Torres1980

Add transaction classifier model, model card, and training code

8c5accf verified 9 days ago

1.64 kB

"""
Classify statement line items and route them: straight-through or human review.
Usage:
    python predict.py "CAP CALL #3 Lakeshore Growth Fund II DUE 10/24/2026"
    python predict.py --file statement_lines.txt
"""

import argparse
import json
from pathlib import Path

import skops.io as sio

HERE = Path(__file__).parent
# Production default. The threshold tuned on the synthetic test set (see metrics.json) is
# optimistic for real feeds, where out-of-taxonomy items appear; 0.90 keeps them in review.
THRESHOLD = 0.90


def load(path=HERE / "model.skops"):
    # skops only loads types you approve; review the list before trusting a file.
    untrusted = sio.get_untrusted_types(file=path)
    return sio.load(path, trusted=untrusted)


def classify(model, lines, threshold=THRESHOLD):
    proba = model.predict_proba(lines)
    out = []
    for text, p in zip(lines, proba):
        top = p.argsort()[::-1][:3]
        conf = float(p[top[0]])
        out.append({
            "text": text,
            "label": model.classes_[top[0]],
            "confidence": round(conf, 4),
            "route": "straight_through" if conf >= threshold else "human_review",
            "alternatives": [{"label": model.classes_[i], "confidence": round(float(p[i]), 4)} for i in top[1:]],
        })
    return out


if __name__ == "__main__":
    ap = argparse.ArgumentParser()
    ap.add_argument("text", nargs="*")
    ap.add_argument("--file")
    args = ap.parse_args()
    lines = Path(args.file).read_text().splitlines() if args.file else args.text
    for r in classify(load(), [l for l in lines if l.strip()]):
        print(json.dumps(r))

	"""
	Classify statement line items and route them: straight-through or human review.

	Usage:
	python predict.py "CAP CALL #3 Lakeshore Growth Fund II DUE 10/24/2026"
	python predict.py --file statement_lines.txt
	"""

	import argparse
	import json
	from pathlib import Path

	import skops.io as sio

	HERE = Path(__file__).parent
	# Production default. The threshold tuned on the synthetic test set (see metrics.json) is
	# optimistic for real feeds, where out-of-taxonomy items appear; 0.90 keeps them in review.
	THRESHOLD = 0.90


	def load(path=HERE / "model.skops"):
	# skops only loads types you approve; review the list before trusting a file.
	untrusted = sio.get_untrusted_types(file=path)
	return sio.load(path, trusted=untrusted)


	def classify(model, lines, threshold=THRESHOLD):
	proba = model.predict_proba(lines)
	out = []
	for text, p in zip(lines, proba):
	top = p.argsort()[::-1][:3]
	conf = float(p[top[0]])
	out.append({
	"text": text,
	"label": model.classes_[top[0]],
	"confidence": round(conf, 4),
	"route": "straight_through" if conf >= threshold else "human_review",
	"alternatives": [{"label": model.classes_[i], "confidence": round(float(p[i]), 4)} for i in top[1:]],
	})
	return out


	if __name__ == "__main__":
	ap = argparse.ArgumentParser()
	ap.add_argument("text", nargs="*")
	ap.add_argument("--file")
	args = ap.parse_args()
	lines = Path(args.file).read_text().splitlines() if args.file else args.text
	for r in classify(load(), [l for l in lines if l.strip()]):
	print(json.dumps(r))