Update README.md
Browse files
README.md
CHANGED
|
@@ -70,17 +70,6 @@ The pipeline is the artifact. Every benchmark image is produced in two determini
|
|
| 70 |
|
| 71 |
The released benchmark contains only the Stage-2 images. Held-out splits with unpublished seeds and additional difficulties can be regenerated from the included generators.
|
| 72 |
|
| 73 |
-
## Anti-shortcut design
|
| 74 |
-
|
| 75 |
-
Every task is constructed to defeat six shortcut classes (full discussion in `code/scope.md`):
|
| 76 |
-
|
| 77 |
-
1. One-shot symbolic perception, then text reasoning.
|
| 78 |
-
2. End-to-end CV scripting (OCR / edge / blob / vector tracing).
|
| 79 |
-
3. Prior leakage / answer-distribution memorisation.
|
| 80 |
-
4. Gestalt heuristics that bypass step-by-step tracing.
|
| 81 |
-
5. Memorisation of the released benchmark (mitigated because the *pipeline* is released, not just the data).
|
| 82 |
-
6. Tool-using zoom/crop/OCR loops.
|
| 83 |
-
|
| 84 |
## Responsible AI
|
| 85 |
|
| 86 |
See `croissant.json` for the full RAI block. Headlines:
|
|
|
|
| 70 |
|
| 71 |
The released benchmark contains only the Stage-2 images. Held-out splits with unpublished seeds and additional difficulties can be regenerated from the included generators.
|
| 72 |
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 73 |
## Responsible AI
|
| 74 |
|
| 75 |
See `croissant.json` for the full RAI block. Headlines:
|