Qdrant
/

constella-zero

Feature Extraction

zero-lookup-table

asymmetric-dual-encoder

research-preview

8-bit precision

Model card Files Files and versions

feat: add static-embeddings like F16 safetensors lookup table

#1

by itsclelia - opened 8 days ago

base: refs/heads/main

←

from: refs/pr/1

Discussion Files changed

feat: add static-embeddings like F16 safetensors lookup table2ac960da

Qdrant org 8 days ago

Hugging Face CLI

No description provided.

chore: conversion and test script, make special tokens in ZeroEncoder configurablef3bea3cc

fix: make it usable by fastembed1365b4aa

Ship int8 safetensors with per-row scales for FastEmbed8312f3fb

Qdrant org 7 days ago

FastEmbed side: https://github.com/qdrant/fastembed/pull/832. Merge this PR first, then that one right after.

Pushed 8312f3f:

model.safetensors now holds the released int8 rows as embeddings and their per-row scales as weights (30 MB). FastEmbed's output matches the released encoder exactly.
zero_encoder.py is back to the released version. FastEmbed now handles special tokens and sqrt pooling itself.
The test script checks the file against the released encoder, with special tokens on.

jmzzomg changed pull request status to merged about 11 hours ago

Upload images, audio, and videos by dragging in the text input, pasting, or clicking here.

Tap or paste here to upload images

· Sign up or log in to comment