This repo now includes a lightweight autoresearch-style workflow adapted from karpathy/autoresearch: the idea is to let an AI agent iterate on train.py, run short fixed-budget experiments, and keep only changes that improve validation quality.

Files that matter

prepare.py - fixed utilities for dataset checks, runtime overrides, and metric extraction
train.py - the single training file the agent edits
program.md - instructions for the research agent

Metric

The primary objective is metrics/mAP50-95(B) from Ultralytics validation results. Higher is better.

Setup

Install dependencies with uv, make sure the dataset YAML exists at ships-aerial-images/data.yaml, then run:

uv sync

Training

Run the baseline or any experiment with:

uv run train.py

By default, the training script uses a fixed 5-minute budget through the Ultralytics time argument and prints a compact summary at the end so an agent can compare runs automatically.

Autoresearch loop

Create a fresh branch such as autoresearch/mar24
Read program.md
Run a baseline with uv run train.py > run.log 2>&1
Iterate only on train.py
Log outcomes to results.tsv
Keep only commits that improve metrics/mAP50-95(B)