Add the scene 56 BenchFlow Oracle run
Browse filesScene 56: a run of the task's training Oracle with one change, iter_cnt 200 -> 4 in its config, which stops the physical-parameter fit before the divergence that stops the unchanged pipeline (NaN at iteration 6, seeds 0 and 1). One H100 SXM5, seed 0, 2026-10-18. Oracle-only; the task's agent sandbox blocks this repository.
- README.md +2 -2
- oracle/README.md +2 -2
- oracle/benchflow-h100/MANIFEST.sha256 +1 -0
- oracle/benchflow-h100/README.md +4 -1
- oracle/benchflow-h100/packages.json +8 -0
- oracle/benchflow-h100/scene_56.zip +3 -0
README.md
CHANGED
|
@@ -46,9 +46,9 @@ The task's agent sandbox blocks this repository.
|
|
| 46 |
| `oracle/retained_current_results.json` | The author's verifier record for both packages |
|
| 47 |
| `oracle/MOSIV_training_pipeline_main.zip` | Training and evaluation source of the task oracle |
|
| 48 |
| `oracle/upstream/` | The author's README, integration notes, licence notes, manifest and agent denylist |
|
| 49 |
-
| `oracle/benchflow-h100/scene_<id>.zip` | Scenes 02, 04, 06, 23, 26, 45 and 67: BenchFlow's retained run of the task's
|
| 50 |
|
| 51 |
-
|
| 52 |
author file and its upstream path; `oracle/benchflow-h100/MANIFEST.sha256` lists BenchFlow's packages. Source archives keep their original licences, including the GRAPHDECO
|
| 53 |
notices for the Gaussian Splatting code.
|
| 54 |
|
|
|
|
| 46 |
| `oracle/retained_current_results.json` | The author's verifier record for both packages |
|
| 47 |
| `oracle/MOSIV_training_pipeline_main.zip` | Training and evaluation source of the task oracle |
|
| 48 |
| `oracle/upstream/` | The author's README, integration notes, licence notes, manifest and agent denylist |
|
| 49 |
+
| `oracle/benchflow-h100/scene_<id>.zip` | Scenes 02, 04, 06, 23, 26, 45, 56 and 67: BenchFlow's retained run of the task's training Oracle (one H100 per scene; unchanged except scene 56's `iter_cnt`) with the verifier's output for it; see `oracle/benchflow-h100/README.md` |
|
| 50 |
|
| 51 |
+
Every scene has a package. `oracle/MANIFEST.sha256` and `oracle/provenance.json` list the size and SHA-256 of every
|
| 52 |
author file and its upstream path; `oracle/benchflow-h100/MANIFEST.sha256` lists BenchFlow's packages. Source archives keep their original licences, including the GRAPHDECO
|
| 53 |
notices for the Gaussian Splatting code.
|
| 54 |
|
oracle/README.md
CHANGED
|
@@ -13,9 +13,9 @@ The task's agent sandbox blocks this repository.
|
|
| 13 |
| `oracle/retained_current_results.json` | The author's verifier record for both packages |
|
| 14 |
| `oracle/MOSIV_training_pipeline_main.zip` | Training and evaluation source of the task oracle |
|
| 15 |
| `oracle/upstream/` | The author's README, integration notes, licence notes, manifest and agent denylist |
|
| 16 |
-
| `oracle/benchflow-h100/scene_<id>.zip` | Scenes 02, 04, 06, 23, 26, 45 and 67: BenchFlow's retained run of the task's
|
| 17 |
|
| 18 |
-
|
| 19 |
author file and its upstream path; `oracle/benchflow-h100/MANIFEST.sha256` lists BenchFlow's packages. Source archives keep their original licences, including the GRAPHDECO
|
| 20 |
notices for the Gaussian Splatting code.
|
| 21 |
|
|
|
|
| 13 |
| `oracle/retained_current_results.json` | The author's verifier record for both packages |
|
| 14 |
| `oracle/MOSIV_training_pipeline_main.zip` | Training and evaluation source of the task oracle |
|
| 15 |
| `oracle/upstream/` | The author's README, integration notes, licence notes, manifest and agent denylist |
|
| 16 |
+
| `oracle/benchflow-h100/scene_<id>.zip` | Scenes 02, 04, 06, 23, 26, 45, 56 and 67: BenchFlow's retained run of the task's training Oracle (one H100 per scene; unchanged except scene 56's `iter_cnt`) with the verifier's output for it; see `oracle/benchflow-h100/README.md` |
|
| 17 |
|
| 18 |
+
Every scene has a package. `oracle/MANIFEST.sha256` and `oracle/provenance.json` list the size and SHA-256 of every
|
| 19 |
author file and its upstream path; `oracle/benchflow-h100/MANIFEST.sha256` lists BenchFlow's packages. Source archives keep their original licences, including the GRAPHDECO
|
| 20 |
notices for the Gaussian Splatting code.
