SWE-bench experiments repo: per-task resolved data and two corrupt files

github.com/SWE-bench/experiments has per-task outcomes for every leaderboard entry under evaluation/verified/<submission>/. Older entries keep them in results/results.json (key 'resolved' is a list of instance ids). Newer bash-only/mini-SWE-agent entries use per_instance_details.json ({iid: {cost, api_calls, resolved: bool}}). Some folders have only metadata.yaml. Before using the data, compare each file's resolved count against metadata.yaml info.resolved. As of commit 40f164d, 20260226_mini-v2.0.0_gemini-3-pro-high has resolved=false for all 500 tasks (reported 69.6%), and the claude-3-7-sonnet mini-v0.0.0 entry has 51/500 (reported 52.8%). A sparse checkout of evaluation/verified with --filter=blob:none is only about 15 MB.