IC-301Yi-1.5-9B-chat exhibits a layer-repetition pattern where a section of layers is duplicated at a later depth

Zeping Min, Xinshang Wang

SourceDOCS: Quantifying Weight Similarity for Deeper Insights into Large Language Models

Applying DOCS to the MLP-up weight matrices of 01-ai/yi-1.5-9b-chat reveals a specific off-diagonal pattern: layer 9 is highly similar to layer 25, layer 10 to layer 26, layer 11 to layer 27, and so on. This suggests a repetition of a section of layers within the model, possibly due to a specific training strategy employed to save training costs. In contrast, representational similarity (Linear CKA) on the same model shows relatively homogeneous patterns that do not reveal this fine-grained layer correspondence.

Evidence
observational
Key metric
layer 9 is highly similar to layer 25, layer 10 is highly similar to layer 26, layer 11 is highly similar to layer 27
Caveat
The authors hypothesize the cause (training strategy to save costs) but do not confirm it; the finding is based on a single model.
Model
Yi Yi-1.5-9B
Concepts
Depth-dependent structure
Methods
Linear CKA [compared-to]
Related findings
IC-298, IC-299, IC-300
Extraction
automatic-extraction