finipher/finipher-tax-document-classifier

Finipher tax document classifier v1.0.0

Names what a US tax or financial document is (W-2, 1098, 1099-INT, mortgage statement, paystub, brokerage statement and so on) from the first 128 tokens of its extracted text. 95 labels.

It is a small head on a frozen MiniLM encoder. The head only works with the exact encoder it was trained on, so this repo carries its own copy of that encoder (model.safetensors, SHA-256 96d95c25…). That file also holds a transaction output layer, which this model ignores.

Internal names: checkpoint finance-multitask-minilm-two-head-v2; head artifact tax-document-head-20261027-v1.1-category-encoder (stored inside tax_document_head_config.json).

How to use

The head is custom, so load it with this code (the same steps Finipher runs in production):

from huggingface_hub import snapshot_download
from safetensors.torch import load_file
from transformers import AutoModel, AutoTokenizer
import json, torch

path = snapshot_download("finipher/finipher-tax-document-classifier", revision="v1.0.0")
tok = AutoTokenizer.from_pretrained(path)
enc = AutoModel.from_pretrained(path).eval()          # ignores the transaction output layer
head = load_file(f"{path}/tax_document_head.safetensors")
cfg = json.load(open(f"{path}/tax_document_head_config.json"))
labels = [cfg["id2label"][str(i)] for i in range(cfg["num_labels"])]

def classify(text: str):
    x = tok([text.strip()[:4000]], truncation=True, max_length=128, return_tensors="pt")
    with torch.inference_mode():
        h = enc(**x).last_hidden_state
        m = x["attention_mask"].unsqueeze(-1).float()
        e = torch.nn.functional.normalize((h * m).sum(1) / m.sum(1), dim=-1)
        p = torch.softmax(e @ head["weight"].T + head["bias"], dim=-1)[0]
    i = int(p.argmax())
    return labels[i], float(p[i]), float(p[i]) >= cfg["review_threshold"]   # 0.76

The third value says whether the label is confident enough to use (0.76 or higher). Below that, send the document to a person to review. OTHER_TAX_FORM and UNKNOWN_DOCUMENT always need review.

Results

Measured offline through Finipher's production code:

• Test set (2,531 documents): 98.5%.
• Same set with the titles cut off (1,203): 94.7%.
• Blind probes (33): 90.9%.

Known mistakes

• Credit card statement CSV exports read as BANK_STATEMENT (confidence 0.51–0.71, so usually sent to review).

Training data

finipher/finipher-us-tax-document-classifier at revision 92c67c9 (v1.1.0), licensed CC-BY-4.0. The encoder underneath also comes from the transaction model, whose training data is still to be confirmed. This repo stays private until that is documented.

Use

The output names a document type. It is not tax or financial advice.

Files and SHA-256

File	SHA-256
config.json	6b8c456706017e7cbddd8a4826a88746294100336be5aa9e40f7730890f1b72b
model.safetensors	96d95c2520f9b8cee0a3e85e1440f541420c4c043e18a0c4579943e08e91a5a3
tokenizer.json	41b9d12a4c559e4098f30173bc0f76d092c43d7b1f873e7027da45a256f30f87
tokenizer_config.json	1128dd50ff3f913671a9405619c1bd8f9e9ddc28f62257cb94ed70683590b715
tax_document_head.safetensors	bb60aaad3734afda9120757dec4e2ed60ae04c6ba93f5993a3cb0c893ee4ac35
tax_document_head_config.json	41600002856c26fc30bb47d2c6e40b3c8a0352b77220c1e21c3d6c3e57594c8b