Hugging Face's logo

rogermt
/
neurogolf-solver

Model card Files
xet
 Community
rogermt commited on
Commit
bc5d5ee
Β·
verified Β·
1 Parent(s): 9282cd1

Update README.md for v5 refactored package

Browse files
Files changed (1) hide show
  1. README.md +108 -647
README.md CHANGED
@@ -1,681 +1,142 @@
1
- # NeuroGolf Solver v3
2
-
3
- Builds minimal ONNX networks for ARC-AGI tasks. Currently solves **306/400** on CPU (15s budget per task).
4
-
5
- ## Results
6
-
7
- | Version | Solved | Key Changes |
8
- |---------|--------|-------------|
9
- | v1 | 128/400 | Conv solver only |
10
- | v2 | 294/400 | Added spatial_gather, variable-shape conv, diff-shape conv |
11
- | **v3** | **306/400** | Fixed opset 10 compat (Gather vs GatherElements), added concat_enhanced, varshape_spatial_gather, conv_var_diff |
12
-
13
- ### v3 Solver Breakdown (306/400)
14
-
15
- | Solver | Count | Description |
16
- |--------|-------|-------------|
17
- | conv_var | 125 | Variable-shape conv on full 30Γ—30 grid |
18
- | conv_fixed | 106 | Fixed-shape conv (Slice→Conv→Pad) |
19
- | conv_diff | 39 | Diff-shape conv (output smaller than input) |
20
- | spatial_gather | 16 | Fixed-shape pixel remapping |
21
- | concat | 5 | Tiled concat with flips |
22
- | color_map | 4 | 1Γ—1 color remapping conv |
23
- | concat_enhanced | 4 | Tiled concat with all 8 dihedral transforms |
24
- | rotate | 3 | 90Β°/180Β°/270Β° rotation |
25
- | transpose | 2 | Matrix transpose |
26
- | varshape_spatial_gather | 1 | Variable-shape pixel remapping |
27
- | upscale | 1 | Nearest-neighbor upscale |
28
-
29
- Current Status: 306/400 βœ… (was 294)
30
- What the model_runner.py code IS
31
- That's the 2024 ARC Prize winning LLM infrastructure β€” it fine-tunes a language model to generate ARC solutions as text.
32
- It's a completely different competition format (text answers, not ONNX models). It won't help us directly.
33
-
34
- What WOULD help β€” things you can search for:
35
- 1. ARC-GEN additional training data The ARC-GEN paper generates extra input/output examples for each task.
36
-
37
- 2. More examples β†’ our lstsq conv solver gets more data β†’ better fitting β†’ more tasks solved. Search Kaggle/GitHub for "arc-gen" data files or the generator code.
38
-
39
- 3. NeuroGolf / Code Golf Championship notebooks Search Kaggle for neurogolf 2026 or code golf championship ARC ONNX. Other competitors' approaches would be gold.
40
-
41
- 4. CompressARC code (paper) This 76K param approach trains a tiny network PER TASK via gradient descent.
42
-
43
- 5. If we could adapt it to produce opset-10 ONNX models, it could solve many of our 94 remaining tasks. Search GitHub for "CompressARC" or the authors' repos.
44
-
45
- What's actually blocking us (94 unsolved tasks)
46
- Category Count Example Why unsolved
47
- Variable diff-shape (output smaller) ~60 Extract subregion from grid Output shape depends on input content β€” can't build static ONNX
48
- Variable diff-shape (output larger) ~17 Tile/upscale by variable factor Same problem
49
- Same-shape but complex ~10 Multi-step reasoning, flood fill Conv can't learn non-local/algorithmic patterns
50
- Fixed diff-shape (output larger) ~7 Input-driven block placement Output depends on input VALUES, not just positions
51
- The fundamental challenge: 94 tasks require reasoning that depends on input content (not just a fixed pixel remapping or local conv pattern).
52
- Our current ONNX opset 10 toolkit (Conv, Gather, ArgMax, etc.) can only express fixed mappings. We'd need to find tasks where the mapping IS fixed but our solver just hasn't found it yet β€” likely by adding more training examples (ARC-GEN) or trying bigger conv kernels with more time budget.
53
-
54
- GPU RUN 307 solved:
55
- /kaggle/working/neurogolf-solver
56
- Using providers: ['CPUExecutionProvider']
57
- Loaded 400 tasks, solving 400
58
- Conv budget: 60.0s per task
59
- ======================================================================
60
- wandb: [wandb.login()] Loaded credentials for https://api.wandb.ai from WANDB_API_KEY.
61
- wandb: Currently logged in as: rogermt23 to https://api.wandb.ai. Use `wandb login --relogin` to force relogin
62
- wandb: β’Ώ Waiting for wandb.init()...
63
- wandb: β£» Waiting for wandb.init()...
64
- wandb: β£½ Waiting for wandb.init()...
65
- wandb: Tracking run with wandb version 0.25.0
66
- wandb: Run data is saved locally in /kaggle/working/neurogolf-solver/wandb/run-20260424_181206-20lzjuuc
67
- wandb: Run `wandb offline` to turn off syncing.
68
- wandb: Syncing run solver_run
69
- wandb: ⭐️ View project at https://wandb.ai/rogermt23/neurogolf
70
- wandb: πŸš€ View run at https://wandb.ai/rogermt23/neurogolf/runs/20lzjuuc
71
- Task 1: UNSOLVED 8.600s
72
- Task 2: conv_var 7.146 56717390 8.270s ( 250,460 bytes)
73
- Task 3: UNSOLVED 0.899s
74
- Task 4: conv_var 10.302 2417390 0.051s ( 10,457 bytes)
75
- Task 5: conv_fixed 8.200 19770030 12.797s ( 176,971 bytes)
76
- Task 6: conv_diff 13.449 103886 0.009s ( 10,703 bytes)
77
- Task 7: conv_fixed 12.362 308014 0.045s ( 20,171 bytes)
78
- Task 8: conv_var 8.777 11105390 0.299s ( 48,860 bytes)
79
- Task 9: conv_var 8.447 15449390 3.087s ( 68,060 bytes)
80
- Task 10: conv_fixed 11.479 744750 0.076s ( 32,971 bytes)
81
- Task 11: conv_fixed 11.112 1074670 0.314s ( 32,971 bytes)
82
- Task 12: conv_fixed 12.029 429974 0.047s ( 10,568 bytes)
83
- Task 13: conv_var 6.993 66129390 11.785s ( 292,060 bytes)
84
- Task 14: UNSOLVED 3.832s
85
- Task 15: conv_fixed 13.260 125550 0.016s ( 4,168 bytes)
86
- Task 16: color_map 13.252 126500 0.002s ( 549 bytes)
