| --- |
| license: other |
| base_model: sentence-transformers/all-MiniLM-L6-v2 |
| pipeline_tag: text-classification |
| language: |
| - en |
| tags: |
| - finance |
| - transactions |
| - text-classification |
| --- |
| |
| # Finipher transaction classifier v1.0.0 |
|
|
| Sorts one bank or card transaction line into one of 17 spending categories. It is a fine-tuned |
| `all-MiniLM-L6-v2` (`BertForSequenceClassification`), so it loads with the standard Transformers |
| classes and needs no custom code. |
|
|
| Internal checkpoint name: `finance-multitask-minilm-two-head-v2`. Finipher's manifest version: |
| `minilm-two-head-20261027-v2`. |
|
|
| ## Labels |
|
|
| Education, Entertainment, Fees, Groceries, Healthcare, Income, Insurance, Mortgage, Personal Care, |
| Rent, Restaurants, Shopping, Subscription, Transfer, Transportation, Travel, Utilities. |
|
|
| ## Input |
|
|
| - One line per transaction, prefixed with its direction: `[debit] STARBUCKS #1234 SEATTLE WA` or |
| `[credit] ACME CORP PAYROLL`. |
| - Maximum length 64 tokens. |
| - Below 0.50 confidence, don't trust the label. |
|
|
| ```python |
| from transformers import AutoModelForSequenceClassification, AutoTokenizer |
| import torch |
| |
| repo = "finipher/finipher-transaction-classifier" |
| tok = AutoTokenizer.from_pretrained(repo, revision="v1.0.0") |
| model = AutoModelForSequenceClassification.from_pretrained(repo, revision="v1.0.0").eval() |
| |
| x = tok(["[debit] NETFLIX.COM 866-579-7172 CA"], truncation=True, max_length=64, return_tensors="pt") |
| with torch.inference_mode(): |
| p = torch.softmax(model(**x).logits, dim=-1)[0] |
| i = int(p.argmax()) |
| print(model.config.id2label[i], float(p[i])) |
| ``` |
|
|
| ## Results |
|
|
| Measured offline through Finipher's production code: |
| - 1,290 hand-written everyday vendor lines (Plaid names, card swipes, POS, ACH): 90.7% correct. |
| - DoDataThings 68k rows, final Finipher category: 98.7%. |
|
|
| ## Known mistakes |
|
|
| - Walmart paychecks (`[credit] WALMART DIR DEP PPD`) read as Shopping. |
| - FasTrak and Circle K card lines read as Groceries. |
| - Mortgage `PRINCIPAL PMT` reads as Utilities. |
|
|
| ## Training data |
|
|
| To be confirmed. This repo stays private until the training data and the rights to it are |
| documented. |
|
|
| ## Use |
|
|
| The output is a category label, 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` | |
|
|
