Songyuan666/TC-pi0.5

TC-π0.5 — hotdog fine-tuned checkpoint + PT50 base

cook_hot_dogs/step5000 is hotdog-only fine-tuning. pi0.5-behaviour-base remains the unchanged 50-task PT50 base; it is not a hotdog-only weight. Original 10-instance results below remain historical; later complete20 results: TC 9/20, meanQ .5000; PT50 5/20, meanQ .2750 (fixed301–320).

Versioned model assets for susongyuan/TC-pi0.5.

Directory	Model	Task coverage
cook_hot_dogs/step5000	TC-π0.5 full fine-tuned checkpoint	cook_hot_dogs
pi0.5-behaviour-base	π0.5-behaviour-base, unchanged PT50	50 BEHAVIOR tasks
common	PaliGemma tokenizer	shared inference asset

Evaluation

Public instances301–310, action seed42, chunk32, official task termination and Q.

Model / input	Success	Mean Q
PT50 base / RGB	2/10 (20%)	0.20
TC-π0.5 / RGB + Oracle graphs, full fine-tuning	5/10 (50%)	0.55

The selected checkpoint is cumulative step5000:4970 formal updates ×48 sequences, approximately4.613 epochs. Original weights, config, normalization and metadata bytes are preserved and covered by release_manifest.json. The code release pins this repository's exact commit; follow its download and inference commands.

Base source: sunshk/openpi_comet_pytorch, revision 5a489fe50fdc8e256533c8c611dfd60c247520cf, folder pi05-b1kpt50-cs32. Renaming the mirror does not change the base model. Upstream model and asset usage terms remain applicable.

Verification: a fresh GitHub clone and fresh Hub download passed GPU inference in the tested dependency environment. See release_verification.json.

Clean-environment verification: a new isolated Python3.11 installation passed dependency checks, GPU model loading and full32×23 synthetic action generation. The complete chunk matched the original environment exactly (maximum difference0). The code release preserves the exact tested PyAV14.4.0 binary as a SHA256-pinned release asset. See release_verification.json.