87
- Task 17: conv_fixed 10.357 2286830 0.266s ( 20,171 bytes)
88
- Task 18: conv_var 8.777 11105390 1.022s ( 48,860 bytes)
89
- Task 19: UNSOLVED 4.225s
90
- Task 20: conv_fixed 12.345 313462 0.043s ( 10,568 bytes)
91
- Task 21: UNSOLVED 0.925s
92
- Task 22: conv_diff 11.594 664286 0.014s ( 20,306 bytes)
93
- Task 23: conv_var 8.447 15449390 0.285s ( 68,060 bytes)
94
- Task 24: conv_var 9.172 7485390 0.153s ( 32,860 bytes)
95
- Task 25: conv_var 8.447 15449390 1.888s ( 68,060 bytes)
96
- Task 26: conv_diff 12.626 236654 0.026s ( 20,306 bytes)
97
- Task 27: conv_fixed 10.583 1825462 0.264s ( 68,171 bytes)
98
- Task 28: conv_fixed 10.583 1825462 0.235s ( 68,171 bytes)
99
- Task 29: UNSOLVED 3.006s
100
- Task 30: conv_var 9.661 4589390 0.066s ( 20,060 bytes)
101
- Task 31: UNSOLVED 0.857s
102
- Task 32: conv_var 10.302 2417390 0.025s ( 10,457 bytes)
103
- Task 33: conv_fixed 10.284 2460334 0.444s ( 32,971 bytes)
104
- Task 34: conv_fixed 11.099 1088750 0.267s ( 48,971 bytes)
105
- Task 35: conv_fixed 11.288 901462 0.184s ( 32,971 bytes)
106
- Task 36: UNSOLVED 0.605s
107
- Task 37: conv_fixed 11.755 565462 0.100s ( 20,171 bytes)
108
- Task 38: conv_diff 12.538 258254 0.012s ( 10,703 bytes)
109
- Task 39: conv_diff 12.363 307706 0.013s ( 10,703 bytes)
110
- Task 40: conv_fixed 11.755 565462 0.091s ( 20,171 bytes)
111
- Task 41: conv_fixed 12.345 313462 0.042s ( 10,568 bytes)
112
- Task 42: conv_fixed 10.303 2413462 0.396s ( 90,571 bytes)
113
- Task 43: conv_fixed 9.839 3841462 0.928s ( 144,971 bytes)
114
- Task 44: conv_fixed 11.755 565462 0.106s ( 20,171 bytes)
115
- Task 45: conv_fixed 11.288 901462 0.185s ( 32,971 bytes)
116
- Task 46: UNSOLVED 1.644s
117
- Task 47: conv_fixed 11.099 1088750 0.136s ( 48,971 bytes)
118
- Task 48: UNSOLVED 0.434s
119
- Task 49: UNSOLVED 0.685s
120
- Task 50: conv_var 7.915 26309390 1.850s ( 116,060 bytes)
121
- Task 51: conv_var 8.447 15449390 0.792s ( 68,060 bytes)
122
- Task 52: conv_fixed 14.178 50094 0.011s ( 4,168 bytes)
123
- Task 53: spatial_gather 12.982 165663 0.006s ( 11,115 bytes)
124
- Task 54: conv_var 7.494 40065390 81.896s ( 176,860 bytes)
125
- Task 55: conv_var 7.693 32825390 2.714s ( 144,860 bytes)
126
- Task 56: conv_diff 14.187 49662 0.010s ( 4,303 bytes)
127
- Task 57: conv_diff 11.692 602198 0.028s ( 33,106 bytes)
128
- Task 58: conv_var 7.693 32825390 2.559s ( 144,860 bytes)
129
- Task 59: spatial_gather 12.982 165663 0.014s ( 11,115 bytes)
130
- Task 60: conv_fixed 11.445 770302 0.083s ( 48,971 bytes)
131
- Task 61: conv_fixed 10.173 2749014 0.665s ( 32,971 bytes)
132
- Task 62: conv_fixed 11.755 565462 0.130s ( 20,171 bytes)
133
- Task 63: conv_var 8.447 15449390 0.661s ( 68,060 bytes)
134
- Task 64: conv_var 8.447 15449390 1.532s ( 68,060 bytes)
135
- Task 65: UNSOLVED 1.051s
136
- Task 66: conv_var 8.447 15449390 1.176s ( 68,060 bytes)
137
- Task 67: UNSOLVED 0.876s
138
- Task 68: conv_fixed 11.755 565462 0.129s ( 20,171 bytes)
139
- Task 69: conv_fixed 11.755 565462 0.112s ( 20,171 bytes)
140
- Task 70: conv_fixed 10.284 2460334 0.509s ( 32,971 bytes)
141
- Task 71: conv_fixed 11.504 726550 0.068s ( 10,568 bytes)
142
- Task 72: conv_diff 10.240 2571674 0.276s ( 145,106 bytes)
143
- Task 73: conv_fixed 13.890 66862 0.012s ( 4,168 bytes)
144
- Task 74: spatial_gather 12.982 165663 0.231s ( 11,115 bytes)
145
- Task 75: conv_fixed 10.155 2798478 0.685s ( 90,571 bytes)
146
- Task 76: conv_var 8.447 15449390 0.862s ( 68,060 bytes)
147
- Task 77: conv_var 8.777 11105390 0.900s ( 48,860 bytes)
148
- Task 78: conv_fixed 11.288 901462 0.132s ( 32,971 bytes)
149
- Task 79: conv_diff 11.773 555386 0.018s ( 10,703 bytes)
150
- Task 80: conv_var 7.693 32825390 24.317s ( 144,860 bytes)
151
- Task 81: conv_fixed 12.908 178414 0.020s ( 10,568 bytes)
152
- Task 82: conv_var 8.777 11105390 0.120s ( 48,860 bytes)
153
- Task 83: concat_enhanced 12.982 165663 0.010s ( 11,113 bytes)
154
- Task 84: conv_var 8.163 20517390 0.514s ( 90,460 bytes)
155
- Task 85: conv_var 8.447 15449390 0.783s ( 68,060 bytes)
156
- Task 86: conv_var 9.661 4589390 0.132s ( 20,060 bytes)
157
- Task 87: rotate 12.982 165663 0.007s ( 11,113 bytes)
158
- Task 88: UNSOLVED 2.159s
159
- Task 89: conv_fixed 10.411 2166574 0.733s ( 48,971 bytes)
160
- Task 90: conv_var 7.915 26309390 0.588s ( 116,060 bytes)
161
- Task 91: UNSOLVED 1.697s
162
- Task 92: conv_var 8.777 11105390 1.093s ( 48,860 bytes)
163
- Task 93: conv_fixed 10.269 2497270 0.688s ( 48,971 bytes)
164
- Task 94: conv_fixed 9.280 6716462 1.584s ( 116,171 bytes)
165
- Task 95: conv_fixed 13.260 125550 0.012s ( 4,168 bytes)
166
- Task 96: UNSOLVED 5.350s
167
- Task 97: conv_var 8.777 11105390 0.877s ( 48,860 bytes)
168
- Task 98: conv_var 9.661 4589390 0.170s ( 20,060 bytes)
169
- Task 99: conv_fixed 11.755 565462 0.097s ( 20,171 bytes)
170
- Task 100: conv_diff 13.157 139166 0.011s ( 4,303 bytes)
171
- Task 101: conv_var 7.915 26309390 2.018s ( 116,060 bytes)
172
- Task 102: conv_fixed 9.293 6628374 2.692s ( 176,971 bytes)
173
- Task 103: conv_diff 13.815 72062 0.011s ( 10,703 bytes)
174
- Task 104: UNSOLVED 2.876s
175
- Task 105: conv_var 8.447 15449390 0.561s ( 68,060 bytes)
176
- Task 106: UNSOLVED 2.142s
177
- Task 107: UNSOLVED 44.310s
178
- Task 108: spatial_gather 12.982 165663 0.050s ( 11,115 bytes)
179
- Task 109: UNSOLVED 5.164s
180
- Task 110: conv_fixed 9.237 7013230 1.083s ( 32,971 bytes)
181
- Task 111: conv_diff 11.765 559706 0.014s ( 20,306 bytes)
182
- Task 112: conv_var 7.915 26309390 3.227s ( 116,060 bytes)
183
- Task 113: conv_var 8.163 20517390 0.187s ( 90,460 bytes)
