What problem does it solve? Manually iterating on model architecture and hyperparameters for small-scale training experiments is slow and unstructured. This Skill lets an agent autonomously edit a single training file, run fixed-budget experiments, and keep or revert changes based on a single comparable metric, enabling overnight unattended optimization loops. ## Core Features & Use Cases - Single-file autonomous editing: The agent edits only tinker/train.py (model, optimizer, hyperparameters) while data prep and validation splits stay immutable, preventing data leakage. - Fixed wall-clock budget per iteration: Each experiment runs under a configurable budget (default 300 seconds) so results are directly comparable, tracked via RESULTS.md, BEST.json, and events.jsonl. - Domain-pluggable packs: Ships with lm-pretrain (val_bpb on TinyStories/FineWeb-edu), vision-classification (CIFAR-10), nlp-classification (20 newsgroups), rl-cartpole, and tabular-classification domains, each with its own metric and templates. - Use Case: Select mode: tinker in a Phase 4 experiment plan, let the agent run an 8-hour overnight loop on a single GPU, then hand RESULTS.md and the best train.py to the paper drafting phase as a tinker journal. ## Quick Start Ask the agent to scaffold tinker mode for your project slug and run a baseline iteration with the tinker_run.sh script, then start the autonomous loop following program.md.
