DeeAxe/business-entity-matching-encoders

Business-entity matching encoders

Eight fine-tuned cross-encoders for scoring whether two business records (name + address) refer to the same organisation. All members live in this single repository as state_dict files (<tag>/model.pt).

They are loads for Hugging Face sequence-classification heads built from intfloat/multilingual-e5-small or intfloat/multilingual-e5-base (MIT). Pair text is "{name} | {address or '-'}" for each record; the two sides are the tokenizer pair.

Members

Path	Backbone	Notes
e5_cont1300k/model.pt	E5-small	1.3M training pairs
kA/model.pt	E5-base	seed 7, 1.9M pairs
kB/model.pt	E5-base	seed 8, sample disjoint from kA
kC/model.pt	E5-base	seed 9, 1.9M pairs
kD/model.pt	E5-base	hard-negative continuation of kA
kE/model.pt	E5-base	hard-negative continuation of kB
kF/model.pt	E5-base	mined-error continuation of kA
kG/model.pt	E5-base	mined-error continuation of kB
deadline_50k/model.cbm	CatBoost	Frozen comparison gate loaded by CPU fit (76 features)
gate0_100k/enhanced.cbm	CatBoost	Same bytes as deadline_50k/model.cbm

Downloads must use a pinned commit, not main. The matching pipeline stores that snapshot in BER_HF_REVISION.

SHA256 checksums are in SHA256SUMS.

Load a checkpoint

import torch
from huggingface_hub import hf_hub_download
from transformers import AutoModelForSequenceClassification, AutoTokenizer

REPO = "DeeAxe/business-entity-matching-encoders"
tag, backbone = "kA", "intfloat/multilingual-e5-base"  # or e5_cont1300k + e5-small
path = hf_hub_download(REPO, f"{tag}/model.pt", revision="<pinned commit>")
tok = AutoTokenizer.from_pretrained(backbone)
model = AutoModelForSequenceClassification.from_pretrained(backbone, num_labels=1)
model.load_state_dict(torch.load(path, map_location="cpu"))
model.eval()

Training recipes that produced these files (data prep, seeds, learning rates, hard-mix and mined-error continuations) are in the matching pipeline that consumes this repository; they can be re-run to rebuild every model.pt.

Licence

MIT, matching the E5 base checkpoints.