184
- Task 114: UNSOLVED 1.415s
185
- Task 115: UNSOLVED 0.432s
186
- Task 116: concat 12.982 165663 0.005s ( 11,113 bytes)
187
- Task 117: conv_var 8.777 11105390 0.733s ( 48,860 bytes)
188
- Task 118: conv_var 7.693 32825390 9.030s ( 144,860 bytes)
189
- Task 119: conv_fixed 10.238 2575574 0.642s ( 68,171 bytes)
190
- Task 120: conv_var 11.216 969390 0.029s ( 4,057 bytes)
191
- Task 121: conv_diff 10.806 1460126 0.024s ( 33,106 bytes)
192
- Task 122: conv_var 9.661 4589390 0.075s ( 20,060 bytes)
193
- Task 123: UNSOLVED 5.802s
194
- Task 124: UNSOLVED 8.161s
195
- Task 125: conv_fixed 9.280 6716462 1.873s ( 116,171 bytes)
196
- Task 126: conv_var 9.661 4589390 0.066s ( 20,060 bytes)
197
- Task 127: conv_var 11.216 969390 0.020s ( 4,057 bytes)
198
- Task 128: conv_fixed 9.527 5244462 1.645s ( 90,571 bytes)
199
- Task 129: conv_fixed 14.178 50094 0.008s ( 4,168 bytes)
200
- Task 130: spatial_gather 12.982 165663 0.008s ( 11,115 bytes)
201
- Task 131: conv_var 8.777 11105390 0.175s ( 48,860 bytes)
202
- Task 132: conv_var 9.172 7485390 0.193s ( 32,860 bytes)
203
- Task 133: conv_var 8.447 15449390 2.697s ( 68,060 bytes)
204
- Task 134: UNSOLVED 0.610s
205
- Task 135: spatial_gather 12.982 165663 0.010s ( 11,115 bytes)
206
- Task 136: conv_fixed 10.906 1321462 0.209s ( 48,971 bytes)
207
- Task 137: UNSOLVED 55.101s
208
- Task 138: UNSOLVED 8.860s
209
- Task 139: conv_fixed 11.940 469550 0.051s ( 20,171 bytes)
210
- Task 140: rotate 12.982 165663 0.007s ( 11,113 bytes)
211
- Task 141: conv_var 7.693 32825390 2.586s ( 144,860 bytes)
212
- Task 142: concat_enhanced 12.982 165663 0.009s ( 11,113 bytes)
213
- Task 143: conv_fixed 11.755 565462 0.114s ( 20,171 bytes)
214
- Task 144: conv_diff 12.604 241742 0.028s ( 20,306 bytes)
215
- Task 145: conv_var 7.693 32825390 3.322s ( 144,860 bytes)
216
- Task 146: conv_diff 13.870 68166 0.013s ( 4,303 bytes)
217
- Task 147: conv_var 10.302 2417390 0.021s ( 10,457 bytes)
218
- Task 148: conv_var 7.693 32825390 2.903s ( 144,860 bytes)
219
- Task 149: conv_diff 11.594 664286 0.022s ( 20,306 bytes)
220
- Task 150: conv_var 10.302 2417390 0.025s ( 10,457 bytes)
221
- Task 151: conv_var 10.302 2417390 0.037s ( 10,457 bytes)
222
- Task 152: concat_enhanced 12.982 165663 0.009s ( 11,113 bytes)
223
- Task 153: conv_diff 11.765 559706 0.018s ( 20,306 bytes)
224
- Task 154: conv_fixed 10.136 2852462 0.743s ( 48,971 bytes)
225
- Task 155: conv_var 10.302 2417390 0.030s ( 10,457 bytes)
226
- Task 156: conv_fixed 11.288 901462 0.097s ( 32,971 bytes)
227
- Task 157: conv_fixed 10.525 1933862 0.263s ( 48,971 bytes)
228
- Task 158: conv_var 7.915 26309390 7.184s ( 116,060 bytes)
229
- Task 159: UNSOLVED 6.813s
230
- Task 160: conv_fixed 12.345 313462 0.028s ( 10,568 bytes)
231
- Task 161: conv_var 8.777 11105390 1.145s ( 48,860 bytes)
232
- Task 162: conv_fixed 8.964 9207862 5.939s ( 90,571 bytes)
233
- Task 163: conv_fixed 11.583 671470 0.185s ( 20,171 bytes)
234
- Task 164: concat 12.982 165663 0.005s ( 11,113 bytes)
235
- Task 165: conv_fixed 9.247 6939862 3.232s ( 68,171 bytes)
236
- Task 166: conv_var 8.777 11105390 0.251s ( 48,860 bytes)
237
- Task 167: conv_fixed 13.809 72494 0.013s ( 10,568 bytes)
238
- Task 168: conv_fixed 11.288 901462 0.122s ( 32,971 bytes)
239
- Task 169: conv_fixed 10.906 1321462 0.223s ( 48,971 bytes)
240
- Task 170: UNSOLVED 0.783s
241
- Task 171: conv_var 11.216 969390 0.022s ( 4,057 bytes)
242
- Task 172: concat 12.982 165663 0.009s ( 11,113 bytes)
243
- Task 173: conv_var 8.447 15449390 1.812s ( 68,060 bytes)
244
- Task 174: UNSOLVED 0.572s
245
- Task 175: spatial_gather 12.982 165663 0.024s ( 11,115 bytes)
246
- Task 176: conv_var 7.494 40065390 0.371s ( 176,860 bytes)
247
- Task 177: UNSOLVED 1.973s
248
- Task 178: UNSOLVED 0.490s
249
- Task 179: transpose 14.509 36000 0.006s ( 127 bytes)
250
- Task 180: conv_diff 12.149 381182 0.032s ( 20,306 bytes)
251
- Task 181: conv_fixed 11.834 522054 0.084s ( 32,971 bytes)
252
- Task 182: conv_fixed 8.964 9207862 6.414s ( 90,571 bytes)
253
- Task 183: UNSOLVED 0.988s
254
- Task 184: UNSOLVED 0.603s
255
- Task 185: UNSOLVED 0.666s
256
- Task 186: UNSOLVED 1.775s
257
- Task 187: conv_var 7.915 26309390 11.362s ( 116,060 bytes)
258
- Task 188: UNSOLVED 0.835s
259
- Task 189: conv_diff 11.944 468002 0.045s ( 20,306 bytes)
260
- Task 190: spatial_gather 12.982 165663 0.011s ( 11,115 bytes)
261
- Task 191: UNSOLVED 90.275s
262
- Task 192: conv_var 11.216 969390 0.036s ( 4,057 bytes)
263
- Task 193: conv_var 10.302 2417390 0.082s ( 10,457 bytes)
264
- Task 194: spatial_gather 12.982 165663 0.008s ( 11,115 bytes)
265
- Task 195: UNSOLVED 4.533s
266
- Task 196: conv_var 8.163 20517390 0.968s ( 90,460 bytes)
267
- Task 197: conv_var 8.447 15449390 0.541s ( 68,060 bytes)
268
- Task 198: UNSOLVED 56.265s
269
- Task 199: conv_var 8.777 11105390 0.188s ( 48,860 bytes)
270
- Task 200: conv_fixed 9.839 3841462 0.843s ( 144,971 bytes)
271
- Task 201: UNSOLVED 2.753s
272
- Task 202: conv_var 8.777 11105390 0.938s ( 48,860 bytes)
273
- Task 203: conv_var 9.661 4589390 0.126s ( 20,060 bytes)
274
- Task 204: conv_var 7.313 48029390 7.958s ( 212,060 bytes)
275
- Task 205: UNSOLVED 3.833s
276
- Task 206: conv_var 9.661 4589390 0.088s ( 20,060 bytes)
277
- Task 207: spatial_gather 12.982 165663 0.007s ( 11,115 bytes)
278
- Task 208: conv_fixed 8.868 10136430 4.659s ( 90,571 bytes)
279
- Task 209: UNSOLVED 10.051s
280
- Task 210: concat 12.982 165663 0.004s ( 11,113 bytes)
281
- Task 211: concat_enhanced 12.982 165663 0.022s ( 11,113 bytes)
282
- Task 212: conv_fixed 11.288 901462 0.107s ( 32,971 bytes)
