Modelpublic

statelet/gte-base-en-v1.5

sourceHugging Faceapache-2.0updated 17d agoView on Hugging Face

0likes110downloads

Model Card

gte-base-en-v1.5 — ONNX build for Statelet

ONNX export of `Alibaba-NLP/gte-base-en-v1.5`, packaged for the dense-retrieval path in Statelet. Weights are unmodified; this repository only re-packages them for ONNX Runtime.

Files

file	precision	notes
`model.onnx`	fp32	default; loaded on the CPU execution provider
`model_gpu_b8_l256_fp16.onnx`	fp16	static shapes, batch 8 × seq 256, for GPU providers
`tokenizer.json`	—	HuggingFace fast tokenizer
`config.json`, `pooling_config.json`	—	pooling is CLS, not mean

Model details

—Embedding dimension 768, 12 layers, vocab 30528, max position 8192.
—Pooling: CLS token (pooling_config.json). Mean-pooling this model gives wrong vectors.

Use in Statelet

STATELETMODELROOT=/absolute/path/to/models

with this repository placed at $STATELET_MODEL_ROOT/gte-base-en-v1.5, or point at it directly with STATELET_EMBEDDING_MODEL.

Changing the embedder changes the vector dimension, so every existing graph must be re-ingested when switching away from a 768-dim model.

License

Apache-2.0, inherited from the upstream model. See the upstream card for the authoritative terms and citation.