|
--- |
|
license: apache-2.0 |
|
datasets: |
|
- HuggingFaceTB/cosmopedia |
|
--- |
|
An untrained precursor MoE created from Cosmo using mergekit. |
|
|
|
Gate routing initialized using prompt hidden state method. Five are based on the visualized topic clusters of Cosmopedia data, three are task-oriented. |
|
|
|
Degenerate layers are 0, 1, and 2 (I believe this means experts will be underutilized for the lowest-level features). |
|
Best I could do with test-and-try prompt-based routing. |
|
Further research might start from the reversed direction, if available in some interpretability tool (activating layer into prompts). |