283
- Task 213: UNSOLVED 1.472s
284
- Task 214: spatial_gather 12.982 165663 0.007s ( 11,115 bytes)
285
- Task 215: conv_var 7.915 26309390 1.406s ( 116,060 bytes)
286
- Task 216: UNSOLVED 3.590s
287
- Task 217: conv_fixed 10.778 1501550 0.275s ( 68,171 bytes)
288
- Task 218: UNSOLVED 0.475s
289
- Task 219: conv_fixed 9.452 5653862 1.743s ( 144,971 bytes)
290
- Task 220: conv_var 11.216 969390 0.031s ( 4,057 bytes)
291
- Task 221: UNSOLVED 30.512s
292
- Task 222: spatial_gather 12.982 165663 0.023s ( 11,115 bytes)
293
- Task 223: upscale 12.982 165663 0.007s ( 11,113 bytes)
294
- Task 224: conv_var 9.172 7485390 0.356s ( 32,860 bytes)
295
- Task 225: conv_fixed 13.123 143990 0.024s ( 10,568 bytes)
296
- Task 226: conv_fixed 10.583 1825462 0.307s ( 68,171 bytes)
297
- Task 227: conv_diff 12.690 221822 0.022s ( 20,306 bytes)
298
- Task 228: conv_fixed 12.345 313462 0.051s ( 10,568 bytes)
299
- Task 229: conv_fixed 14.178 50094 0.007s ( 4,168 bytes)
300
- Task 230: conv_var 11.216 969390 0.028s ( 4,057 bytes)
301
- Task 231: UNSOLVED 4.200s
302
- Task 232: conv_var 8.777 11105390 0.210s ( 48,860 bytes)
303
- Task 233: UNSOLVED 6.093s
304
- Task 234: conv_var 8.777 11105390 0.764s ( 48,860 bytes)
305
- Task 235: conv_diff 11.114 1072586 0.032s ( 68,306 bytes)
306
- Task 236: conv_diff 12.604 241742 0.022s ( 20,306 bytes)
307
- Task 237: conv_var 9.172 7485390 0.081s ( 32,860 bytes)
308
- Task 238: UNSOLVED 1.918s
309
- Task 239: UNSOLVED 1.157s
310
- Task 240: conv_fixed 7.755 30870190 20.204s ( 336,971 bytes)
311
- Task 241: transpose 14.509 36000 0.003s ( 127 bytes)
312
- Task 242: conv_diff 12.413 292586 0.012s ( 4,303 bytes)
313
- Task 243: conv_var 9.172 7485390 0.484s ( 32,860 bytes)
314
- Task 244: UNSOLVED 0.774s
315
- Task 245: conv_var 8.777 11105390 0.223s ( 48,860 bytes)
316
- Task 246: conv_var 7.915 26309390 1.546s ( 116,060 bytes)
317
- Task 247: UNSOLVED 0.935s
318
- Task 248: UNSOLVED 1.187s
319
- Task 249: UNSOLVED 1.262s
320
- Task 250: conv_fixed 11.288 901462 0.105s ( 32,971 bytes)
321
- Task 251: conv_var 9.172 7485390 0.120s ( 32,860 bytes)
322
- Task 252: conv_var 8.777 11105390 0.163s ( 48,860 bytes)
323
- Task 253: conv_diff 11.285 904082 0.017s ( 20,306 bytes)
324
- Task 254: conv_fixed 11.479 744750 0.078s ( 32,971 bytes)
325
- Task 255: UNSOLVED 113.706s
326
- Task 256: conv_var 9.172 7485390 0.092s ( 32,860 bytes)
327
- Task 257: conv_diff 11.948 465842 0.031s ( 20,306 bytes)
328
- Task 258: conv_var 11.216 969390 0.017s ( 4,057 bytes)
329
- Task 259: UNSOLVED 0.456s
330
- Task 260: conv_fixed 11.288 901462 0.162s ( 32,971 bytes)
331
- Task 261: conv_var 11.216 969390 0.014s ( 4,057 bytes)
332
- Task 262: UNSOLVED 0.639s
333
- Task 263: UNSOLVED 0.672s
334
- Task 264: UNSOLVED 3.545s
335
- Task 265: conv_fixed 8.157 20646614 10.371s ( 250,571 bytes)
336
- Task 266: conv_fixed 14.060 56382 0.011s ( 4,168 bytes)
337
- Task 267: conv_fixed 13.570 92014 0.015s ( 4,168 bytes)
338
- Task 268: conv_var 9.172 7485390 0.117s ( 32,860 bytes)
339
- Task 269: UNSOLVED 8.091s
340
- Task 270: spatial_gather 12.982 165663 0.020s ( 11,115 bytes)
341
- Task 271: conv_diff 12.537 258686 0.016s ( 10,703 bytes)
342
- Task 272: conv_var 11.216 969390 0.017s ( 4,057 bytes)
343
- Task 273: conv_fixed 10.303 2413462 0.397s ( 90,571 bytes)
344
- Task 274: UNSOLVED 0.848s
345
- Task 275: UNSOLVED 13.033s
346
- Task 276: color_map 13.252 126500 0.004s ( 549 bytes)
347
- Task 277: conv_fixed 10.583 1825462 0.343s ( 68,171 bytes)
348
- Task 278: conv_var 8.163 20517390 1.166s ( 90,460 bytes)
349
- Task 279: conv_var 7.915 26309390 2.053s ( 116,060 bytes)
350
- Task 280: conv_var 7.146 56717390 16.851s ( 250,460 bytes)
351
- Task 281: conv_var 8.777 11105390 0.480s ( 48,860 bytes)
352
- Task 282: conv_fixed 13.260 125550 0.016s ( 4,168 bytes)
353
- Task 283: conv_fixed 11.755 565462 0.059s ( 20,171 bytes)
354
- Task 284: conv_var 8.163 20517390 1.040s ( 90,460 bytes)
355
- Task 285: conv_var 8.447 15449390 3.120s ( 68,060 bytes)
356
- Task 286: conv_var 8.447 15449390 0.869s ( 68,060 bytes)
357
- Task 287: spatial_gather 12.982 165663 0.041s ( 11,115 bytes)
358
- Task 288: conv_var 9.661 4589390 0.050s ( 20,060 bytes)
359
- Task 289: UNSOLVED 9.816s
360
- Task 290: UNSOLVED 1.343s
361
- Task 291: UNSOLVED 0.330s
362
- Task 292: conv_var 6.993 66129390 0.889s ( 292,060 bytes)
363
- Task 293: conv_var 10.302 2417390 0.056s ( 10,457 bytes)
364
- Task 294: conv_fixed 13.112 145462 0.017s ( 4,168 bytes)
365
- Task 295: UNSOLVED 3.396s
366
- Task 296: conv_diff 13.151 140006 0.015s ( 10,703 bytes)
367
- Task 297: conv_var 8.777 11105390 0.091s ( 48,860 bytes)
368
- Task 298: conv_var 11.216 969390 0.024s ( 4,057 bytes)
369
- Task 299: conv_fixed 12.602 242390 0.029s ( 20,171 bytes)
370
- Task 300: UNSOLVED 0.750s
371
- Task 301: conv_var 10.302 2417390 0.039s ( 10,457 bytes)
372
- Task 302: conv_fixed 10.238 2575574 0.585s ( 68,171 bytes)
373
- Task 303: conv_var 8.163 20517390 3.015s ( 90,460 bytes)
374
- Task 304: UNSOLVED 4.951s
375
- Task 305: conv_fixed 10.882 1352950 0.213s ( 20,171 bytes)
376
- Task 306: conv_var 7.494 40065390 5.846s ( 176,860 bytes)
377
- Task 307: UNSOLVED 2.677s
378
- Task 308: UNSOLVED 1.412s
379
- Task 309: color_map 13.252 126500 0.003s ( 549 bytes)
380
- Task 310: UNSOLVED 2.443s
381
- Task 311: concat 12.982 165663 0.008s ( 11,113 bytes)
382
- Task 312: conv_fixed 10.564 1860374 0.266s ( 48,971 bytes)
383
- Task 313: conv_var 8.447 15449390 0.624s ( 68,060 bytes)
384
- Task 314: conv_fixed 12.142 383734 0.065s ( 20,171 bytes)
