Fine-Tune a Small Transformer for Legal-Domain EN-DE Translation
Overview
What this challenge is about.
You receive a 120,000-segment parallel EN-DE legal corpus and a held-out 1,000-segment test set with reference translations. Fine-tune a small pretrained Transformer (e.g., NLLB-200-distilled or Opus-MT-en-de) using Hugging Face transformers. Compare against the unfinetuned baseline on BLEU, chrF++, COMET-22, and a curated 200-term legal-glossary recall. Discuss whether automatic metrics align with human reading of 30 random outputs. Deliver a 4-page recommendation memo for the head of product.
The Brief
What you'll do, and what you'll demonstrate.
Decide whether a fine-tuned small Transformer beats a generic baseline well enough to ship as the in-house legal-translation engine.
Earning criteria — what you'll demonstrate
- Fine-tune a pretrained encoder-decoder Transformer for a specific domain
- Apply multiple automatic MT metrics and understand their limitations
- Quantify domain-glossary recall as a domain-specific MT metric
- Recommend a ship/no-ship decision based on a mixed metric + qualitative view
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.
Career paths this builds toward
Canonical rolesNLP Engineer
Fine-tuning encoder-decoder Transformers and evaluating with both automatic and human metrics is the NLP-engineer's headline portfolio piece at any translation-adjacent product.
This challenge sharpens
- neural-mt
- transformer
- fine-tuning
Applied AI Scientist
Pairing automatic and human evaluation for a ship/no-ship decision is the applied-AI-scientist's daily craft in regulated-domain AI.
This challenge sharpens
- mt-evaluation
- domain-adaptation
- fine-tuning
Machine Learning Engineer
Reproducible fine-tuning runs with checkpoints and metric tracking are the MLE-grade portfolio piece for any productionized NLP system.
This challenge sharpens
- pytorch
- transformer
- fine-tuning