queryn-adapter-me5-large_to_bge-m3

Queryn adapter — me5-large → bge-m3

Translates an embedding produced by me5-large into the embedding space of bge-m3, so a corpus already embedded with me5-large can be served against a bge-m3 index without re-embedding it. Part of the Queryn embedding-translation engine.

Specs

Architecture ablation (best test cosine): linear 0.8918 ← saved, deep 0.8825.

Input / output contract

• Input source_embedding — float32, shape [batch, 1024]. Raw me5-large embeddings; the graph L2-normalizes them itself, so pre-normalization is neither required nor harmful.
• Output target_embedding — float32, shape [batch, 1024], unit-normalized, in bge-m3 space.
• Batch axis is dynamic.

Usage

import numpy as np, onnxruntime as ort
from huggingface_hub import hf_hub_download

path = hf_hub_download("QuerynAi/queryn-adapter-me5-large_to_bge-m3", "model.onnx")
sess = ort.InferenceSession(path, providers=["CPUExecutionProvider"])

src = np.random.rand(4, 1024).astype(np.float32)   # your me5-large embeddings
tgt = sess.run(["target_embedding"], {"source_embedding": src})[0]
assert tgt.shape == (4, 1024)                     # unit vectors in bge-m3 space

Training

Trained on paired embeddings over a unified multi-domain corpus — arXiv abstracts, Australian case law, SQuAD passages, PubMed abstracts, and crypto/markets news (~350k rows spanning science, legal, QA, medical, and finance). Loss: 1 - mean cosine similarity, Adam, ReduceLROnPlateau, best-epoch checkpoint. Both a linear baseline and the MLP are trained for every pair; the higher-scoring one is published (ties go to linear).

Plots

me5-large → all targets: learning curves and best scores (this pair included).

Linear vs. deep for every pair (black ring = saved architecture).

Full adapter set: Queryn Embedding Adapters

Provenance

• Source checkpoint: models/v1/me5-large_to_bge-m3.pt (sha256 484d8a67f2740752…)
• Converted: 2026-08-30T20:45:22+00:00 · torch 2.13.0 · ptConverter.py

License

Released under the MIT license.