Academic coursework · Natural Language Processing / Project 08
Computer Science · Ton Duc Thang University
LawMate
An academic interface for asking questions and comparing Vietnamese tax-law models
Project focus
An academic Q&A demo with explicit configuration choices, loading and generation feedback, and expandable supporting passages. Single-answer and comparison views make the experimental differences easier to inspect.
- Comparison flow
- Loading states
- Readable results
A flow in motion
One question. Four ways to explore it.
Follow configuration selection, a shared question, supporting context, and comparison across the academic demo’s four approaches.
Four explicit configurations
A shared question → visible context
How do the four configurations handle the same question?
Make the configuration visible.
Base model, retrieval, fine-tuning, and the combined approach are explicit choices, rather than invisible changes behind an answer.
Plays while in view. Pause or choose a stage to explore at your own pace. A conceptual visualization using sample data, independent of the original application.
- Context
- Natural Language Processing · Course project
- Project status
- Completed
- Tools & technologies
- Python · Streamlit · RAG · QLoRA
01 / The project
The project
in context.
LawMate is an end-of-semester NLP project comparing a base language model, retrieval, fine-tuning, and their combination for Vietnamese tax-law questions. Its Streamlit demo makes those experimental choices visible through configuration cards, example questions, answer views, and side-by-side comparisons. It is presented as an academic model-comparison tool.
The user flow
- 01Choose a configuration
- 02Enter or select a question
- 03Wait for model loading and generation
- 04Read the answer and retrieved context
- 05Compare configurations
02 / The making
What went
into it.
Make the active configuration explicit
Four configuration cards distinguish the base model, retrieval, fine-tuning, and combined approach. The interface shows which model is loaded and separates single-answer exploration from comparison mode.
Explain the waiting states
Model loading and answer generation have distinct progress messages. Comparison runs configurations sequentially, with one model loaded at a time, so the interface reflects the real processing sequence.
Keep supporting context within reach
Retrieved passages are available in expandable sections beside the answer and comparison results. Example questions provide an entry point without requiring the visitor to invent a suitable test query.