385
- Task 315: UNSOLVED 3.655s
386
- Task 316: conv_diff 12.363 307706 0.011s ( 10,703 bytes)
387
- Task 317: conv_fixed 13.260 125550 0.015s ( 4,168 bytes)
388
- Task 318: conv_diff 13.127 143342 0.017s ( 10,703 bytes)
389
- Task 319: UNSOLVED 1.183s
390
- Task 320: conv_var 9.172 7485390 0.096s ( 32,860 bytes)
391
- Task 321: conv_diff 12.259 341342 0.028s ( 20,306 bytes)
392
- Task 322: conv_fixed 14.178 50094 0.010s ( 4,168 bytes)
393
- Task 323: conv_fixed 9.804 3976174 0.771s ( 90,571 bytes)
394
- Task 324: conv_var 8.447 15449390 1.332s ( 68,060 bytes)
395
- Task 325: UNSOLVED 0.748s
396
- Task 326: varshape_spatial_gather 12.982 165663 0.034s ( 11,115 bytes)
397
- Task 327: UNSOLVED 1.620s
398
- Task 328: conv_var 7.146 56717390 4.966s ( 250,460 bytes)
399
- Task 329: conv_var 11.216 969390 0.021s ( 4,057 bytes)
400
- Task 330: conv_fixed 10.583 1825462 0.344s ( 68,171 bytes)
401
- Task 331: conv_fixed 13.112 145462 0.014s ( 4,168 bytes)
402
- Task 332: conv_var 7.693 32825390 0.301s ( 144,860 bytes)
403
- Task 333: conv_fixed 11.755 565462 0.098s ( 20,171 bytes)
404
- Task 334: conv_diff 13.354 114206 0.017s ( 10,703 bytes)
405
- Task 335: conv_var 8.447 15449390 0.502s ( 68,060 bytes)
406
- Task 336: conv_fixed 10.906 1321462 0.146s ( 48,971 bytes)
407
- Task 337: color_map 13.252 126500 0.003s ( 549 bytes)
408
- Task 338: conv_var 7.693 32825390 4.005s ( 144,860 bytes)
409
- Task 339: UNSOLVED 0.114s
410
- Task 340: conv_var 9.172 7485390 0.360s ( 32,860 bytes)
411
- Task 341: conv_fixed 11.288 901462 0.174s ( 32,971 bytes)
412
- Task 342: conv_fixed 11.288 901462 0.174s ( 32,971 bytes)
413
- Task 343: conv_fixed 10.571 1847262 0.297s ( 90,571 bytes)
414
- Task 344: conv_var 10.302 2417390 0.026s ( 10,457 bytes)
415
- Task 345: conv_fixed 10.906 1321462 0.146s ( 48,971 bytes)
416
- Task 346: UNSOLVED 0.349s
417
- Task 347: conv_diff 13.073 151346 0.020s ( 20,306 bytes)
418
- Task 348: conv_var 9.661 4589390 0.040s ( 20,060 bytes)
419
- Task 349: UNSOLVED 32.901s
420
- Task 350: conv_var 7.693 32825390 2.218s ( 144,860 bytes)
421
- Task 351: conv_diff 11.524 711914 0.024s ( 10,703 bytes)
422
- Task 352: conv_var 11.216 969390 0.017s ( 4,057 bytes)
423
- Task 353: conv_var 9.172 7485390 0.198s ( 32,860 bytes)
424
- Task 354: conv_fixed 11.288 901462 0.159s ( 32,971 bytes)
425
- Task 355: UNSOLVED 0.353s
426
- Task 356: conv_fixed 10.906 1321462 0.236s ( 48,971 bytes)
427
- Task 357: UNSOLVED 1.202s
428
- Task 358: conv_var 7.915 26309390 1.847s ( 116,060 bytes)
429
- Task 359: conv_var 10.302 2417390 0.075s ( 10,457 bytes)
430
- Task 360: conv_diff 12.439 285254 0.024s ( 10,703 bytes)
431
- Task 361: conv_fixed 12.345 313462 0.042s ( 10,568 bytes)
432
- Task 362: conv_fixed 11.755 565462 0.115s ( 20,171 bytes)
433
- Task 363: conv_fixed 11.288 901462 0.141s ( 32,971 bytes)
434
- Task 364: conv_var 8.163 20517390 1.431s ( 90,460 bytes)
435
- Task 365: UNSOLVED 0.935s
436
- Task 366: UNSOLVED 11.200s
437
- Task 367: conv_var 7.915 26309390 2.131s ( 116,060 bytes)
438
- Task 368: conv_fixed 11.288 901462 0.117s ( 32,971 bytes)
439
- Task 369: conv_fixed 10.906 1321462 0.212s ( 48,971 bytes)
440
- Task 370: conv_var 8.777 11105390 0.745s ( 48,860 bytes)
441
- Task 371: conv_var 8.777 11105390 0.195s ( 48,860 bytes)
442
- Task 372: conv_diff 11.586 669254 0.058s ( 20,306 bytes)
443
- Task 373: spatial_gather 12.982 165663 0.004s ( 11,115 bytes)
444
- Task 374: UNSOLVED 16.522s
445
- Task 375: conv_var 9.661 4589390 0.106s ( 20,060 bytes)
446
- Task 376: UNSOLVED 12.250s
447
- Task 377: UNSOLVED 2.078s
448
- Task 378: conv_var 8.777 11105390 0.359s ( 48,860 bytes)
449
- Task 379: conv_var 8.163 20517390 1.794s ( 90,460 bytes)
450
- Task 380: rotate 12.982 165663 0.005s ( 11,113 bytes)
451
- Task 381: conv_fixed 10.906 1321462 0.193s ( 48,971 bytes)
452
- Task 382: conv_var 8.163 20517390 1.191s ( 90,460 bytes)
453
- Task 383: conv_var 8.447 15449390 1.650s ( 68,060 bytes)
454
- Task 384: UNSOLVED 2.067s
455
- Task 385: spatial_gather 12.982 165663 0.009s ( 11,115 bytes)
456
- Task 386: conv_diff 12.787 201470 0.021s ( 20,306 bytes)
457
- Task 387: conv_var 8.447 15449390 1.863s ( 68,060 bytes)
458
- Task 388: UNSOLVED 2.976s
459
- Task 389: conv_var 11.216 969390 0.014s ( 4,057 bytes)
460
- Task 390: conv_fixed 9.809 3956462 1.066s ( 68,171 bytes)
461
- Task 391: UNSOLVED 0.395s
462
- Task 392: conv_fixed 11.288 901462 0.163s ( 32,971 bytes)
463
- Task 393: conv_diff 12.051 420578 0.012s ( 10,703 bytes)
464
- Task 394: UNSOLVED 0.389s
465
- Task 395: conv_diff 13.526 96146 0.015s ( 10,703 bytes)
466
- Task 396: UNSOLVED 2.566s
467
- Task 397: conv_fixed 11.755 565462 0.103s ( 20,171 bytes)
468
- Task 398: UNSOLVED 21.737s
469
- Task 399: UNSOLVED 1.155s
470
- Task 400: conv_diff 10.109 2931914 0.040s ( 20,306 bytes)
471
- wandb: β’Ώ updating run metadata (0.0s)
472
- wandb: β£» updating run metadata (0.0s)
473
- wandb: β£½ updating run metadata (0.0s)
474
- wandb: β£Ύ updating run metadata (0.0s)
475
- wandb: β£· uploading history steps 397-399, summary, console lines 397-399 (0.1s)
476
- wandb: β£― uploading history steps 397-399, summary, console lines 397-399 (0.1s)
477
- wandb:
478
- wandb: Run history:
479
- wandb: cost β–‚β–β–β–ƒβ–β–‚β–β–β–‚β–β–β–β–β–β–ˆβ–β–β–β–β–…β–β–β–β–‡β–β–β–ƒβ–β–β–‡β–β–β–β–β–β–‚β–β–β–‚β–
480
- wandb: macs β–‚β–β–ƒβ–β–β–ƒβ–β–β–ƒβ–‚β–β–β–β–β–‚β–…β–β–β–β–β–β–„β–β–β–β–„β–‡β–‚β–β–ˆβ–β–β–β–β–β–β–ƒβ–β–β–
481
- wandb: memory β–†β–‚β–…β–‚β–ƒβ–‚β–‚β–ƒβ–β–…β–†β–…β–†β–ƒβ–„β–…β–β–ˆβ–…β–†β–†β–‡β–…β–†β–β–…β–ƒβ–‚β–β–…β–†β–…β–β–β–‚β–‚β–β–‡β–‡β–
