File size: 5,627 Bytes
4c27541
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
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
{
  "actions": [
    "left",
    "right",
    "forward",
    "pickup",
    "drop",
    "toggle",
    "done"
  ],
  "evidence": {
    "ids": {
      "description": "one opaque id per row",
      "keys": [
        "split_ids_new",
        "split_ids_old"
      ]
    },
    "logits": "logits_new and logits_old hold the model's action logits (float32 [N, 7]) on each row; the model takes the argmax action",
    "new": {
      "count": 37,
      "demonstrations": 8,
      "description": "every step of 8 successful demonstrations of the new mission by a planner that sees the whole grid, in held-out rooms; episode_new numbers the demonstration (0 to 7) and step_new the step in it (0 is the first)",
      "keys": [
        "x_new",
        "y_new",
        "logits_new",
        "episode_new",
        "step_new"
      ]
    },
    "old": {
      "count": 64,
      "description": "single steps of the trained mission, a fixed random draw from the model's training demonstrations, with the planner's action",
      "keys": [
        "x_old",
        "y_old",
        "logits_old"
      ]
    }
  },
  "input": "observation rows int64 [N, 197]: the grid and the instruction; model_spec.json input gives the layout, the cell encoding and every instruction word's meaning",
  "limit_definitions": {
    "max_changed_frac": "changed entries over all parameter entries; an entry is changed if its float32 value differs from the parent's; null: this limit is not applied",
    "max_edit_frac": "numbers the patch specifies, counted per op (add_dense: tensor size; index-form set/add: number of values; slice-form set/add: 1 with one value, the block size with a values file; scale_rows/cols: positions; project_out: direction length; low_rank_add: r*(rows+cols); tie: group-map length), over all parameter entries",
    "max_file_bytes": "total bytes of the distinct .npy files one patch references",
    "max_ops": "operations per patch",
    "max_rank": "rank r of u and v in low_rank_add",
    "max_rel_frob": "per tensor of rank >= 2, ||child - parent||_F / ||parent||_F; 1-D tensors are exempt; null: this limit is not applied",
    "rel_frob_min_ndim": "smallest tensor rank that max_rel_frob applies to; tensors of lower rank are exempt"
  },
  "limits": {
    "max_changed_frac": null,
    "max_edit_frac": 0.05,
    "max_file_bytes": 67108864,
    "max_ops": 64,
    "max_rank": 8,
    "max_rel_frob": null,
    "rel_frob_min_ndim": 2
  },
  "new_mission": {
    "facts": "the model was never trained on a row holding these words: their embedding rows are untrained, and the action pickup (index 3) was never a training target, so the model does not perform the new mission. The rooms, objects and colors are those of the trained mission; an episode ends when the mission is done or after 64 steps. What completes the new mission is not stated; the demonstrations show it",
    "instruction": "pick up {the|a} {color} {type}",
    "words": [
      {
        "id": 2,
        "word": "pick"
      },
      {
        "id": 3,
        "word": "up"
      }
    ]
  },
  "patch_check": {
    "budget_queries": 8,
    "final_selection": "kept at the end: among the checked patches and the unedited classifier (S = 0), the one with the highest S on the check split; ties go to fewer changed parameters, then to the earlier query; a patch with S <= 0 and an unchecked patch are never kept",
    "invalid_patches": "a patch that fails validation consumes no query",
    "parts": {
      "new": "500 held-out episodes of the new mission",
      "old": "500 held-out episodes of the trained mission"
    },
    "readout": "each query applies one patch and runs the patched model greedily (argmax action at every step, at most 64 steps) on every held-out episode of both parts, and returns aggregate numbers, each against the unedited model: the success rate, fixed (episodes the unedited model failed that the patched one completes), broken (episodes the unedited model completed that the patched one fails), the changed-parameter count, S_new and S_old (fixed - broken on each part) and S = S_new + S_old; no per-episode information"
  },
  "patch_format": {
    "allowed_ops": [
      "scale_rows",
      "scale_cols",
      "set",
      "add",
      "project_out",
      "low_rank_add",
      "add_dense",
      "tie"
    ],
    "schema": "wrsi.patch.v0"
  },
  "rows": {
    "layout": "int64 [197]: the grid [x, y, channel] in C order (192 values), then 5 word ids",
    "targets": "y holds the planner's action index (0 to 6) at that step"
  },
  "schema": "wrsi.task.policyskill.v1",
  "trained_mission": "go to {the|a} {color} {type}",
  "training_free_rule": {
    "allowed": [
      "forward passes, activations and weight inspection",
      "label-conditioned descriptive statistics (class means and their differences, SVD/PCA, correlations)",
      "single-element ablation sweeps for attribution",
      "comparing up to 8 explicit candidate patches on evidence"
    ],
    "audit": "every executed cell is logged and audited against these rules",
    "forbidden": [
      "gradients of any loss with respect to parameters (autograd, hand-written backpropagation, finite-difference gradients)",
      "solvers whose targets derive from evidence labels (lstsq, solve, pinv, inv, closed-form regression or a linear probe)",
      "any .fit( call",
      "iterative parameter updates driven by a label-dependent objective (gradient descent, coordinate descent, evolutionary or random search that keeps the best)",
      "scoring more than 8 distinct full candidate patches against evidence labels in one session"
    ]
  }
}