Pieces/embeddinggemma-300m-distilled-width10pct-128dim-best

Distilled Backbone: embeddinggemma-300m-distilled-10pct-128dim

This is a distilled/compressed version of google/embeddinggemma-300m.

Compression Details

• Base model: google/embeddinggemma-300m
• Width reduction factor: 0.1
• Target hidden size: 72
• Final embedding dimension: 128
• Had projection layer: True
• Projection incorporated: True

Usage

To use this as a pretrained backbone in validation:

python -m playground.validate_from_checkpoint \
    --pretrained_backbone_model "Pieces/embeddinggemma-300m-distilled-width10pct-128dim-best" \
    --backbone_embedding_dim 128 \
    --backbone_pooling_mode "mean" \
    --num_samples 100 \
    --dataset_split "val" \
    --num_general_tags_high 100 \
    --num_general_tags_low 100

Important: The path must point to the sentence_transformer subdirectory (not the parent directory).

The exported model has been configured to output embeddings of dimension 128 directly, incorporating any projection layers from the distillation process.