Update README.md

README.md

@@ -1,3 +1,91 @@
----
-license: mit
----
+---
+license: mit
+task_categories:
+- image-to-text
+- text-to-image
+language:
+- en
+pretty_name: simons ARC (abstraction & reasoning corpus) task version 11
+size_categories:
+- 10K<n<100K
+configs:
+  - config_name: default
+    data_files:
+      - split: train
+        path: data.jsonl
+---
+
+# Version 1
+
+Have dataset items that are somewhat evenly of each type. The LLM validation loss went down and then continue to rise afterwards,
+so I guess the complexity of the dataset was too high.
+The image sizes are between 1 and 10 pixels.
+
+# Version 2
+
+Here the majority of dataset items are histograms. 
+Smaller images. Here the image sizes are between 1 and 5 pixels.
+Let's see if the LLM does better on this one.
+
+# Version 3
+
+Focus on pair comparisons finding color intersection and color union.
+
+# Version 4
+
+image size 3-10.
+This converged slowly. I suspect the bigger image size was too optimistic. Next time try with a smaller image size.
+
+# Version 5
+
+image size 3-8.
+Added `pair_histogram_add`, `pair_histogram_max`, `pair_histogram_min`.
+This converged slowly. I suspect the big image size is still too optimistic. Next time try with an even smaller image size.
+
+# Version 6
+
+I'm trying even smaller images.
+
+image width 1-5, this is what it has been trained on in the past.
+
+image height 3-8, this is slightly bigger than what it has been trained on.
+
+This did the LLM make sense of quite fast. Next I'm goint to try with a wider image.
+
+# Version 7
+
+image width 3-8, this is slightly bigger than what it has been trained on, and something that the model has struggled learning.
+
+image height 1-5, this is what it has been trained on in the past.
+
+# Version 8
+
+image size 1-8.
+
+I have been training on image pairs, so the model is now somewhat ok at making sense on image pairs.
+So I hope it will be better at processing tasks, since a task is a list of image pairs.
+
+# Version 9
+
+image size 1-10.
+
+same weight to all transformations.
+
+Making use of TaskFormatterRLECompact` for the first time.
+Oh, the number of tokens per row is too high. I'm going to use smaller images.
+
+# Version 10
+
+image size 1-5.
+
+same weight to all transformations.
+
+Making use of TaskFormatterRLECompact`.
+
+# Version 11
+
+image size 1-5.
+
+same weight to all transformations.
+
+Making use of TaskFormatterRLECompact`.