Table of Contents
🌏 中文版
Lab 3 in the official MIT 6.S191 2026 repository is Lab 3: LoRA Fine-Tuning and LLM-as-a-Judge Evaluation. It builds chat templates and generation with LFM2-1.2B, adapts style through LoRA, and combines OpenRouter with Opik for a judge workflow. This article pins the 2026 branch so later changes to master do not silently alter the exercise.
Before you begin
The official 2026 README specifies Google Colab, Python 3, and a GPU runtime. Copy the notebook to your Drive and run it from the beginning. Put API keys in the notebook's secret manager—never in a shareable cell or Git commit.
Recommended sequence
- Fix three prompts and record base-model outputs
- Complete the LoRA configuration and training TODOs
- Write a human rubric before spending API credit on a judge
Solve one TODO at a time. Write the expected input and output shapes before executing the cell; when something fails, preserve the error and your reason for the fix. Public solutions are for final comparison, not initial copying.
Expected outputs are base-versus-LoRA responses for fixed prompts, a training record, and human or judge scores under one rubric. Common failures include mismatched chat-template/tokenizer formatting and invoking the judge before freezing the rubric, which makes results incomparable.
Completion criteria
Keep a notebook copy, one reproducible end-to-end run, and a short conclusion: what the model did correctly, where it failed, and which variable you would change next. A service-dashboard screenshot does not replace model outputs and an experiment record.
Limits
This lab has the most dependencies: a Colab GPU, Comet/Opik, and an OpenRouter key. The official Lab 3 notebook warns that capable judges may cost money and free models may be rate-limited; recheck current terms.
References
Loading...