|
| 21 |
|
oracle/benchflow-h100/MANIFEST.sha256
CHANGED
|
@@ -4,4 +4,5 @@ f8f85f56a229d340c347381879a316870b068220f1c24673de147e32e9b11f39 oracle/benchfl
|
|
| 4 |
32f94a5998be3bb7a5871dab36d69e19c5c2b9a15986dbe0b53431cb5532d0c8 oracle/benchflow-h100/scene_23.zip
|
| 5 |
10bfdda0f727c1b2e3d35a1c33a2b32de24817fa3f588cb9b7cc0840450967f9 oracle/benchflow-h100/scene_26.zip
|
| 6 |
058795342ed3fd41d165489972bb9c77d27b6030bb5ed58315f517ce5087190e oracle/benchflow-h100/scene_45.zip
|
|
|
|
| 7 |
c500f7939f5be5f2603e363c2ed637409d781e66496468fe8fe4f91fab93f641 oracle/benchflow-h100/scene_67.zip
|
|
|
|
| 4 |
32f94a5998be3bb7a5871dab36d69e19c5c2b9a15986dbe0b53431cb5532d0c8 oracle/benchflow-h100/scene_23.zip
|
| 5 |
10bfdda0f727c1b2e3d35a1c33a2b32de24817fa3f588cb9b7cc0840450967f9 oracle/benchflow-h100/scene_26.zip
|
| 6 |
058795342ed3fd41d165489972bb9c77d27b6030bb5ed58315f517ce5087190e oracle/benchflow-h100/scene_45.zip
|
| 7 |
+
4a2d1510fee7d6c1be39dca071372dcb8ef901e6e1fda7abab22c3cd513c3e1e oracle/benchflow-h100/scene_56.zip
|
| 8 |
c500f7939f5be5f2603e363c2ed637409d781e66496468fe8fe4f91fab93f641 oracle/benchflow-h100/scene_67.zip
|
oracle/benchflow-h100/README.md
CHANGED
|
@@ -1,6 +1,6 @@
|
|
| 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 |
|
|
@@ -10,6 +10,8 @@ Each `scene_<id>.zip` holds:
|
|
| 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 |
|
|
@@ -20,6 +22,7 @@ Scene 06 was trained twice that day. This package is the run that kept `/root/so
|
|
| 20 |
| 23 | 14333724 | `32f94a5998be3bb7a5871dab36d69e19c5c2b9a15986dbe0b53431cb5532d0c8` |
|
| 21 |
| 26 | 16589092 | `10bfdda0f727c1b2e3d35a1c33a2b32de24817fa3f588cb9b7cc0840450967f9` |
|
| 22 |
| 45 | 15555049 | `058795342ed3fd41d165489972bb9c77d27b6030bb5ed58315f517ce5087190e` |
|
|
|
|
| 23 |
| 67 | 20633043 | `c500f7939f5be5f2603e363c2ed637409d781e66496468fe8fe4f91fab93f641` |
|
| 24 |
|
| 25 |
Oracle-only material: the task's agent sandbox blocks this repository.
|
|
|
|
| 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 (scene 56: 2026-10-18, see below). 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 |
|
|
|
|
| 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 56's package comes from the training Oracle with one change: `iter_cnt` 200 -> 4 in `source/mosiv/config/genesismo/56.json`, which stops the physical-parameter fit after 4 iterations. The unchanged pipeline stops on NaN parameters at iteration 6 with seeds 0 and 1; the maintainers approved this run for the Oracle. Its `solution/` carries the changed config, and the verifier rejected it at capture only on three scene bounds.
|
| 14 |
+
|
| 15 |
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.
|
| 16 |
|
| 17 |
| Scene | Bytes | SHA-256 |
|
|
|
|
| 22 |
| 23 | 14333724 | `32f94a5998be3bb7a5871dab36d69e19c5c2b9a15986dbe0b53431cb5532d0c8` |
|
| 23 |
| 26 | 16589092 | `10bfdda0f727c1b2e3d35a1c33a2b32de24817fa3f588cb9b7cc0840450967f9` |
|
| 24 |
| 45 | 15555049 | `058795342ed3fd41d165489972bb9c77d27b6030bb5ed58315f517ce5087190e` |
|
| 25 |
+
| 56 | 17049899 | `4a2d1510fee7d6c1be39dca071372dcb8ef901e6e1fda7abab22c3cd513c3e1e` |
|
| 26 |
| 67 | 20633043 | `c500f7939f5be5f2603e363c2ed637409d781e66496468fe8fe4f91fab93f641` |
|
| 27 |
|
| 28 |
Oracle-only material: the task's agent sandbox blocks this repository.
|
oracle/benchflow-h100/packages.json
CHANGED
|
@@ -47,6 +47,14 @@
|
|
| 47 |
"sha256": "058795342ed3fd41d165489972bb9c77d27b6030bb5ed58315f517ce5087190e",
|
| 48 |
"members": 472
|
| 49 |
},
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 50 |
{
|
| 51 |
"scene": "67",
|
| 52 |
"label": "67",
|
|
|
|
| 47 |
"sha256": "058795342ed3fd41d165489972bb9c77d27b6030bb5ed58315f517ce5087190e",
|
| 48 |
"members": 472
|
| 49 |
},
|
| 50 |
+
{
|
| 51 |
+
"scene": "56",
|
| 52 |
+
"label": "56-cap4",
|
| 53 |
+
"path": "oracle/benchflow-h100/scene_56.zip",
|
| 54 |
+
"bytes": 17049899,
|
| 55 |
+
"sha256": "4a2d1510fee7d6c1be39dca071372dcb8ef901e6e1fda7abab22c3cd513c3e1e",
|
| 56 |
+
"members": 472
|
| 57 |
+
},
|
| 58 |
{
|
| 59 |
"scene": "67",
|
| 60 |
"label": "67",
|
oracle/benchflow-h100/scene_56.zip
ADDED
|
@@ -0,0 +1,3 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
version https://git-lfs.github.com/spec/v1
|
| 2 |
+
oid sha256:4a2d1510fee7d6c1be39dca071372dcb8ef901e6e1fda7abab22c3cd513c3e1e
|
| 3 |
+
size 17049899
|