BiyiVLA pi05 Stage-1 checkpoints
Four full training checkpoints from pi05_biyi_283ep_fft_bs12_15256_20261102_v1.
| Directory | Optimizer updates |
|---|---|
4000/ |
4,000 |
8000/ |
8,000 |
12000/ |
12,000 |
15256/ |
15,256 (final) |
Each directory contains model.safetensors, optimizer.pt, metadata.pt,
and assets/biyi_stage1/norm_stats.json. These are full checkpoints, not only
inference weights. PyTorch .pt files should only be loaded from trusted sources.
Training used pi05 full fine-tuning, global batch 12 on one A800 80GB, BF16, seed 42, action horizon 50, 1,525 warmup updates, peak/end learning rates 2.5e-5 / 2.5e-6, and AdamW with weight decay 0.05 and betas (0.9, 0.95).
Dataset: PPPPPilot/BiyiVLA, frozen revision
bb88c0febefae680612652bf51d4d9e2515cecd5 (283 episodes, 183,073 frames).
There were 15,256 optimizer updates and 183,072 sample exposures.
checkpoints_manifest.json records the byte size and SHA-256 of every uploaded
checkpoint and provenance file. Original run and normalization-statistics
manifests are included for reproducibility. Dataset videos, credentials, cache
files, and smoke-test checkpoints are not included.
Training metrics: https://wandb.ai/y8hong-heymandex/BiyiVLA-Stage1/runs/meqlersd
Training loss is not an evaluation of real-world robot performance.