Build a Domain Instruction-Tuning Recipe for a Legal Coach
Overview
What this challenge is about.
You will source instruction data from three streams: ~3,000 synthetic paralegal Q&A generated by a frontier model (anonymized prompts), ~1,500 curated examples from public legal-coaching forums, and ~500 hand-written gold examples. Format them as ShareGPT-style conversations. Fine-tune a 7B open-weight model with LoRA. Evaluate on a 50-question paralegal benchmark (LLM-as-judge with rubric) plus a 30-question MT-Bench-like generic helpfulness check to detect capability regression. Deliver: dataset, training notebook, benchmark report, and a 3-page recipe doc.
The Brief
What you'll do, and what you'll demonstrate.
Build an instruction-tuning recipe for a legal coach that lifts domain helpfulness without regressing generic capability.
Earning criteria — what you'll demonstrate
- Design a multi-source instruction-tuning dataset
- Apply LoRA fine-tuning to a 7B model on a domain dataset
- Evaluate domain lift alongside generic capability regression
- Communicate fine-tuning recipes for internal reuse
Program Fit
Where this fits in your program.
Sharpens the same skills your degree expects you to demonstrate.
Skills
Skills you'll demonstrate.
Each one shows up on your verified credential.
Careers
Roles this prepares you for.
Real titles. Real skill bridges. Pick the one closest to your trajectory.
Applied AI Scientist
Owning a domain instruction-tuning recipe end-to-end is the day-to-day of applied AI scientists at vertical-AI startups.
This challenge sharpens
- instruction-tuning
- data-curation
- llm-evaluation
Machine Learning Engineer
Building the training pipeline plus the regression-testing harness is core MLE work for any fine-tuning team.
This challenge sharpens
- lora-fine-tuning
- regression-testing
- instruction-tuning
Data Engineer
Curating, deduplicating, and licensing a multi-source instruction dataset is the data-engineering skillset every fine-tuning team needs.
This challenge sharpens
- data-curation
- synthetic-data
- instruction-tuning