pi05_l90clean_from_bowl5

π0.5 LIBERO — continued from bowl5_full, on the libero_90-cleaned dataset

日本語版: README.ja.md

Full-parameter continued fine-tuning of siruku6/pi05_bowl5_full (005500) for 8,000 further steps on siruku6/libero_combined_bowl5_initpose_l90clean, a 111-task dataset that adds libero_90 and the standard LIBERO suites to the 40-task LIBERO-plus branch its parent was trained on.

Checkpoints: 002000, 004000, 006000, 008000.

Read this before using it. This checkpoint trades away performance on its parent's task distribution to gain performance on libero_90. Whether that is a good trade depends entirely on what you are evaluating. For us it was not.

What it is

Evaluation

008000, measured against its own parent. Perturbed LIBERO rows × 5 episodes, EGL rendering.

The whole gain is on libero_90, which this model trained on and its parent did not. On every set drawn from the task distribution we cared about, the two are indistinguishable (differences of 0.005–0.025, against measured run-to-run noise of about 0.05).

About the competition score

This checkpoint was submitted to the PARC2026 finals and scored 0.075, against 0.498 for its parent on the same track with byte-identical inference settings.

That score is not a measurement of this checkpoint. The submission archive was built with the wrong tokenizer in model_weights/vlm/ — a SmolVLM2 (GPT2Tokenizer) directory instead of the π0.5 Gemma one — so every instruction was tokenized into the wrong token IDs. We reproduced it afterwards: same weights, same settings, same two tasks, swapping only the tokenizer directory gave 0.000 (wrong) versus 0.500 (correct).

So the honest summary is: this checkpoint has no valid competition score. What the local evaluation says stands: comparable to its parent on the target distribution, better on libero_90.

Usage

from lerobot.policies.pi05.modeling_pi05 import PI05Policy
from lerobot.configs.policies import PreTrainedConfig

config = PreTrainedConfig.from_pretrained("path/to/008000")
config.load_vlm_weights = False        # VLM weights are in this checkpoint
config.vlm_model_name = "path/to/008000/vlm"
policy = PI05Policy.from_pretrained("path/to/008000", config=config)

vlm/ is not included in this repository and must be put in place yourself — it is the 4-file Gemma tokenizer (tokenizer_class: GemmaTokenizer). See Getting the vlm/ tokenizer below. Getting this wrong is not hypothetical: it is exactly what invalidated this checkpoint's competition score.

Getting the vlm/ tokenizer

vlm/ is the PaliGemma tokenizer, unmodified. Its canonical source is google/paligemma-3b-pt-224. That repository is gated: sign in, open the model page, accept the Gemma license and wait for access to be granted, then authenticate locally with hf auth login before the download will work.

hf download google/paligemma-3b-pt-224 \
    tokenizer.json tokenizer_config.json special_tokens_map.json added_tokens.json \
    --local-dir <checkpoint>/vlm

Those four files, about 17 MB, are byte-for-byte what vlm/ holds; tokenizer.json has md5 e00c4623084b205875ed738d5685f24b. Put them in place as real files, not symlinks.

Note that the tokenizer is not in lerobot/pi05_libero_base — that repository ships only model.safetensors, config.json and the two pre/post-processor JSON files.

Check that tokenizer_config.json says "tokenizer_class": "GemmaTokenizer". A directory containing merges.txt and vocab.json is a SmolVLM2 tokenizer, not this one; the policy will run and silently fail to follow instructions.

License and provenance

Derived from lerobot/pi05_libero_base, which is published under the Gemma Terms of Use.

Gemma is provided under and subject to the Gemma Terms of Use found at ai.google.dev/gemma/terms

Section 3.2 (Use Restrictions) and the Gemma Prohibited Use Policy apply. LeRobot is Apache 2.0; the LIBERO / LIBERO-plus demonstration data is MIT.

Related

• Parent: siruku6/pi05_bowl5_full
• Dataset: siruku6/libero_combined_bowl5_initpose_l90clean
• Index of all published checkpoints: PARC2026 model index