Overview
What this challenge is about.
You receive a 1.5B base model (e.g., SmolLM-1.7B or Qwen-1.8B), permission to use 2 hours of a rented A100, and a curated seed of around 5,000 math-tutoring dialogues. Augment with synthetic dialogues (write a generator using an open 7B teacher and validate quality), build a final 20,000-example instruction dataset, and supervise-fine-tune the 1.5B base. Evaluate on GSM8K-style accuracy, pedagogy rubric (Socratic vs. answer-giving), and an unsafe-content probe set. Write the dataset-curation playbook the next team member will follow.
The Brief
What you'll do, and what you'll demonstrate.
Instruction-tune a 1.5B model into a pedagogy-aware math tutor under tight compute and produce the dataset-curation playbook for ongoing iteration.
Earning criteria — what you'll demonstrate
- Curate an instruction dataset with explicit quality controls
- Run supervised fine-tuning on a small open model
- Evaluate LLM outputs against both accuracy and pedagogy rubrics
- Document a dataset-curation process the next team member can follow
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.
NLP Engineer
Curating instruction data and supervised fine-tuning a small open model is exactly the day-one work of an NLP engineer at any product team shipping their own LLM.
This challenge sharpens
- instruction-tuning
- dataset-curation
- supervised-fine-tuning
AI Engineer
Shipping a tuned on-device-ready model with pedagogy rubrics is core AI-engineer work at any consumer LLM product.
This challenge sharpens
- instruction-tuning
- llm-evaluation
- huggingface
Machine Learning Engineer
Documenting a dataset-curation pipeline for ongoing iteration is the MLE craft of building processes that survive team turnover.
This challenge sharpens
- dataset-curation
- synthetic-data
- huggingface