bingran-you commited on
Commit
d9462cf
·
verified ·
1 Parent(s): 011bfff

Correct the scene 02 provenance of the BenchFlow Oracle runs

Browse files

Scene 02's capture-time verifier stopped on two SSIM self-consistency rows before the bounds, not on a bound. scene_02.zip changes only in provenance.json; submission/ is unchanged. The folder README also notes that scene 06 ran twice.

oracle/benchflow-h100/MANIFEST.sha256 CHANGED
@@ -1,4 +1,4 @@
1
- fb3ea31531140d27bac380c54027978bc0640205a3b6aec213c1616c0dfbe570 oracle/benchflow-h100/scene_02.zip
2
  f8f85f56a229d340c347381879a316870b068220f1c24673de147e32e9b11f39 oracle/benchflow-h100/scene_04.zip
3
  8336b6a8e35b7b84bb96df1bf5cfde944db45b8bfcf3987998ffb839ab6a9ddb oracle/benchflow-h100/scene_06.zip
4
  32f94a5998be3bb7a5871dab36d69e19c5c2b9a15986dbe0b53431cb5532d0c8 oracle/benchflow-h100/scene_23.zip
 
1
+ 633d6ff53c759704a4e8914b5670bfbcdf6f45b919b9a4bb43207d8e0ad5ea36 oracle/benchflow-h100/scene_02.zip
2
  f8f85f56a229d340c347381879a316870b068220f1c24673de147e32e9b11f39 oracle/benchflow-h100/scene_04.zip
3
  8336b6a8e35b7b84bb96df1bf5cfde944db45b8bfcf3987998ffb839ab6a9ddb oracle/benchflow-h100/scene_06.zip
4
  32f94a5998be3bb7a5871dab36d69e19c5c2b9a15986dbe0b53431cb5532d0c8 oracle/benchflow-h100/scene_23.zip
oracle/benchflow-h100/README.md CHANGED
@@ -1,18 +1,20 @@
1
  # BenchFlow Oracle runs
2
 
3
- Retained runs of the task's training Oracle, unchanged, at FrontierPhysics main `9787d07908d94e8093a18f4dc3b120c0db5c871e` (`tasks/multi-object-visual-inverse-dynamics/oracle/`, since moved to `oracle/reference/training-pipeline/`). Each scene ran once on one H100 in the task image; the run's sandbox hostname drew the scene. The task's `oracle/replay.py` restores `submission/` and recomputes `results/metrics.json` with the verifier's metric code.
4
 
5
  Each `scene_<id>.zip` holds:
6
 
7
  - `submission/`: the run's `/root/paper.pdf`, `/root/solution/` and `/root/results/`, byte for byte;
8
  - `evidence/verifier/`: the unmodified verifier's output for the run at that revision;
9
- - `provenance.json`: seed, GPU, date, sandbox hostname, the verifier's interval measurements and the task's bounds after recalibration.
10
 
11
- At that revision these runs failed the scene bounds, which were the MOSIV paper's material-pair means; every other verifier check passed. The task's bounds for these scenes were then moved to the runs' measurements (`oracle/README.md` in the task). The trained model checkpoints were not retained.
 
 
12
 
13
  | Scene | Bytes | SHA-256 |
14
  |---|---|---|
15
- | 02 | 53933577 | `fb3ea31531140d27bac380c54027978bc0640205a3b6aec213c1616c0dfbe570` |
16
  | 04 | 16407865 | `f8f85f56a229d340c347381879a316870b068220f1c24673de147e32e9b11f39` |
17
  | 06 | 19858330 | `8336b6a8e35b7b84bb96df1bf5cfde944db45b8bfcf3987998ffb839ab6a9ddb` |
18
  | 23 | 14333724 | `32f94a5998be3bb7a5871dab36d69e19c5c2b9a15986dbe0b53431cb5532d0c8` |
 
1
  # BenchFlow Oracle runs
2
 
3
+ Retained runs of the task's training Oracle, unchanged, at FrontierPhysics main `9787d07908d94e8093a18f4dc3b120c0db5c871e` (`tasks/multi-object-visual-inverse-dynamics/oracle/`, since moved to `oracle/reference/training-pipeline/`), seed 0, on one H100 per scene in the task image, on 2026-10-17. The run's sandbox hostname drew the scene. The task's `oracle/replay.py` restores `submission/` and recomputes `results/metrics.json` with the verifier's metric code.
4
 
5
  Each `scene_<id>.zip` holds:
6
 
7
  - `submission/`: the run's `/root/paper.pdf`, `/root/solution/` and `/root/results/`, byte for byte;
8
  - `evidence/verifier/`: the unmodified verifier's output for the run at that revision;
9
+ - `provenance.json`: seed, GPU, date, sandbox hostname, why the verifier rejected the run, the verifier's interval measurements and the task's bounds after recalibration.
10
 
11
+ At that revision the verifier rejected every run. For 04, 06, 23, 26, 45 and 67 it stopped on a scene bound; the bounds were the MOSIV paper's material-pair means. For 02 it stopped earlier, on two SSIM rows of the pipeline's own metrics.json that differ from its measurement by 2.1e-6 and 2.6e-6 (tolerance 2e-6); replay recomputes those rows. Every other verifier check passed. The task's bounds for these scenes were then moved to the runs' measurements (`oracle/README.md` in the task).
12
+
13
+ Scene 06 was trained twice that day. This package is the run that kept `/root/solution/`, which replay needs; the other run (future CD 19.8 against 42.8 here) did not keep it. The trained model checkpoints were not retained.
14
 
15
  | Scene | Bytes | SHA-256 |
16
  |---|---|---|
17
+ | 02 | 53933744 | `633d6ff53c759704a4e8914b5670bfbcdf6f45b919b9a4bb43207d8e0ad5ea36` |
18
  | 04 | 16407865 | `f8f85f56a229d340c347381879a316870b068220f1c24673de147e32e9b11f39` |
19
  | 06 | 19858330 | `8336b6a8e35b7b84bb96df1bf5cfde944db45b8bfcf3987998ffb839ab6a9ddb` |
20
  | 23 | 14333724 | `32f94a5998be3bb7a5871dab36d69e19c5c2b9a15986dbe0b53431cb5532d0c8` |
oracle/benchflow-h100/packages.json CHANGED
@@ -3,8 +3,8 @@
3
  "scene": "02",
4
  "label": "02",
5
  "path": "oracle/benchflow-h100/scene_02.zip",
6
- "bytes": 53933577,
7
- "sha256": "fb3ea31531140d27bac380c54027978bc0640205a3b6aec213c1616c0dfbe570",
8
  "members": 472
9
  },
10
  {
 
3
  "scene": "02",
4
  "label": "02",
5
  "path": "oracle/benchflow-h100/scene_02.zip",
6
+ "bytes": 53933744,
7
+ "sha256": "633d6ff53c759704a4e8914b5670bfbcdf6f45b919b9a4bb43207d8e0ad5ea36",
8
  "members": 472
9
  },
10
  {
oracle/benchflow-h100/scene_02.zip CHANGED
@@ -1,3 +1,3 @@
1
  version https://git-lfs.github.com/spec/v1
2
- oid sha256:fb3ea31531140d27bac380c54027978bc0640205a3b6aec213c1616c0dfbe570
3
- size 53933577
 
1
  version https://git-lfs.github.com/spec/v1
2
+ oid sha256:633d6ff53c759704a4e8914b5670bfbcdf6f45b919b9a4bb43207d8e0ad5ea36
3
+ size 53933744