Per-puzzle records, bootstrap summaries and analysis outputs backing
connections-rl, a two-scale (Qwen2.5-1.5B / 7B), three-seed study of
what verifiable-reward RL actually transfers.
This is an artifact bundle for auditing published numbers, not a loadable
training dataset, so the dataset viewer is disabled.
Two conventions in these files are easy to misread. Both have bitten this
project… See the full description on the dataset page:
https://huggingface.co/datasets/jacksonlukas/connections-rl-results.
