static WEIGHTS_BYTES: &[u8]Expand description
Bundled sparse-linear projection weights for BGE-M3 sparse embedding.
§Provenance
Extracted from the fastembed-rs crate (v4) which bundles the same file
from the BAAI/bge-m3 checkpoint. The weights implement the sparse-linear
layer described in the BGE-M3 paper: a single linear projection
hidden_size → 1 that maps each token’s 1024-d hidden state to a scalar
relevance score, followed by ReLU activation and max-pooling by vocab ID.
- Source checkpoint:
BAAI/bge-m3(HF commit5617a9f61b02800) - Tensors:
weightshape[1024](F32),biasscalar (F32) - File SHA-256:
a2601321f01abbb696d171a58a65ff35be1603d9cbc22c647dfe34b4568dd690 - File size: 4,236 bytes