482
- wandb: onnx_bytes β–ˆβ–ƒβ–‚β–β–‚β–‚β–β–„β–†β–‚β–ˆβ–β–‚β–β–β–…β–β–β–‚β–β–β–‚β–β–‚β–β–ƒβ–…β–β–β–β–β–‡β–ƒβ–‚β–β–ƒβ–ƒβ–β–β–‚
483
- wandb: params β–β–‚β–‚β–‚β–ƒβ–‚β–β–β–ƒβ–β–‚β–β–ƒβ–β–‚β–…β–ƒβ–β–β–β–‚β–β–ˆβ–‚β–β–β–β–‚β–β–β–‚β–β–β–β–‚β–ƒβ–‚β–ƒβ–‚β–
484
- wandb: score β–‚β–β–ˆβ–ƒβ–‡β–…β–ƒβ–ƒβ–ƒβ–…β–ƒβ–ƒβ–ƒβ–‚β–ˆβ–‚β–ƒβ–ƒβ–‚β–ˆβ–†β–ˆβ–ˆβ–†β–ƒβ–‡β–†β–…β–ˆβ–ˆβ–ƒβ–ƒβ–„β–†β–ƒβ–‡β–ƒβ–ƒβ–ƒβ–†
485
- wandb: task_id β–β–β–β–‚β–‚β–‚β–ƒβ–ƒβ–ƒβ–ƒβ–ƒβ–„β–„β–„β–„β–„β–…β–…β–…β–…β–…β–…β–…β–…β–…β–…β–…β–†β–†β–†β–†β–†β–‡β–‡β–‡β–‡β–ˆβ–ˆβ–ˆβ–ˆ
486
- wandb: task_time_sec β–β–β–β–β–β–β–β–β–‚β–β–β–β–β–β–β–β–β–‚β–β–β–β–β–β–β–β–β–β–β–β–ˆβ–β–β–β–β–β–β–β–β–β–
487
- wandb:
488
- wandb: Run summary:
489
- wandb: cost 2931914
490
- wandb: macs 2822650
491
- wandb: memory 104338
492
- wandb: onnx_bytes 20306
493
- wandb: params 4926
494
- wandb: score 10.10883
495
- wandb: solver conv_diff
496
- wandb: task_id 400
497
- wandb: task_time_sec 0.04023
498
- wandb:
499
- wandb: πŸš€ View run solver_run at: https://wandb.ai/rogermt23/neurogolf/runs/20lzjuuc
500
- wandb: ⭐️ View project at: https://wandb.ai/rogermt23/neurogolf
501
- wandb: Synced 5 W&B file(s), 0 media file(s), 0 artifact file(s) and 0 other file(s)
502
- wandb: Find logs at: ./wandb/run-20260424_181206-20lzjuuc/logs
503
-
504
- ======================================================================
505
- Solved: 307/400 in 1073s
506
- conv_var: 125
507
- conv_fixed: 107
508
- conv_diff: 39
509
- spatial_gather: 16
510
- concat: 5
511
- color_map: 4
512
- concat_enhanced: 4
513
- rotate: 3
514
- transpose: 2
515
- upscale: 1
516
- varshape_spatial_gather: 1
517
-
518
- 365 ONNX files, Total local estimated score: 3267.0 total 33385.0 KB
519
- Created submission.zip
520
-
521
- ## Mjor Issues
522
- i submited the notebook with LB score:
523
- Latest Score
524
- 501.42 V10
525
-
526
- Best Score
527
- 501.42 V10
528
-
529
- Daily Submissions
530
- 1 / 100 used
531
- We nned check submission file process
532
- ONNX network being s4ved
533
- I h4ve 4ttched notebooks with LB 4000+ (neurogolf-2026-solver-notebooks.zip)
534
- use neurogolf-2026-solver-notebooks.zip so you c4n compre why the diff between my LS (using offi4l K4ggle utils lib) 3267 & 501.42
535
-
536
- I hve 4tched ARC-GEN d4t4set ARC-GEN-100K.zip
537
- until you figure why submissions is re4lly low not in step with the LB
538
-
539
-
540
- (i noticed you r not usiiing the v4lidtef netwrok from utils m4ybe there iis iissues with network)
541
- the submission fiile needs to veryfied before it submitted use these from utils lib i supplied
542
- ef verify_network(network, task_num, examples):
543
- filename = "task{:03d}.onnx".format(task_num)
544
- onnx.save(network, filename)
545
- if not check_network(filename): return
546
- try:
547
- session = onnxruntime.InferenceSession(filename)
548
- except onnxruntime.ONNXRuntimeError as e:
549
- print(f"Error: Unable to load ONNX model: {e}")
550
- return
551
- arc_agi_right, arc_agi_wrong, arc_agi_expected = verify_subset(session, examples["train"] + examples["test"])
552
- arc_gen_right, arc_gen_wrong, arc_gen_expected = verify_subset(session, examples["arc-gen"])
553
- print(f"Results on ARC-AGI examples: {arc_agi_right} pass, {arc_agi_wrong} fail")
554
- print(f"Results on ARC-GEN examples: {arc_gen_right} pass, {arc_gen_wrong} fail")
555
- print()
556
- macs, memory, params = score_network(filename)
557
- if macs is None or memory is None or params is None:
558
- print("Error: Your network performance could not be measured")
559
- elif arc_agi_wrong + arc_gen_wrong == 0:
560
- print("Your network IS READY for submission!")
561
- print()
562
- print("Performance stats:")
563
- onnx_tool.model_profile(filename)
564
- points = max(1.0, 25.0 - math.log(macs + memory + params))
565
- print()
566
- print(f"It appears to require {macs} MACs + {memory} bytes + {params} params, yielding {points:.3f} points.")
567
- print()
568
- print("Next steps:")
569
- print(f" * Click the link below to download {filename} onto your local machine.")
570
- print(" * Create a zip file containing that network along with all others.")
571
- print(" * Submit that zip file to the Kaggle competition so that it can be officially scored.")
572
- print()
573
- display(FileLink(filename))
574
- else:
575
- print("Your network IS NOT ready for submission.")
576
- expected = None
577
- expected = arc_agi_expected if arc_agi_expected is not None else expected
578
- expected = arc_gen_expected if arc_gen_expected is not None else expected
579
- if expected is None: return
580
- benchmark = convert_to_numpy(expected)
581
- actual = {}
582
- actual["input"] = expected["input"]
583
- actual["output"] = convert_from_numpy(run_network(session, benchmark["input"]))
584
- print("The expected result is shown in green; your actual result is shown in red.")
585
- show_examples([expected], bgcolor=(200, 255, 200))
586
- show_examples([actual], bgcolor=(255, 200, 200))
587
-
588
-
589
- def verify_subset(session, example_subset):
590
- right, wrong, expected, error = 0, 0, None, ""
591
- for example in example_subset:
592
- benchmark = convert_to_numpy(example)
593
- if not benchmark: continue
594
- try:
595
- user_output = run_network(session, benchmark["input"])
596
- if np.array_equal(user_output, benchmark["output"]):
597
- right += 1
598
- else:
599
- expected = example
600
- wrong += 1
601
- except onnxruntime.ONNXRuntimeError:
602
- error = traceback.format_exc()
603
- wrong += 1
604
- if error: print(f"Error: {error}")
605
- return right, wrong, expected
606
 
