ctr-e018-pi05-athenb-lora-10000

CTR experiment E018, optimizer step 10000.

• Experiment: PI0.5 general-checkpoint A-then-B 30K training
• Question: How does PI0.5 fine-tuning from optimizer step zero on the frozen PutCab A-then-B dataset evolve across 10K, 20K, and 30K optimizer updates?
• Content: Inference checkpoint (parameters, assets/normalization, metadata); optimizer state excluded.
• Source path on training host: /root/autodl_oss_runs/E108-autodl4090/checkpoints/pi05_putcab_athenb_fullhorizon_mb16_ga1_lora3ep/ctr-e018-general-athenb-30k-s87431-wdk/10000
• Integrity: every file is listed with SHA-256 in SHA256SUMS.

Published 2026-09-27 at the user's request to preserve all checkpoints before the host was shut down. Evaluation status and caveats are recorded in the CTR experiment record E018.