AaronLeeky commited on
Commit
d936f07
Β·
verified Β·
1 Parent(s): 06a8841

Publish formal 300M-176k and 500M-170k inference checkpoints

Browse files
checkpoints/updated/300m/README.md CHANGED
@@ -1,46 +1,11 @@
1
- | Task | Metric | Value | Samples |
2
- |---|---|---:|---:|
3
- | Text-to-Motion | FID ↓ | 0.245161 | 1024 |
4
- | Text-to-Motion | KID mean ↓ | 3.0646918e-04 | 1024 |
5
- | Text-to-Motion | KID std | 0.000000 | 1024 |
6
- | Text-to-Motion | R@1 ↑ | 0.646484 | 1024 |
7
- | Text-to-Motion | R@2 ↑ | 0.807617 | 1024 |
8
- | Text-to-Motion | R@3 ↑ | 0.872070 | 1024 |
9
- | Text-to-Motion | Diversity | 1.375774 | 1024 |
10
- | Text-to-Motion | Matching Score ↓ | 1.100267 | 1024 |
11
- | Text-to-Motion | Contact Sliding Speed ↓ | 0.356658 | 1024 |
12
- | Text-to-Motion | Foot-Ground Error ↓ | 0.041224 | 1024 |
13
- | Text-to-Motion | Body Jerk ↓ | 36.577579 | 1024 |
14
- | Text-to-Motion | HoloMotion g-MPJPE ↓ | 113.133088 | 1024 |
15
- | Text-to-Motion | HoloMotion MPJPE ↓ | 41.761806 | 1024 |
16
- | Text-to-Motion | HoloMotion E-Vel ↓ | 5.131106 | 1024 |
17
- | Text-to-Motion | HoloMotion E-Acc ↓ | 1.348361 | 1024 |
18
- | Audio-to-Motion | BeatAlign ↑ | 0.494386 | 512 |
19
- | Audio-to-Motion | FID-k ↓ | 25.749748 | 512 |
20
- | Audio-to-Motion | FID-g ↓ | 0.350600 | 512 |
21
- | Audio-to-Motion | Diversity-k | 14.195702 | 512 |
22
- | Audio-to-Motion | Diversity-g | 4.883447 | 512 |
23
- | Audio-to-Motion | PFC ↓ | 0.642937 | 512 |
24
- | Audio-to-Motion | Contact Sliding Speed ↓ | 0.437157 | 512 |
25
- | Audio-to-Motion | Foot-Ground Error ↓ | 0.041582 | 512 |
26
- | Audio-to-Motion | Body Jerk ↓ | 92.607785 | 512 |
27
- | Audio-to-Motion | HoloMotion g-MPJPE ↓ | 105.708219 | 512 |
28
- | Audio-to-Motion | HoloMotion MPJPE ↓ | 45.704789 | 512 |
29
- | Audio-to-Motion | HoloMotion E-Vel ↓ | 6.512291 | 512 |
30
- | Audio-to-Motion | HoloMotion E-Acc ↓ | 2.006182 | 512 |
31
- | Human-Reference-to-Motion | g-MPJPE ↓ | 33.864086 | 512 |
32
- | Human-Reference-to-Motion | MPJPE ↓ | 15.050118 | 512 |
33
- | Human-Reference-to-Motion | End-Effector Error ↓ | 38.754670 | 512 |
34
- | Human-Reference-to-Motion | E-Vel ↓ | 4.443727 | 512 |
35
- | Human-Reference-to-Motion | E-Acc ↓ | 3.592424 | 512 |
36
- | Human-Reference-to-Motion | FID ↓ | 0.007115 | 512 |
37
- | Human-Reference-to-Motion | KID mean ↓ | -1.9663787e-05 | 512 |
38
- | Human-Reference-to-Motion | KID std | 0.000000 | 512 |
39
- | Human-Reference-to-Motion | Diversity | 1.359622 | 512 |
40
- | Human-Reference-to-Motion | Contact Sliding Speed ↓ | 0.301409 | 512 |
41
- | Human-Reference-to-Motion | Foot-Ground Error ↓ | 0.045747 | 512 |
42
- | Human-Reference-to-Motion | Body Jerk ↓ | 61.724432 | 512 |
43
- | Human-Reference-to-Motion | HoloMotion g-MPJPE ↓ | 113.139903 | 512 |
44
- | Human-Reference-to-Motion | HoloMotion MPJPE ↓ | 43.409275 | 512 |
45
- | Human-Reference-to-Motion | HoloMotion E-Vel ↓ | 5.483321 | 512 |
46
- | Human-Reference-to-Motion | HoloMotion E-Acc ↓ | 1.593818 | 512 |
 
