DistilBERT-Reconciler

DistilBERT-Reconciler (v1)

Fine-tuned DistilBERT on 3.2 M labelled post-trade break descriptions + resolution actions (ISO 20022 & proprietary logs).

Figure 1 – DistilBERT-Reconciler: end-to-end training & inference pipeline, showing fine-tuning loop (dashed) and production-time text-to-root-cause flow.

Intended use

Automated classification of reconciliation exceptions in fixed-income settlement workflows (CUSIP/ISIN). Produces label_id then mapped to human root-cause & recommended next action. Not for retail investment advice.

Not for retail investment advice.

Training details

• Base : distilbert-base-uncased
• Epochs : 4 • lr = 3e-5 • max_len = 256
• Hardware : 2× A100 40 GB
• Loss curve & confusion matrix : see /training_artifacts/.

Usage

from transformers import AutoTokenizer, AutoModelForSequenceClassification
tok = AutoTokenizer.from_pretrained("kelvi23/DistilBERT-Reconciler")
mdl = AutoModelForSequenceClassification.from_pretrained("kelvi23/DistilBERT-Reconciler")

text = "COAF: partial collateral received awaiting tri-party"
inputs = tok(text, return_tensors="pt")
pred  = mdl(**inputs).logits.argmax(-1).item()

Limitations & bias

Labels derived from North-American corporate-bond desks (2019–2025). May under-perform on equities or non-USD/CAD repos without re-training.

Citation

Musodza, K. (2025). Bond Settlement Automated Exception Handling and Reconciliation. Zenodo. https://doi.org/10.5281/zenodo.16828730

➡️ Technical white-paper & notebooks: https://github.com/Coreledger-tech/Exception-handling-reconciliation.git