+ 1000094,
2731
+ 1000095,
2732
+ 1000096,
2733
+ 1000097,
2734
+ 1000098,
2735
+ 1000099
2736
+ ],
2737
+ "issued_actions": 127448,
2738
+ "action_counts": {
2739
+ "6": 41388,
2740
+ "3": 9645,
2741
+ "7": 10390,
2742
+ "4": 2204,
2743
+ "5": 4494,
2744
+ "1": 3244,
2745
+ "8": 30040,
2746
+ "0": 25337,
2747
+ "2": 706
2748
+ },
2749
+ "largest_action_fraction": 0.32474420940305065,
2750
+ "raw_score_min": [
2751
+ -0.08517017960548401,
2752
+ -0.10428443551063538,
2753
+ -0.08317434787750244,
2754
+ -0.09850794076919556,
2755
+ -0.09002876281738281,
2756
+ -0.07311639189720154,
2757
+ -0.26507842540740967,
2758
+ -0.08630828559398651,
2759
+ -0.16656984388828278
2760
+ ],
2761
+ "raw_score_max": [
2762
+ 1.112355351448059,
2763
+ 1.2103047370910645,
2764
+ 1.1933003664016724,
2765
+ 1.1528035402297974,
2766
+ 1.1515668630599976,
2767
+ 1.1958045959472656,
2768
+ 1.1107509136199951,
2769
+ 1.2056286334991455,
2770
+ 1.1282782554626465
2771
+ ],
2772
+ "raw_score_mean": [
2773
+ 0.19860854102436143,
2774
+ 0.022158051565253748,
2775
+ 0.00810792909272446,
2776
+ 0.06813755351500361,
2777
+ 0.014625414859000431,
2778
+ 0.03114063633102587,
2779
+ 0.3257789390649875,
2780
+ 0.0847539822291615,
2781
+ 0.23716105913420343
2782
+ ]
2783
+ },
2784
+ "checkpoint_sha256": "381d1ed73a6f3949cb1880d6f1feb18db13068aeb6db39a7974ae9fb1fba0f56",
2785
+ "training_step": 5000,
2786
+ "profile_sha256": null,
2787
+ "hf_revision": "bacdab2bdeb46480955b4ec03d1f097162928c6d",
2788
+ "hf_path": "zero-latency/asterix/vla/starvla-qwenpi_v3-h1/asterix-qwenpi_v3-hist8-zero-s42-v1",
2789
+ "wandb_url": "https://wandb.ai/saberrr-zju/starvla_tasks/runs/ec20df07",
2790
+ "wandb_readback": {
2791
+ "eval_atari_hist8_v1/zero_to_zero/fixed/latency_0/mean_return": 5758.5
2792
+ },
2793
+ "status": "verified"
2794
+ },
2795
+ {
2796
+ "task": "asterix",
2797
+ "model": "qwenpi_v3",
2798
+ "run_id": "asterix-qwenpi_v3-hist8-zero-s42-v1",
2799
+ "condition": "zero_to_profile",
2800
+ "episodes": 100,
2801
+ "mean_return": 143.5,
2802
+ "std_return": 138.50180504238924,
2803
+ "mean_raw_frames": 1064.13,
2804
+ "true_episode_end_count": 100,
2805
+ "frame_cap_count": 0,
2806
+ "action_audit": {
2807
+ "eval_protocol_version": "hist8-history01-sf-render-v4",
2808
+ "history_encoding": "binary_0_1",
2809
+ "episodes": 100,
2810
+ "episode_seeds": [
2811
+ 1000000,
2812
+ 1000001,
2813
+ 1000002,
2814
+ 1000003,
2815
+ 1000004,
2816
+ 1000005,
2817
+ 1000006,
2818
+ 1000007,
2819
+ 1000008,
2820
+ 1000009,
2821
+ 1000010,
2822
+ 1000011,
2823
+ 1000012,
2824
+ 1000013,
2825
+ 1000014,
2826
+ 1000015,
2827
+ 1000016,
2828
+ 1000017,
2829
+ 1000018,
2830
+ 1000019,
2831
+ 1000020,
2832
+ 1000021,
2833
+ 1000022,
2834
+ 1000023,
2835
+ 1000024,
2836
+ 1000025,
2837
+ 1000026,
2838
+ 1000027,
2839
+ 1000028,
2840
+ 1000029,
2841
+ 1000030,
2842
+ 1000031,
2843
