File size: 5,769 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
{
  "classes": [
    "I_WALK",
    "I_LOOK",
    "I_RUN",
    "I_JUMP",
    "I_TURN_LEFT",
    "I_TURN_RIGHT",
    "END"
  ],
  "evidence": {
    "ids": {
      "description": "one opaque id per example",
      "keys": [
        "split_ids_prim",
        "split_ids_old"
      ]
    },
    "old": {
      "count": 64,
      "description": "rows of training commands, a fixed random draw from the model's training rows, with their answers",
      "keys": [
        "x_old",
        "y_old",
        "pred_old"
      ]
    },
    "predictions": "pred_prim and pred_old hold the model's argmax at each OUT position and -1 elsewhere",
    "prim": {
      "count": 1,
      "description": "the command jump alone, the one training command that holds jump, as it appears in training, with its answer",
      "keys": [
        "x_prim",
        "y_prim",
        "pred_prim"
      ]
    }
  },
  "input": "token rows int64 [N, seq_len]; model_spec.json input.tokens gives every token's meaning (the interpretation rules) and input.layout the row format and the grammar",
  "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",
    "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": 0.5,
    "rel_frob_min_ndim": 2
  },
  "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": "held-out commands that hold jump together with other words: jump with a direction, opposite or around, repeated by twice or thrice, and joined to other commands by and or after",
      "old": "held-out commands of the training distribution, never in the model's training rows"
    },
    "readout": "each query applies one patch and returns aggregate numbers on a held-out check split, each against the unedited model, for the two parts: exact-match accuracy, fixed (rows the unedited model got wrong that the patched one gets right), broken (rows the unedited model got right that the patched one gets wrong), the changed-parameter count, S_new and S_old (fixed - broken on each part) and S = S_new + S_old; no per-row 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": "[BOS, command words, SEP, OUT x 48], right-padded with PAD to seq_len; a command has 1 to 9 words of the grammar below; the answer at the k-th OUT slot is the k-th action of the command's meaning, END at every slot after the last action. Grammar: commands: C -> S | S and S | S after S; S -> V | V twice | V thrice; V -> D opposite left | D opposite right | D around left | D around right | D left | D right | U; D -> U | turn; U -> walk | look | run | jump (turn appears only with a direction). [[.]] is a command's meaning, a sequence of actions; meanings concatenate",
    "targets": "y holds the class index (0 to 6: the six actions, then END) at each OUT position and -1 elsewhere; a row is right when every OUT position is right, END included"
  },
  "schema": "wrsi.task.composeskill.v1",
  "target": {
    "commands": "commands that hold jump together with other words: jump with a direction, opposite or around, repeated by twice or thrice, and joined to other commands by and or after",
    "training": "the model was trained on commands without jump and on the command jump alone; it was never trained on a command that holds jump with another word"
  },
  "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"
    ]
  }
}