CoolFace
Modelpublic

Chaman1234/Sparse-AST-BWM

sourceHugging Faceapache-2.0updated 8d agoView on Hugging Face
0likes1.7kdownloads
config.json88 linesDownload Raw Back to root
1{
2  "model_family": "Sparse-AST-BWM",
3  "description": "Sparse-AST Blender World Model (BWM) Family with Recurrent Leaky Gating and Top-K Mixture-of-Experts Gating Router for Procedural 3D Mathematics and Blender Python.",
4  "author": "Chaman1234",
5  "default_model": "100M-32",
6  "models": {
7    "3M-32": {
8      "subfolder": "3M-32",
9      "status": "ready",
10      "params": 3276368,
11      "native_seq_len": 64,
12      "layers": 4,
13      "d_model": 256,
14      "loss": 16.2013,
15      "perplexity": 10867358.11
16    },
17    "10M-32": {
18      "subfolder": "10M-32",
19      "status": "ready",
20      "params": 10517352,
21      "native_seq_len": 64,
22      "layers": 6,
23      "d_model": 384,
24      "loss": 21.2309,
25      "perplexity": 485165195.41
26    },
27    "100M-32": {
28      "subfolder": "100M-32",
29      "status": "ready",
30      "params": 100893176,
31      "native_seq_len": 32,
32      "layers": 18,
33      "d_model": 704,
34      "loss": 3.9078,
35      "perplexity": 49.79
36    },
37    "200M-32": {
38      "subfolder": "200M-32",
39      "status": "ready",
40      "params": 201530672,
41      "native_seq_len": 32,
42      "layers": 28,
43      "d_model": 800,
44      "loss": 4.4454,
45      "perplexity": 85.24
46    },
47    "200M-512": {
48      "subfolder": "200M-512",
49      "status": "pending",
50      "target_spec": {
51        "pt_filename": "final_sparse_ast_200m_512.pt",
52        "alt_patterns": [
53          "checkpoint_200m_512_1000.pt",
54          "checkpoint_200m_512_*.pt"
55        ],
56        "d_model": 800,
57        "d_hidden": 1600,
58        "num_layers": 28,
59        "num_heads": 8,
60        "al_hidden": 24,
61        "vocab_size": 512,
62        "native_seq_len": 512,
63        "target_eval_seq_len": 512,
64        "params": 201121072,
65        "loss": null,
66        "perplexity": null,
67        "eval_time_s": null,
68        "description": "Native 512-token context window 201.1M Sparse-AST model, supporting long-horizon Blender scripts and complex geometry pipelines natively.",
69        "type": "backbone"
70      }
71    },
72    "TopK-MoE": {
73      "subfolder": "TopK-MoE",
74      "status": "ready",
75      "params": 37316,
76      "native_seq_len": 32,
77      "layers": null,
78      "d_model": null,
79      "loss": null,
80      "perplexity": null
81    },
82    "500M-32": {
83      "subfolder": "500M-32",
84      "status": "pending",
85      "params": 501300512
86    }
87  }
88}