Downloads · 30 days
13
35% of all-time downloads
Elvinky/pi05-piperx-round2-full558-value-human-1to1-step800-seed1000
pi05-piperx-round2-full558-value-human-1to1-step800-seed1000 is a robotics model from Elvinky. Use it for the robotics task on the model card, and read the license before you ship it in a product. It is set up for lerobot.
Fresh fine-tuning from the original full558 PI05 SFT 50,000-step initializer; not a continuation of the earlier 1511-step policy. Seed 1000.
Downloads · 30 days
13
35% of all-time downloads
All-time downloads
37
Public
Parameters
4.1B
16.6 GB on disk
Likes
0
Public
Click a slice to open those files.
.safetensors16.6 GB · 100%
From the Hugging Face model README
Fresh fine-tuning from the original full558 PI05 SFT 50,000-step initializer; not a continuation of the earlier 1511-step policy. Seed 1000.
Strict global sample batch 64 = 32 demonstration anchors + 32 selected/human anchors. Pool A: 502 training demonstration episodes, 1,132,703 frames. Pool B: 491,500 frames across 7,567 segments: 291,245 demonstration Value-selected frames (positive A50, global Top10%, W20 windows), 124,147 round1 human-intervention frames, and 76,108 round2 human-intervention frames. Only intervention=true human fragments are used. Value-heldout episodes are excluded from fine-tuning pools; the original SFT initializer nevertheless saw full558.
Four A800 GPUs, all-parameter FP32, AMP/TF32/compile off, gradient checkpointing on, Adam foreach=False, microbatch4 x accumulation4 per rank, 800 updates. Learning rate 2.5e-6, warmup200, cosine decay to 2.5e-7 at800. Action horizon50, physical action dimensions14; padded actions are excluded from both loss and denominator. Original initializer normalization is unchanged. LeRobot0.6.1, PyTorch2.11.0+cu128, Transformers5.5.4.
Inference weights, tokenizer, processors and configs only; no optimizer or raw data. Not an exact-resume checkpoint. All original model/config bytes are unchanged. The custom training-only mask_action_padding_loss config field may require a compatible config loader; absolute tokenizer paths must be pointed to the bundled tokenizer directory on the deployment machine. These are loader compatibility considerations, not a request to change model weights or inference calculations.
Strict reload of813 tensors, unchanged normalization and finite real-input action decoding passed. This is numerical/loading validation, not a robot success-rate claim. Model SHA256: 5e7af5ae6958a177e07b89d01b5450fafd16bcb67bf53c58ed45e25c4e735e04