EHTB/erp-classifier

ERP GL Account Classifier

Multimodal classifier for ERP GL account classification. Achieves 90.8% test accuracy by learning from:

• Text: vendor + description
• Numerical: transaction amount + log(amount)
• Categorical: department embeddings

Model Details

• Base model: DistilBERT (distilbert-base-uncased)
• Architecture: 3-branch fusion (text 768-d + numerical 64-d + categorical 32-d)
• Parameters: 66.6M (all trainable)
• Training data: 659 samples (synthetic ERP transactions)
• Test accuracy: 90.8%
• Test F1: 90.6%

Usage

import json
import numpy as np
import torch
from pathlib import Path
from transformers import AutoTokenizer, AutoConfig
from huggingface_hub import hf_hub_download

# Download auxiliary files
scaler_path = hf_hub_download(repo_id="EHTB/erp-classifier-dataset", filename="scaler.json", repo_type="dataset")
vocab_path = hf_hub_download(repo_id="EHTB/erp-classifier-dataset", filename="dept_vocab.json", repo_type="dataset")

with open(scaler_path) as f:
    scaler_params = json.load(f)
    scaler_mean = np.array(scaler_params['mean'])
    scaler_scale = np.array(scaler_params['scale'])

with open(vocab_path) as f:
    dept_vocab = json.load(f)['dept2id']

# Load model (requires custom ERPClassifierModel class)
# See: https://github.com/yourusername/erp-classifier/blob/main/src/model.py

tokenizer = AutoTokenizer.from_pretrained("EHTB/erp-classifier")
config = AutoConfig.from_pretrained("EHTB/erp-classifier")

# Note: This model uses a custom architecture (ERPClassifierModel).
# You need to import the model class from the repository to load weights.

Training Procedure

• Optimizer: AdamW (lr=3e-5, weight_decay=0.01)
• Warmup: 10% of training steps
• Batch size: 16
• Epochs: 10 (early stopping patience=5)
• Loss: CrossEntropyLoss
• Training time: ~3 minutes (Apple M1/M2)

Performance

GL Account	F1 Score	Support
GL_6410_FACILITY_PANTRY	100%	18
GL_6450_CLEANING_MAINTENANCE	100%	19
GL_6400_OFFICE_CONSUMABLES	96.8%	15
GL_6250_TECH_BOOKS_TRAINING	95.0%	19
GL_6150_MARKETING_PROMO	94.1%	18
GL_6310_WORKPLACE_SAFETY_ERGONOMICS	91.9%	19
GL_1400_IT_HARDWARE_CAPEX	79.1%	19
GL_6220_IT_PERIPHERALS_OPEX	64.0%	15

Limitations

• Trained on synthetic data (942 samples)
• GL_6220 (IT peripherals < 800 EUR) has lower accuracy due to limited training examples
• Requires custom model class (not compatible with standard AutoModelForSequenceClassification)

Citation

@misc{erp-classifier,
  title={ERP GL Account Classifier},
  author={Your Name},
  year={2024},
  publisher={HuggingFace},
  howpublished={\url{https://huggingface.co/EHTB/erp-classifier}}
}