Update README.md
Browse files
README.md
CHANGED
@@ -5,4 +5,8 @@ datasets:
|
|
5 |
---
|
6 |
An untrained precursor MoE created from Cosmo using mergekit.
|
7 |
|
8 |
-
Gate routing initialized using prompt hidden state method. Five are based on the visualized topic clusters of Cosmopedia data, three are task-oriented.
|
|
|
|
|
|
|
|
|
|
5 |
---
|
6 |
An untrained precursor MoE created from Cosmo using mergekit.
|
7 |
|
8 |
+
Gate routing initialized using prompt hidden state method. Five are based on the visualized topic clusters of Cosmopedia data, three are task-oriented.
|
9 |
+
|
10 |
+
Degenerate layers are 0, 1, and 2 (I believe this means experts will be underutilized for the lowest-level features).
|
11 |
+
Best I could do with test-and-try prompt-based routing.
|
12 |
+
Further research might start from the reversed direction, if available in some interpretability tool (activating layer into prompts).
|