TGCC-P — hotdog partial fine-tuning

Task: hotdog (cook_hot_dogs) only. This is the original partial step05708 checkpoint (10/20 successes, meanQ .5), not the later full-model or corrected v3fast weights. Original model tensor bytes are unchanged.

This release contains the evaluated partial fine-tuning checkpoint, its success-demonstration retrieval support, calibration, and the complete 20-instance cook_hot_dogs evaluation. 10/20 successes (50%); mean official Q = 0.5000. All 20 instances reached official task terminals: 10 successes and 10 timeouts. No capability stops or infrastructure interruptions occurred in this evaluated suite. No videos were recorded.

This checkpoint is separate from the subsequent TGCC-F full-model5000-update experiment (11/20, meanQ .55) and corrected v3fast release. Their results are not attributed to TGCC-P.

Training scope and provenance

• Parent: TC-π0.5 selected_step5000_user_stop, SHA256 80436905542b0ce75d77d17cf908b29846f5497b58189177d406daebdd3e1239.
• Actual completed schedule: 1000 critic-only warmup updates + 4708 partial joint updates = 5708. The original 6000-update schedule stopped early; original training_complete=false metadata is preserved.
• Jointly trained: ActionExpert, retrieval adapter, and three consistency scoring heads. Student vision/VLM, graph and semantic/memory modules were frozen. Full-model updates: 0 for this release.
• Effective batch: 48. Training support: 5888 fixed anchors from 184 curated successful demonstration episodes; 16 held-out episodes / 512 anchors for calibration. This is not an exhaustive pass over all 51714 original causal sequences.
• Scores measure demonstrated action, execution and planning consistency, not calibrated task-success or failure probabilities. No collected failure examples or synthetic negative targets were used.
• Retrieval matches visual features, robot state, three-level graphs and causal history. Query-episode exclusion is enforced; validation/public-test examples do not enter the training support. Source demonstrations do not have independently verified terminal-success receipts for every episode.

Evaluation protocol

cook_hot_dogs, public-test instances 301–320, action seed42, environment seed0, 32-action chunks, official physics and scoring. Inputs include privileged graph/state information: Oracle Upper Bound. Server inference with local BEHAVIOR simulation.

Native controller input/joint-target limits are retained. The extra TGCC whole-chunk range rejection is disabled. Statistical score deviations trigger feedback/retrieval, not early task termination; actual skill legality, finite-value and interface checks remain. One unconditioned candidate plus at most one fresh reference-conditioned candidate is used per request. A separate 320-step physical interface preflight passed before the scored suite and is not counted in the 20 outcomes.

Camera captures: head720×720, wrists480×480, model224×224. Renderer texture-streaming budget0.05; texture residency may differ from earlier baseline runs. Physics and camera resolutions are unchanged; cross-configuration images are not guaranteed pixel-identical. Official configured limit13717 produces recorded timeout terminal step13718; raw step counts are retained.

Model	Instances	Success	Mean Q
This TGCC partial checkpoint	301–320	10/20 (50%)	0.5000

Complete per-instance table: evaluation/results.md. Machine-readable: CSV, JSON. Original result JSONs, exact server identity, preflight and final verification receipts are under evaluation/. Wall times are original simulation-run wall times, including inference waits.

Contents

• checkpoint/student.safetensors: complete exported student state, including frozen parameters; this does not mean every parameter was fine-tuned.
• checkpoint/distribution.safetensors: scoring heads and retrieval adapter.
• checkpoint/metadata.json: original export and training provenance, unchanged.
• checkpoint/calibration.json, calibration_samples.json: held-out calibration and evidence, unchanged.
• retrieval_library/: final runtime feature/support artifacts, frozen visual encoder and identity. Intermediate per-episode extraction caches and raw demonstration videos are not included.
• source/evaluated_source.tar.gz: exact files listed in the evaluated server's source manifest; source hashes are preserved. runtime_extensions/: the native-limit/advisory runtime and client extensions used for this suite. These contain project-specific launch paths and are not presented as a turnkey portable deployment.
• MANIFEST.json, SHA256SUMS: release file sizes and SHA256 digests.

Download and inspect

from huggingface_hub import snapshot_download
from safetensors import safe_open

root = snapshot_download("Songyuan666/TGCC-P")
with safe_open(f"{root}/checkpoint/distribution.safetensors", framework="pt", device="cpu") as weights:
    print(list(weights.keys())[:10])

Pin the release commit for reproducibility. This is a custom robotics checkpoint, not a Transformers AutoModel repository. Actual deployment additionally requires the matching base configuration/tokenizer/normalizer, TC graph model interfaces, BEHAVIOR/OmniGibson assets and the evaluated advisory runtime policy. Parent/base resources: Songyuan666/TC-pi0.5; parent release revision 7ee996d9184511cff6d9720e73862ae8ee12a664. Normalizer SHA256 d66ed16830a98f90dde8a315058b4a0df59f5e05734c1686d8b3f66787d0a929.

The published results are from the original frozen evaluation. Publication verification covers file identities and evaluation accounting; it does not constitute a fresh 20-instance test of downloaded artifacts or a portable-environment reproduction.

Terms and limitations

Upstream code, weights, demonstration data and simulator/asset usage terms remain applicable. See THIRD_PARTY_NOTICES.md and UPSTREAM_CODE_LICENSE; the latter is the upstream code license and is not a blanket relicensing of all models, data or assets. Simulator assets and raw datasets are external. Single-task fixed-seed results do not establish broad generalization or real-robot safety.

中文：本仓库为TGCC局部微调step5708权重，非全模型微调；完整20例成功10例、平均Q0.500。保留真实训练计数、原始逐例结果和校准信息，不包含当前另行训练的全量模型成绩。