What problem does it solve? Reproducing a machine learning paper requires fetching the right datasets and model checkpoints before any experiment can run, and downloading tens of gigabytes of weights only to discover they cannot fit in VRAM wastes bandwidth and time. This Skill checks hardware feasibility per experiment first, then fetches, verifies, and records an absolute on-disk path for every asset the experiment plan needs. ## Core Features & Use Cases - Hardware feasibility checks: Probes GPU VRAM, RAM, and free disk, then computes per-experiment memory requirements (fp16 inference vs. training with optimizer states) with auditable arithmetic before downloading anything. - Asset downloading and cache pre-seeding: Fetches Hugging Face models and datasets, plain-URL files, and pre-seeds framework caches (e.g. keras MNIST at ~/.keras/datasets/) so later stages run deterministically and offline-safe. - Escalation for gated assets: For license-walled or registration-required datasets, it exhausts autonomous routes, then hands the user a concrete task (exact URL, destination path) and iterates across turns until the asset is obtained or the user explicitly gives up. - Use Case: After scoping a paper's experiments, run this stage to confirm a 124M-parameter GPT-2 fine-tune fits on a 24 GB GPU, download the PyTorch weights and E2E NLG dataset, and produce a placeholder mapping table the next pipeline stage consumes mechanically. ## Quick Start Run stage 2 of the reproduction pipeline: check this machine's hardware against the experiments in .paper-reproduction/experiment-scoping.md, download the datasets and checkpoints they need, and write a report with an absolute verified path for every asset.