607
  ## Quick Start
608
 
609
  ```bash
610
- # 1. Clone
611
  git clone https://huggingface.co/rogermt/neurogolf-solver
612
  cd neurogolf-solver
613
 
614
- # 2. Install deps
615
  pip install numpy onnx onnxruntime
616
 
617
- # 3. Get ARC data
618
  git clone --depth 1 https://github.com/fchollet/ARC-AGI.git
619
 
620
- # 4. Run
621
- python neurogolf_solver.py --data_dir ARC-AGI/data/training/ --output_dir submission --conv_budget 15
622
 
623
- # 5. Results
624
- ls submission/*.onnx | wc -l
625
- ```
626
 
627
- ## Create submission.zip for Kaggle
628
- ```python
629
- import zipfile, os
630
- with zipfile.ZipFile('submission.zip', 'w', zipfile.ZIP_DEFLATED) as zf:
631
- for f in sorted(os.listdir('submission')):
632
- if f.endswith('.onnx'):
633
- zf.write(os.path.join('submission', f), f)
634
- print(f"Created submission.zip: {os.path.getsize('submission.zip')/1024:.0f} KB")
635
  ```
636
 
637
- ## Key Parameters
638
 
639
  | Flag | Default | Description |
640
  |------|---------|-------------|
641
- | `--conv_budget` | `30` | Seconds per task for conv solver. More = more tasks solved |
642
  | `--data_dir` | `ARC-AGI/data/training/` | Path to task JSONs |
643
- | `--output_dir` | `submission` | Where to save .onnx files |
644
- | `--kaggle` | off | Use Kaggle task format (task001.json) |
 
 
645
  | `--tasks` | all | Comma-separated task numbers (e.g., `1,2,3`) |
 
646
  | `--use_wandb` | off | Enable W&B logging |
647
 
648
  ## How It Works
649
 
650
- **Format:** Input/output = `[1, 10, 30, 30]` one-hot float32. ONNX opset 10, IR version 10.
651
 
652
- **Solver pipeline:**
653
- 1. **Analytical solvers** (instant, zero-cost): identity, constant, color_map, transpose, flip, rotate, tile, upscale, kronecker, concat, concat_enhanced, diagonal_tile, spatial_gather, varshape_spatial_gather
654
- 2. **Conv solvers** (learned via least-squares):
655
- - Fixed-shape: `Slice β†’ Conv β†’ ArgMax β†’ Equal+Cast β†’ Pad`
656
- - Variable-shape: `Conv(30Γ—30) β†’ ArgMax β†’ Equal+Cast β†’ Mul(mask)`
657
- - Diff-shape: `Slice β†’ Conv β†’ Slice(crop) β†’ ArgMax β†’ Equal+Cast β†’ Pad`
658
- - Variable diff-shape: `Conv(30Γ—30) β†’ ArgMax β†’ Equal+Cast β†’ Mul(input_mask)`
659
 
660
- **Key design decisions:**
661
- - Uses `Gather` instead of `GatherElements` for opset 10 compatibility
662
- - Uses `Equal + Cast` instead of `OneHot` (avoids CUDA kernel issues)
663
- - CPU-only inference (GPU has no benefit for tiny 30Γ—30 grids)
664
- - Least-squares fitting finds optimal conv weights analytically (no gradient descent)
665
 
666
- ## What's NOT solved yet (94 tasks)
667
 
668
- - Tasks with input-driven output structure (output layout depends on input content)
669
- - Tasks requiring multi-step reasoning (flood fill, gravity, counting)
670
- - Variable diff-shape tasks where output extends beyond input bounds
671
- - Tasks needing very large conv kernels (>29Γ—29)
 
 
 
 
 
 
672
 
673
  ## Scoring
674
 
675
- NeuroGolf scoring: `Score = MACs + memory_bytes + params`
676
- - Analytical solvers β†’ near-zero cost
677
- - Conv solvers β†’ cost proportional to kernel size
678
- - Lower score = better
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
679
 
