Saeedhoj/spendcoder-distilroberta

SpendCoder: purchase line to spend category

Fine-tuned distilroberta-base that reads a short purchase or invoice line (for example "ANNUAL MAINTENANCE OF HV SWITCHGEAR") and returns one of 14 spend categories, with a confidence score. Low-confidence lines are meant to go to a person. Code, data preparation and tests: https://github.com/saeedm31/SpendCoder

Results (test year, contracts the model never saw)

Model	Accuracy	Macro F1	Weighted F1
TF-IDF + logistic regression (baseline)	0.673	0.616	0.675
Fine-tuned distilroberta-base	0.665	0.569	0.664

At the chosen review threshold (0.99), 29.0% of test lines are coded automatically with 91.3% accuracy; the rest go to a person.

Train: 17,098 lines (contracts starting before Oct 2023). Validation: 2,065 (Oct 2023 to Sep 2024). Test: 1,547 (from Oct 2024). Test lines whose exact text appears in training were removed.

Categories

• Engineering & professional services
• Research & development
• Facility operations & management
• Construction & building maintenance
• Equipment maintenance & repair
• IT & telecom services
• Environmental services
• Utilities, security & housekeeping
• Other services
• IT hardware & software
• Electrical & power equipment
• Lab instruments & chemicals
• Vehicles & machinery
• Other goods

Data

US Department of Energy contract descriptions and their Product and Service Codes from USAspending.gov (public domain), grouped into the 14 categories above.

Use

Suggest a spend category for purchase and invoice lines in accounts payable or procurement; send lines below the review threshold to a person. Not for automatic posting without review. Trained on one public-sector buyer's wording: before using it in another company, check it on that company's labelled lines and fine-tune further.

Files

• PyTorch / Transformers weights at the repo root: AutoModelForSequenceClassification.from_pretrained(...)
• onnx/: model.onnx, tokenizer.json, spendcoder.json (labels, review threshold) for CPU serving with onnxruntime, no PyTorch needed