LoRAFleet/openvla-oft-libero-reconstructed-r64

OpenVLA-OFT LIBERO reconstructed rank-64 adapters

Four approximate rank-64 adapters reconstructed by LoRAFleet from the published OpenVLA-OFT merged checkpoints by moojink: spatial, object, goal, and 10 (LIBERO-10). These are reconstructed adapters, not the publisher's original training adapters.

Shared base and method

All four target the official openvla/openvla-7b weights at revision 47a0ec7fc4ec123775a391911046cf33cf9ed83f. All three local base weight shards were SHA-256 verified against that revision's Hub LFS hashes; see base_verification.json.

For each of 439 target modules, factorize W_published - W_base using randomized SVD with our custom FlashTSQR GPU implementation, sketch width 192, two power iterations, seed 0, output rank 64. The projected small core uses SVD. Factors are BF16, B = U*S, A = Vh, with scale 1 (r = lora_alpha = 64). The other 543 backbone tensors match the base exactly for every suite. Each adapter has 221,656,576 A/B parameters, stored in 443,446,440 bytes. This release reconstructs a common-base representation; it is not a further rank-budget compression of those reconstructed adapters and is not lossless.

Each suite contains lora_adapter/, its original action head and proprioception projector, dataset_statistics.json, reconstruction diagnostics, and evaluation records. Task heads and statistics are suite-specific and are required. Their bytes were verified against the pinned publisher revisions in manifest.json. The 7B base weights are not duplicated here.

Download and integration

from huggingface_hub import snapshot_download
release = snapshot_download("LoRAFleet/openvla-oft-libero-reconstructed-r64")
base = snapshot_download(
    "openvla/openvla-7b",
    revision="47a0ec7fc4ec123775a391911046cf33cf9ed83f",
)

Use the OpenVLA-OFT LIBERO runtime and setup, including its continuous action head, proprioception projector, image processing, and action normalization. A stock OpenVLA action decoder is insufficient. Publisher repository IDs and fixed revisions for fetching their runtime metadata are recorded per suite in manifest.json.

apply_adapter.py provides apply_adapter(model, base_directory, suite_directory): pass an instantiated OFT-compatible backbone with the published module names. It explicitly replaces each target weight with the official base tensor plus the FP32 product of the saved factors, then casts to the model weight dtype. This matches the reconstruction arithmetic used in our evaluation and also handles Conv2d factors. It leaves non-target backbone tensors untouched, so those must already contain the official base weights. Load the suite-specific heads and statistics separately through the OFT runtime. Do not apply this adapter on top of the publisher's merged target weights by ordinary additive merging. The PEFT-format files have portable base ID/revision fields; standalone stock PEFT/OFT end-to-end loading has not been validated by this release.

Limited rollout validation

Only task ID 0, eight initial states per suite, was evaluated. This is a smoke test, not a full LIBERO benchmark or proof of equivalent action distributions.

Evaluation instantiated the published OFT model and replaced all 439 target weights with W_base + B@A; the remaining 543 tensors were audited as exactly base-equal. A live multi-adapter shared-base serving system was not benchmarked. Policy-call counts are batched runner calls, not per-episode environment steps.

Attribution

Base model: OpenVLA authors, https://github.com/openvla/openvla (MIT). Source fine-tuned models and OFT runtime: OpenVLA-OFT authors / moojink, https://github.com/moojink/openvla-oft (MIT). Reconstruction and FlashTSQR implementation: LoRAFleet. Please also credit the upstream OpenVLA and OpenVLA-OFT work when using these artifacts.