What problem does it solve? Reproducing a machine learning paper requires deciding exactly which experiments to run and with which hyperparameters, but papers and codebases routinely disagree on flag names, defaults, and values. This Skill reads the paper PDF and the cloned codebase together and produces a concrete, copy-paste-runnable experiment plan that downstream pipeline stages can execute without further interpretation. ## Core Features & Use Cases - Paper-code reconciliation: Maps each experiment in the paper to a real entry point in the repo, resolving hyperparameters by precedence (author scripts, paper values, code defaults, justified inference) and flagging discrepancies. - Mock and full commands per experiment: Writes a GPU-first scaled-down mock command for smoke testing plus the full paper-faithful reproduction command, with sweep tables of valid parameter tuples instead of free cross products. - Structured scoping report: Saves a standardized report to .paper-reproduction/experiment-scoping.md listing datasets, checkpoints, baseline coverage, and risks for downstream stages. - Use Case: You clone the LoRA repository, drop the paper PDF inside, and ask what to run. The Skill identifies the GPT-2 E2E fine-tuning experiment, emits commands using only flags that exist in gpt2_ft.py, cites the paper's LoRA hyperparameters, and lists the checkpoint and dataset assets to fetch. ## Quick Start Clone the paper's official codebase, place the paper PDF inside it, and ask the agent to scope out which experiments to reproduce with exact commands and hyperparameters.
