lydongcanh/tectonic-doctype

Document Type Classifier

Classifies a legal / deal document into one of nine types from its text. A logistic-regression head on frozen BAAI/bge-m3 embeddings, so it is multilingual (100+ languages, 8192-token context) and embeds whole documents.

Labels: acquisition_agreement, commercial_agreement, constitutional, employment_agreement, financial_statements, financing_agreement, ip_agreement, lease_agreement, nda (commercial_agreement is the catch-all for "some other contract").

Evaluation

Measured (English held-out test): macro-F1 0.970. This is the only labelled test set that exists (all training data is English: EDGAR / CUAD / ContractNLI), so it is the only computed benchmark. Per-class F1:

• acquisition_agreement: 1.000
• commercial_agreement: 0.914
• constitutional: 1.000
• employment_agreement: 1.000
• financial_statements: 1.000
• financing_agreement: 0.983
• ip_agreement: 0.875
• lease_agreement: 0.980
• nda: 0.975

Other languages: a capability, not a measured result. The head is trained only on English, and there is no labelled non-English test set, so a score for other languages cannot be reported honestly. Other languages work zero-shot through bge-m3's shared multilingual space, which performs well on public multilingual benchmarks but is unvalidated for this task. Treat non-English predictions as usable but unverified. Confidence is not calibrated, so set any accept/escalate threshold empirically, and note training documents are US-filing-style, so non-US document structures may differ.

Usage

import numpy as np, skops.io as sio
from sentence_transformers import SentenceTransformer
from huggingface_hub import hf_hub_download

REPO = "lydongcanh/tectonic-doctype"
enc = SentenceTransformer("BAAI/bge-m3")
enc.max_seq_length = 8192
head = sio.load(hf_hub_download(REPO, "classifier.skops"), trusted=[])

def classify(text: str):
    words = text.split()
    chunks = [" ".join(words[i:i+2000]) for i in range(0, len(words), 2000)][:6] or [""]
    v = enc.encode(chunks).mean(0); v = v / np.linalg.norm(v)
    p = head.predict_proba([v])[0]; i = int(p.argmax())
    return {"label": head.classes_[i], "confidence": float(p[i])}

Data & license

Built from CUAD (© The Atticus Project, CC BY 4.0), ContractNLI (CC BY 4.0), and SEC EDGAR (public). Released under CC BY 4.0.