+ 1000032,
2844
+ 1000033,
2845
+ 1000034,
2846
+ 1000035,
2847
+ 1000036,
2848
+ 1000037,
2849
+ 1000038,
2850
+ 1000039,
2851
+ 1000040,
2852
+ 1000041,
2853
+ 1000042,
2854
+ 1000043,
2855
+ 1000044,
2856
+ 1000045,
2857
+ 1000046,
2858
+ 1000047,
2859
+ 1000048,
2860
+ 1000049,
2861
+ 1000050,
2862
+ 1000051,
2863
+ 1000052,
2864
+ 1000053,
2865
+ 1000054,
2866
+ 1000055,
2867
+ 1000056,
2868
+ 1000057,
2869
+ 1000058,
2870
+ 1000059,
2871
+ 1000060,
2872
+ 1000061,
2873
+ 1000062,
2874
+ 1000063,
2875
+ 1000064,
2876
+ 1000065,
2877
+ 1000066,
2878
+ 1000067,
2879
+ 1000068,
2880
+ 1000069,
2881
+ 1000070,
2882
+ 1000071,
2883
+ 1000072,
2884
+ 1000073,
2885
+ 1000074,
2886
+ 1000075,
2887
+ 1000076,
2888
+ 1000077,
2889
+ 1000078,
2890
+ 1000079,
2891
+ 1000080,
2892
+ 1000081,
2893
+ 1000082,
2894
+ 1000083,
2895
+ 1000084,
2896
+ 1000085,
2897
+ 1000086,
2898
+ 1000087,
2899
+ 1000088,
2900
+ 1000089,
2901
+ 1000090,
2902
+ 1000091,
2903
+ 1000092,
2904
+ 1000093,
2905
+ 1000094,
2906
+ 1000095,
2907
+ 1000096,
2908
+ 1000097,
2909
+ 1000098,
2910
+ 1000099
2911
+ ],
2912
+ "issued_actions": 5171,
2913
+ "action_counts": {
2914
+ "6": 2869,
2915
+ "7": 440,
2916
+ "1": 244,
2917
+ "3": 426,
2918
+ "2": 73,
2919
+ "4": 140,
2920
+ "8": 931,
2921
+ "5": 48
2922
+ },
2923
+ "largest_action_fraction": 0.5548249854960355,
2924
+ "raw_score_min": [
2925
+ -0.11920478940010071,
2926
+ -0.07697911560535431,
2927
+ -0.08089113235473633,
2928
+ -0.12366610765457153,
2929
+ -0.035442501306533813,
2930
+ -0.050102174282073975,
2931
+ -0.25782644748687744,
2932
+ -0.11103026568889618,
2933
+ -0.22623014450073242
2934
+ ],
2935
+ "raw_score_max": [
2936
+ 0.08351439237594604,
2937
+ 1.0634565353393555,
2938
+ 1.0248469114303589,
2939
+ 1.109860897064209,
2940
+ 1.0878527164459229,
2941
+ 1.0884151458740234,
2942
+ 1.0750973224639893,
2943
+ 1.0779128074645996,
2944
+ 1.0730949640274048
2945
+ ],
2946
+ "raw_score_mean": [
2947
+ -0.00917475129739012,
2948
+ 0.04486093732473096,
2949
+ 0.018818002592468228,
2950
+ 0.07282999691439407,
2951
+ 0.02743620055716314,
2952
+ 0.0065598727885242515,
2953
+ 0.5550956811714112,
2954
+ 0.08844435388464397,
2955
+ 0.18066050732672917
2956
+ ]
2957
+ },
2958
+ "checkpoint_sha256": "381d1ed73a6f3949cb1880d6f1feb18db13068aeb6db39a7974ae9fb1fba0f56",
2959
+ "training_step": 5000,
2960
+ "profile_sha256": "a5acfbb259cc3431bcce9dedef862a1e3895bf81659aa3c1707fa54c277852fc",
2961
+ "hf_revision": "8336631d8221b15728cd5c32ff242c339499276f",
2962
+ "hf_path": "zero-latency/asterix/vla/starvla-qwenpi_v3-h1/asterix-qwenpi_v3-hist8-zero-s42-v1",
2963
+ "wandb_url": "https://wandb.ai/saberrr-zju/starvla_tasks/runs/ec20df07",
2964
+ "wandb_readback": {
2965
+ "eval_atari_hist8_v1/zero_to_profile/profile_sample/profile_sample/mean_return": 143.5
2966
+ },
2967
+ "status": "verified"
2968
+ }
2969
+ ]
2970
+ }