lesson_38_toy_yaml_constrained_prompting

Lesson 38: Toy YAML Constrained Prompting

This lesson extends the structured-output prompting block with a YAML-style decoder target. Each sequence contains an instruction prefix and then switches at an explicit yaml_token_id marker into the supervised YAML region.

The toy task keeps the setup tiny: the model predicts two key-value lines after the marker, and the prompt-side labels are masked with ignore_index so only the constrained output segment contributes to the loss.

Run

python -m tracks.llm.lesson_38_toy_yaml_constrained_prompting.train --epochs 1 --device cpu

Outputs are written to outputs/llm/lesson_38_toy_yaml_constrained_prompting/<run_name>/.