Hugging Face's logo

mindlab-research
/
MJev-Spatial-4B

Safetensors
Model card Files
xet
 Community
2
MJev-Spatial-4B
File size: 4,617 Bytes
0657319
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
50
51
52
53
54
55
56
57
58
59
60
61
62
63
64
65
66
67
68
69
70
71
72
73
74
75
76
77
78
79
80
81
82
83
84
85
86
87
88
89
90
91
92
93
94
95
96
97
98
99
100
101
102
103
104
105
106
107
108
109
110
111
112
113
114
115
116
117
118
119
120
121
122
123
124
125
126
127
128
129
130
131
132
133
134
135
136
137
138
139
140
141
142
143
144
145
146
147
148
149
150
151
152
153
154
155
156
157
158
159
160
161
162
163
164
165
166
167
168
169
170
171
172
173
174
175
176
177
178
179
{
  "schema_version": "spatial-decision/v3",
  "architecture": "kev-spatial/v1",
  "base_model": "Qwen/Qwen3.5-4B",
  "base_revision": "851bf6e806efd8d0a36b00ddf55e13ccb7b8cd0a",
  "seed": 3407,
  "data_seed": 3407,
  "head_seed": 3407,
  "train_mode": "full_language",
  "readout_kind": "native_last_qk_mean",
  "readout_layer_index": 31,
  "native_readout_fp32": true,
  "epochs": 1,
  "pilot_steps": 500,
  "global_batch_size": 512,
  "micro_batch_size": 16,
  "expected_world_size": 8,
  "head_learning_rate": 0.000005,
  "backbone_learning_rate": 0.000005,
  "head_warmup_steps": 0,
  "warmup_steps": 0,
  "scheduler": {
    "kind": "constant"
  },
  "weight_decay": 0.01,
  "head_weight_decay": 0.0,
  "adam_betas": [
    0.9,
    0.95
  ],
  "gradient_clip": 1.0,
  "optimizer_fused": true,
  "precision": "BF16 compute; FP32 language weights, gradients, Adam states and head; frozen BF16 vision",
  "trainable_dtype": "float32",
  "gradient_checkpointing": true,
  "max_choices": 32,
  "max_images": 5,
  "max_sequence_length": 4096,
  "min_image_tokens": 64,
  "max_image_tokens": 512,
  "point_variant": "crop_refill",
  "point_depth": 3,
  "scalar_codebook": {
    "unit": "m",
    "edges_m": [
      0.0,
      0.006111921607101044,
      0.01595940178733695,
      0.031825584656481165,
      0.05738905411563928,
      0.098576715708787,
      0.1649379531856473,
      0.27185865878338616,
      0.4441284614600559,
      0.7216882170597828,
      1.168890319410675,
      1.889418840971512,
      3.0503287464783715,
      4.920777685521732,
      7.934430353116933,
      12.790003906249407,
      20.61326595080775,
      33.21804442818109,
      53.52676469841237,
      86.24801559186353,
      138.96823852002032,
      223.910648412756,
      360.7691933437561,
      581.2746080627744,
      936.5512036521882,
      1508.970069349626,
      2431.2468584039048,
      3917.2119909330504,
      6311.387363552563,
      10168.863955282299,
      16384.000000000004
    ],
    "fine_bins": 8,
    "fine_bins_by_root": [
      4,
      4,
      4,
      4,
      4,
      4,
      8,
      8,
      8,
      8,
      8,
      8,
      8,
      8,
      8,
      8,
      8,
      8,
      8,
      8,
      8,
      8,
      8,
      8,
      8,
      8,
      8,
      8,
      8,
      8
    ],
    "zero_class": 0,
    "overflow_class": 31,
    "overflow_representative_m": 16384.0,
    "scale_m": 0.01,
    "fine_interpolation": "linear_within_log_spaced_root",
    "version": "metric-log1p30-fine48-16384-v1",
    "positive_leaf_cells": 216,
    "fit": "frozen_previous_Molmo_train_only_proposal_not_refit_on_72k_or_test",
    "activation": "user_selected_full_data_once_launch_by_user_only",
    "proposal_sha256": "fb9089b65e5db06693abbf1190dc0b75f2de5ccb515679f66860095577bb7770",
    "precision_design": {
      "absolute_floor_m": 0.01,
      "relative_fraction": 0.05,
      "criterion": "max(0.01 m, 0.05 * value)",
      "qualification": "Engineering quantization tolerance, NOT official accuracy requirement or model-error guarantee",
      "selection": "Smallest power-of-two candidate count >=4 satisfying midpoint half-width bound at each coarse lower edge"
    }
  },
  "numeric_codebooks": {
    "count": {
      "version": "integer-32x32-v1",
      "unit": "count",
      "fine_bins": 32,
      "maximum": 1023
    },
    "deg": {
      "version": "degrees-30x32-v1",
      "unit": "deg",
      "fine_bins": 32,
      "edges_m": [
        0,
        30,
        60,
        90,
        120,
        150,
        180,
        210,
        240,
        270,
        300,
        330,
        360
      ],
      "overflow_representative_m": 360.0
    }
  },
  "data_directory": "data/processed/molmo_priority_20260928-r12-content-only",
  "expected_train_qa": 12304058,
  "sampler": "flat_qa",
  "task_sampling": "All 12304058 QA exactly once in seed3407 flat permutation; no task/source resampling; zero-weight tail padding",
  "loss_weights": {
    "choice": 1.0,
    "scalar": 1.0,
    "point": 1.0
  },
  "micro_batch_layout": "global_length_balanced",
  "deduplicate_vision": false,
  "shared_prefix": false,
  "num_workers": 2,
  "checkpoint_policy": "full",
  "save_every": 1000,
  "diagnostic_note": "Full r12 one-pass resumable recipe: global512/micro16, non-reentrant gradient checkpointing, full model/optimizer/RNG checkpoints every 1000 optimizer steps.",
  "distill_teacher": "Qwen3-VL-32B-Instruct",
  "distill_test_set_leakage": true,
  "distill_batch_ratio": "one_global512_base_plus_one_global512_distill_per_optimizer_step",
  "fixed_learning_rate": 0.000005
}