Datasets:
Update README
Browse files
README.md
CHANGED
|
@@ -31,12 +31,17 @@ Each composite crest is paired with:
|
|
| 31 |
- an English translation,
|
| 32 |
- a non-linguistic program code over the generator factors.
|
| 33 |
|
| 34 |
-
Because every crest is synthesized from a known triple of generator factors
|
| 35 |
-
container `C`, modifier `R`, motif `M`
|
| 36 |
metrics, controlled factor-pair recombination splits, counterfactual motif-
|
| 37 |
-
sensitivity tests under fixed container
|
| 38 |
-
of factor accessibility from frozen representations. See the
|
| 39 |
-
paper for details and baselines.
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 40 |
|
| 41 |
## Quick start
|
| 42 |
|
|
@@ -103,8 +108,8 @@ Each Croissant record in the `images` record set has these fields:
|
|
| 103 |
For program-label experiments, the same images are paired with non-linguistic
|
| 104 |
codes for the container (`C:NNN`), modifier (`X:N`), and motif (`M:NNN`); the
|
| 105 |
three `*_holdout.json` Croissant variants reassign splits so that whole factor
|
| 106 |
-
combinations
|
| 107 |
-
the underlying primitive tokens still appear individually in training.
|
| 108 |
|
| 109 |
## Recombination splits
|
| 110 |
|
|
@@ -132,9 +137,8 @@ crests.
|
|
| 132 |
|
| 133 |
The dataset is released under
|
| 134 |
[CC BY-NC 4.0](https://creativecommons.org/licenses/by-nc/4.0/);
|
| 135 |
-
see `LICENSE.txt` for the full legal code. The companion code
|
| 136 |
-
|
| 137 |
-
during anonymous review) is released under the MIT License.
|
| 138 |
|
| 139 |
The component images bundled with KamonBench (one isolated motif per
|
| 140 |
composite and one container per contained composite) are repackaged in PNG
|
|
|
|
| 31 |
- an English translation,
|
| 32 |
- a non-linguistic program code over the generator factors.
|
| 33 |
|
| 34 |
+
Because every crest is synthesized from a known triple of generator factors
|
| 35 |
+
(container `C`, modifier `R`, motif `M`), KamonBench supports direct factor
|
| 36 |
metrics, controlled factor-pair recombination splits, counterfactual motif-
|
| 37 |
+
sensitivity tests under fixed (container, modifier) contexts, and linear
|
| 38 |
+
probes of factor accessibility from frozen representations. See the
|
| 39 |
+
accompanying paper for details and baselines.
|
| 40 |
+
|
| 41 |
+
The companion code (package, training and evaluation pipelines, and the
|
| 42 |
+
generator) lives at
|
| 43 |
+
[`anon-submittere/KamonBench`](https://github.com/anon-submittere/KamonBench)
|
| 44 |
+
during anonymous review.
|
| 45 |
|
| 46 |
## Quick start
|
| 47 |
|
|
|
|
| 108 |
For program-label experiments, the same images are paired with non-linguistic
|
| 109 |
codes for the container (`C:NNN`), modifier (`X:N`), and motif (`M:NNN`); the
|
| 110 |
three `*_holdout.json` Croissant variants reassign splits so that whole factor
|
| 111 |
+
combinations (`(C, M)`, `(R, M)`, or `(C, R, M)`) are absent from training,
|
| 112 |
+
while the underlying primitive tokens still appear individually in training.
|
| 113 |
|
| 114 |
## Recombination splits
|
| 115 |
|
|
|
|
| 137 |
|
| 138 |
The dataset is released under
|
| 139 |
[CC BY-NC 4.0](https://creativecommons.org/licenses/by-nc/4.0/);
|
| 140 |
+
see `LICENSE.txt` for the full legal code. The companion code is released
|
| 141 |
+
under the MIT License.
|
|
|
|
| 142 |
|
| 143 |
The component images bundled with KamonBench (one isolated motif per
|
| 144 |
composite and one container per contained composite) are repackaged in PNG
|