680
  ## Repo
 
681
  https://huggingface.co/rogermt/neurogolf-solver
 
1
+ # NeuroGolf Solver v5
2
+
3
+ Builds minimal ONNX networks for ARC-AGI tasks. Modular Python package with opset 17, zero-cost Slice-based transforms.
4
+
5
+ **Currently running on Kaggle β€” results pending.**
6
+
7
+ ## Version History
8
+
9
+ | Version | Date | Solved (local) | Arc-gen Validated | Est LB | Key Changes |
10
+ |---------|------|----------------|-------------------|--------|-------------|
11
+ | **v5** | **2026-04-26** | **TBD** | **TBD** | **TBD** | Refactored to package, opset 17, Slice-based flip/rotate (0 MACs), lstsq crash fix, tensor-based Pad & ReduceSum |
12
+ | v4.3 | 2026-04-25 | 307 | 50 | ~670 | Methodology docs, no code changes |
13
+ | v4.0 | 2026-04-24 | 307 | 50 | ~656 | ARC-GEN validation, static profiler |
14
+ | v3 | 2026-04-24 | 307 | ~40 | 501 | concat_enhanced, varshape_spatial_gather |
15
+ | v2 | prior | 294 | β€” | β€” | Spatial_gather, variable-shape conv |
16
+ | v1 | prior | 128 | β€” | β€” | Conv solver only |
17
+
18
+ ## Project Structure
19
+
20
+ ```
21
+ neurogolf_solver/ # Python package (v5)
22
+ β”œβ”€β”€ __init__.py # Package marker
23
+ β”œβ”€β”€ config.py # Runtime config (providers, opset)
24
+ β”œβ”€β”€ constants.py # All constants (grid dims, excluded tasks, limits)
25
+ β”œβ”€β”€ data_loader.py # Task loading, one-hot encoding, example extraction
26
+ β”œβ”€β”€ gather_helpers.py # Gather-based ONNX model builders
27
+ β”œβ”€β”€ main.py # Entry point with W&B init
28
+ β”œβ”€β”€ onnx_helpers.py # Opset 17 builders (Slice, Pad, ReduceSum, mk)
29
+ β”œβ”€β”€ profiler.py # Static cost profiler (fallback for onnx_tool)
30
+ β”œβ”€β”€ submission.py # run_tasks with W&B logging, zip/csv generation
31
+ β”œβ”€β”€ validators.py # Model validation against train+test+arc-gen
32
+ └── solvers/
33
+ β”œβ”€β”€ __init__.py # Exports solve_task, ANALYTICAL_SOLVERS
34
+ β”œβ”€β”€ analytical.py # identity, constant, color_map, transpose
35
+ β”œβ”€β”€ conv.py # lstsq conv solvers (fixed, variable, diffshape, var_diff)
36
+ β”œβ”€β”€ geometric.py # flip, rotate, shift, crop, gravity
37
+ β”œβ”€β”€ solver_registry.py # Solver ordering + solve_task orchestration
38
+ └── tiling.py # tile, upscale, mirror, concat, spatial_gather
39
+
40
+ neurogolf_solver.py # Legacy monolith (v4, kept for reference)
41
+ neurogolf_utils.py # Official Kaggle scoring library
42
+ ARC-GEN-100K.zip # 400 files Γ— ~250 examples synthetic data
43
+ neurogolf-2026-solver-notebooks.zip # 5 reference notebooks (LB 4000+)
44
+ ```
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
45
 
46
  ## Quick Start
47
 
48
  ```bash
49
+ # Clone
50
  git clone https://huggingface.co/rogermt/neurogolf-solver
51
  cd neurogolf-solver
52
 
53
+ # Install deps
54
  pip install numpy onnx onnxruntime
55
 
56
+ # Get ARC data
57
  git clone --depth 1 https://github.com/fchollet/ARC-AGI.git
58
 
59
+ # Run (local)
60
+ python -m neurogolf_solver.main --data_dir ARC-AGI/data/training/ --output_dir submission --conv_budget 30
61
 
62
+ # Run (Kaggle)
63
+ python -m neurogolf_solver.main --kaggle --data_dir /kaggle/input/competitions/neurogolf-2026/ --output_dir /kaggle/working/submission --conv_budget 60
 
64
 
65
+ # With ARC-GEN data and W&B logging
66
+ python -m neurogolf_solver.main --data_dir ARC-AGI/data/training/ --arcgen_dir ARC-GEN-100K/ --output_dir submission --use_wandb
 
 
 
 
 
 
67
  ```
68
 
69
+ ## Parameters
70
 
71
  | Flag | Default | Description |
72
  |------|---------|-------------|
 
73
  | `--data_dir` | `ARC-AGI/data/training/` | Path to task JSONs |
74
+ | `--arcgen_dir` | `` | Path to ARC-GEN-100K/ directory |
75
+ | `--output_dir` | `/kaggle/working/submission` | Where to save .onnx files |
76
+ | `--kaggle` | off | Use Kaggle task format (task001.json with embedded arc-gen) |
77
+ | `--conv_budget` | `30` | Seconds per task for conv solver |
78
  | `--tasks` | all | Comma-separated task numbers (e.g., `1,2,3`) |
79
+ | `--device` | `auto` | `auto`, `cpu`, or `cuda` |
80
  | `--use_wandb` | off | Enable W&B logging |
81
 
82
  ## How It Works
83
 
84
+ **Format:** Input/output = `[1, 10, 30, 30]` one-hot float32. ONNX opset 17, IR version 8.
85
 
86
+ **Solver pipeline (in order):**
87
+ 1. **Analytical solvers** (instant, near-zero cost):
88
+ identity β†’ constant β†’ color_map β†’ transpose β†’ flip β†’ rotate β†’ shift β†’ tile β†’ upscale β†’ kronecker β†’ nonuniform_scale β†’ mirror_h β†’ mirror_v β†’ quad_mirror β†’ concat β†’ concat_enhanced β†’ diagonal_tile β†’ fixed_crop β†’ spatial_gather β†’ varshape_spatial_gather
 
 
 
 
89
 
90
+ 2. **Conv solvers** (learned via least-squares, validated against arc-gen):
91
+ - `conv_fixed` β€” Slice β†’ Conv β†’ ArgMax β†’ Equal+Cast β†’ Pad
92
+ - `conv_variable` β€” Conv(30Γ—30) β†’ ArgMax β†’ Equal+Cast β†’ Mul(mask)
93
+ - `conv_diffshape` β€” Slice β†’ Conv β†’ Slice(crop) β†’ ArgMax β†’ Equal+Cast β†’ Pad
94
+ - `conv_var_diff` β€” Conv(30Γ—30) β†’ ArgMax β†’ Equal+Cast β†’ Mul(input_mask)
95
 
96
+ ## v5 Changes from v4
97
 
98
+ | Change | Impact |
99
+ |--------|--------|
100
+ | **Opset 10 β†’ 17, IR 10 β†’ 8** | Enables Slice(step=-1) for zero-cost transforms |
101
+ | **s_flip: Slice(step=-1)** | 0 MACs (was ~165K with Gather) |
102
+ | **s_rotate k=2: double Slice reverse** | 0 MACs (was ~165K) |
103
+ | **s_rotate k=1,3 (square): Slice+Transpose** | 0 MACs (was ~165K) |
104
+ | **All Pad nodes: tensor-based pads input** | Required for opset 17 compatibility |
105
+ | **All ReduceSum nodes: axes as tensor input** | Required for opset 13+ compatibility |
106
+ | **lstsq crash fix: try/except LinAlgError** | Prevents SVD non-convergence crash (task 313) |
107
+ | **Refactored to 16-file package** | Maintainable, testable, no more monolith |
108
 
109
  ## Scoring
110
 
111
+ ```
112
+ Score per task = max(1.0, 25.0 - ln(MACs + memory_bytes + params))
113
+ ```
114
+
115
+ - Analytical solvers (Slice/Transpose/Gather) β†’ near-zero cost β†’ ~20-25 pts
116
+ - Conv solvers β†’ cost proportional to kernel size β†’ ~7-14 pts
117
+ - Unsolved β†’ 1.0 pt minimum
118
+
119
+ ## Competition Rules
120
+
121
+ | Item | Value |
122
+ |------|-------|
123
+ | Input/Output | float32 `[1,10,30,30]` one-hot |
124
+ | Opset | 10 or 17 (both accepted on Kaggle) |
125
+ | Max file size | 1.44 MB per model |
126
+ | Banned ops | Loop, Scan, NonZero, Unique, Script, Function |
127
+ | Excluded tasks | {21, 55, 80, 184, 202, 366} |
128
+ | Validation | Models checked against train + test + arc-gen (ALL splits) |
129
+
130
+ ## Key Docs
131
+
132
+ - **SKILL.md** β€” Competition rules, architecture, methodology, checklist
133
+ - **LEARNING.md** β€” All mistakes, research findings, what works/doesn't
134
+ - **TODO.md** β€” Roadmap, experiment queue, status tracking
135
+
136
+ ## Strategy
137
+
138
+ We build our own solver. No blending. No public datasets. See LEARNING.md for competitive intelligence on what others do (for awareness only).
139
 
140
  ## Repo
141
+
142
  https://huggingface.co/rogermt/neurogolf-solver