PI0.5 sortletters CASM-LAN independent-arm v25 — step 30000
This repository contains the full-state step-30000 training checkpoint for pi05_sortletters_casm_lan_independent_3x.
- Dataset:
Shiki42/sortletters_parallel_121_casm_lan_3x_v25
- Dataset revision:
sha256:5914c89f4d9bcb6e522c1c11d6d332d9b33f70da11e9ba603d6b749365902021
- parallelVLA commit:
5c05847c1363d93b37be960b735bcb19db3e7197
- OpenPI commit:
50025322f29614ba4124af715f9bf15fa97e5759
- Base model SHA-256:
0eb11ca9587678c1d2ef8cf32807c29f8ce53a2bfdfc1aa4a4c96f16fca59b0f
- Global batch: 32
- Optimizer updates: 30000
- Seed: 87431
- Runtime: BF16, SDPA, compile none, full gradient checkpointing, non-fused AdamW
- Checkpoint cadence: every 5000 updates
The checkpoint contains model parameters and full optimizer/training state. Training metrics and the raw performance receipt are included. This artifact has passed producer-side file, provenance, finite-metric, and checksum gates. It has not yet been qualified by a real-robot rollout success-rate evaluation; training loss is not a behavior claim.
Use FILES.tsv, SHA256SUMS, DELIVERY_RECEIPT.json, and CONTROL_SHA256SUMS to verify a download. The model preserves the lan_independent restriction: each arm consumes only its own wrist/state/language stream, with zero cross-arm communication structure.
