reproducible-ai/lda-1b-robocasa

LDA RoboCasa demo reproducibility canary

This is a one optimizer step fine-tuning canary for Wayer2/LDA-robocasa. It demonstrates that pinned LDA source, model dependencies, and the four-episode demo dataset can run through TReqs, produce a changed checkpoint, record a GLaaS lineage DAG, and publish the verified release with roar put.

It is not a quality, accuracy, convergence, or deployment-safety claim.

Reproducible inputs

• Source: reproducible-ai/LDA-1B@fd78edbc1942903fe8dd073a208996bed014849e
• Starting checkpoint: Wayer2/LDA-robocasa@811d14d8c22d3e98021c035948118143f53dd312
• VLM: Qwen/Qwen3-VL-4B-Instruct@ebb281ec70b05090aa6165b016eac8ec08e71b17
• Vision encoder architecture: facebook/dinov3-vits16-pretrain-lvd1689m@114c1379950215c8b35dfcd4e90a5c251dde0d32; weights come from the strict-loaded starting checkpoint
• Dataset: four committed playground/demo_data/sim_pick_place episodes
• Training: one optimizer step, global batch size 4, BF16, one 96 GB RTX PRO 6000 Blackwell GPU, Qwen and DINO frozen
• Release path: robocasa-demo-canary-0.0.1

evaluation.json records both checkpoint SHA-256 digests, tensor metadata, and the name of an action_model.* tensor that changed. input-manifest.json records pinned revisions and hashes for the downloaded inputs and demo files.

Loading layout

The public package includes loader resources under checkpoints/loader/. Copy those resources to a new release root and place the published state dictionary under its checkpoints/ directory to restore LDA's required layout:

config.yaml
dataset_statistics.json
checkpoints/LDA-robocasa-treqs-canary.pt

LDA resolves the Qwen dependency at the exact revision recorded above and in config.yaml; loading must retain that revision rather than using the model's mutable default branch. The DINO architecture config is bundled under pretrained/dinov3-vits16-pretrain-lvd1689m; no separate DINO weight download is needed because strict loading obtains every DINO tensor from this checkpoint.

License

Licensing is component-specific; no single license supersedes every upstream term. The LDA source and bundled demo dataset are identified as CC BY-NC 4.0 in CC-BY-NC-4.0.md. The Wayer2 checkpoint and Qwen model metadata identify Apache 2.0; its terms are included as APACHE-2.0.txt. Embedded DINOv3 materials remain subject to the DINOv3 License included as DINOv3-LICENSE.md.

Use and redistribution must comply with all applicable component terms. Upstream model cards are included under upstream/ for provenance and attribution.

Built with DINOv3.

The demo training adapter preserves the 12 source state values and right-pads them with 46 zeros for the checkpoint's 58-wide state encoder. It routes demo NEW_EMBODIMENT ID 32 to existing Franka slot 4 (the committed data identifies franka_robotiq). Actions retain the loader's 138-wide padding and mask. Video and state observations sample offsets [-5, 0], with the future video at offset 16, matching the pinned checkpoint's two-frame history and prediction horizon. Normal dataset boundary padding still applies. This shape-only canary adaptation does not establish RoboCasa feature semantics or policy quality and does not resize checkpoint tensors. It is recorded as datasets.vla_data.demo_canary_adapter: true in the packaged config.

Public verification

• Immutable checkpoint
• Public training lineage
• Run notes

The host anonymously streamed every checkpoint byte and verified the published inventory and SHA-256 hashes. The worker's load, finite-tensor and changed-parameter checks are bound to the same checkpoint digest. Cold replay and held-out policy quality have not been tested. Roar reported untracked artifact directories.