Downloads · 30 days
7
21% of all-time downloads
Elvinky/pi05-piperx-full558-intervention-1to1-mask-10822s-seed1000-fp32
pi05-piperx-full558-intervention-1to1-mask-10822s-seed1000-fp32 is a machine learning model from Elvinky. Use it for the machine learning task on the model card, and read the license before you ship it in a product.
Inference files only; no optimizer state. Use both saved processor pipelines.
Downloads · 30 days
7
21% of all-time downloads
All-time downloads
34
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
Inference files only; no optimizer state. Use both saved processor pipelines.
Tokenizer files are included at the repository root. The exported preprocessor references this repository rather than a machine-specific path. For a fully offline local snapshot, override tokenizer_processor.tokenizer_name with the local snapshot directory when creating the processors. Weights and normalization tensors are unchanged.
This is the final 10,822-step, seed-1000 intervention fine-tune initialized from the full558 50k baseline, not the original pi05_base and not a D1/D3/D7 demonstration-only model. Training used a 0.5 intervention sampling ratio (1:1 demonstration replay and pure human-intervention segments) with temporal action-padding loss masking enabled. The intervention collection contains 616 episodes and 138,514 frames at 30 FPS (about 1 hour 17 minutes); 2.5 intervention epochs is an expected sampling exposure, not 2.5 epochs over the entire combined dataset.
The original train_config.json is included as a record of the custom training configuration; machine-local paths in it are historical references, not portable commands. The inference config is the existing exported config, which omits the training-only mask_action_padding_loss field. No inference-time loss masking or 1:1 sampler is required. This package does not claim full training-code/data reproducibility. No optimizer or RNG state is included.