1
+ # OMG 300M formal checkpoint
2
+
3
+ Validation-loss-selected checkpoint at step 176000, within the 180k training budget.
4
+
5
+ File: `sstep=176000.ckpt` (PyTorch checkpoint, inference copy).
6
+ Optimizer and scheduler states are omitted. The saved provenance records exact equality of all state-dict tensors with the original training checkpoint; the inference-copy SHA-256 was reverified before upload.
7
+
8
+ See `provenance.json` for source-run identification and checksums.
9
+ Sinusoidal history positional encoding; self- and cross-attention Q/K normalization enabled.
10
+
11
+ The previous checkpoint's benchmark scores must not be attributed to this replacement. No benchmark table is included here pending publication of the matching evaluation protocol and artifacts.
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
checkpoints/updated/300m/provenance.json ADDED
@@ -0,0 +1,9 @@
 
 
 
 
 
 
 
 
 
 
1
+ {
2
+ "source_sha256": "8f93be0766e6b32e4b08b68dc5b9f5e326203e280e7b4907c54c299ca6650796",
3
+ "copy_sha256": "6c081cbc6c642547c5e44ba956ac61c656d2d49ccff3de5e2947d141e3484a4f",
4
+ "global_step": 176000,
5
+ "all_state_dict_tensors_exact": true,
6
+ "omitted": "optimizer and scheduler states only",
7
+ "source_run": "300m_scaling_formal_20260910",
8
+ "file": "checkpoints/updated/300m/sstep=176000.ckpt"
9
+ }
checkpoints/updated/300m/{sstep=180000.ckpt β†’ sstep=176000.ckpt} RENAMED
@@ -1,3 +1,3 @@
1
  version https://git-lfs.github.com/spec/v1
2
- oid sha256:571acb20831a8fa600b892ebd5a5f18b42da2dec8bb42eddf614da42ac2dc48d
3
- size 5065534266
 
1
  version https://git-lfs.github.com/spec/v1
2
+ oid sha256:6c081cbc6c642547c5e44ba956ac61c656d2d49ccff3de5e2947d141e3484a4f
3
+ size 1980827761
checkpoints/updated/500m/README.md ADDED
@@ -0,0 +1,11 @@
 
 
 
 
 
 
 
 
 
 
 
 
1
+ # OMG 500M formal checkpoint
2
+
3
+ Validation-loss-selected checkpoint at step 170000, within the 180k training budget.
4
+
5
+ File: `sstep=170000.ckpt` (PyTorch checkpoint, inference copy).
6
+ Optimizer and scheduler states are omitted. The saved provenance records exact equality of all state-dict tensors with the original training checkpoint; the inference-copy SHA-256 was reverified before upload.
7
+
8
+ See `provenance.json` for source-run identification and checksums.
9
+ Sinusoidal history positional encoding; self- and cross-attention Q/K normalization enabled.
10
+
11
+ The previous checkpoint's benchmark scores must not be attributed to this replacement. No benchmark table is included here pending publication of the matching evaluation protocol and artifacts.
checkpoints/updated/500m/provenance.json ADDED
@@ -0,0 +1,9 @@
 
 
 
 
 
 
 
 
 
 
1
+ {
2
+ "source_sha256": "3fe20f5a08113e8532b7e0d8a26784c2971c8d2e08a570912f7441f1cf53067b",
3
+ "copy_sha256": "5eda5b3daa28cc1037eda1b8e74f760a28c37f83c75bb7e5004a790df5106ed9",
4
+ "global_step": 170000,
5
+ "all_state_dict_tensors_exact": true,
6
+ "omitted": "optimizer and scheduler states only",
7
+ "source_run": "500m_scaling_formal_20260908",
8
+ "file": "checkpoints/updated/500m/sstep=170000.ckpt"
9
+ }
checkpoints/updated/500m/sstep=170000.ckpt ADDED
@@ -0,0 +1,3 @@
 
 
 
 
1
+ version https://git-lfs.github.com/spec/v1
2
+ oid sha256:5eda5b3daa28cc1037eda1b8e74f760a28c37f83c75bb7e5004a790df5106ed9
3
+ size 2987797171