RoBERTa & SBERT Models Train & Test

[Hyrin-mansoor]

Overview

We need to implement a RoBERTa-based classifier to predict the correct ERPNext Doctype from a user's natural language query. This is the first stage of our multi-step query generation pipeline and SBERT for predicting top fields.

---

Tasks

• Finalize the training dataset in RoBERTa & SBERT format
• Train or fine-tune a RoBERTa & SBERT model.
• Evaluate classification accuracy on dev/test sets.
• Add inference script and API wrapper.
• Integrate with overall pipeline.
• Add logging and error handling.

Notes

This component is expected to evolve continuously — we will likely update it with more training data and re-tune periodically as new doctypes or user question patterns emerge.

Labels

NLP, model, pipeline, stage-1, continual-improvement

[Hyrin-mansoor]

Phase 1 training was completed with good precision.
Continuous training is needed.