dimitrisdais/laya-banking77-typed-decisions

Language Aware Decisions for Banking

This model adapts the open Laya language aware decision model for banking intent classification.

Rather than asking one question with 77 possible answers, it uses a two step decision hierarchy:

1. Select one of eight banking families
2. Select the specific intent within that family

Training

• Base model: convaiinnovations/laya
• Dataset: BANKING77
• Training examples: 7,994
• Validation examples: 999
• Calibration examples: 1,000
• Training method: frozen encoder with trained typed decision heads
• Trainable parameters: approximately 26.5 million
• Selected checkpoint: epoch 30
• Fine tuning method: supervised head training, not LoRA

Results

Model	Validation macro F1
Laya, unchanged	0.447
Laya, banking fine tuned	0.740
TF IDF + logistic regression	0.860

All results use the same 999 message validation split.

Files

• model.safetensors: complete model checkpoint
• rl_agent_config.json: Laya model configuration
• encoder/: encoder configuration
• tokenizer/: tokenizer files
• questions.json: banking decision hierarchy
• artifacts/epoch_history.csv: training history

Source code

GitHub: https://github.com/dimitrisdais/language-aware-decisions-for-banking

The GitHub repository contains the training notebooks, evaluation code, classical machine learning baseline and reproducible inference examples.

Limitations

This is an experimental demonstration trained on customer service intent data. It is not suitable for consequential banking decisions without additional evaluation, calibration, monitoring and human oversight.

Confidence calibration has not been finalized. Confidence values should therefore be treated as experimental.

Attribution

Built from the Apache 2.0 licensed Laya checkpoint.

BANKING77 was introduced by PolyAI and is distributed under CC BY 4.0.