CoFL-S

This release contains a sanitized CoFL-S checkpoint for non-commercial academic research. Its licensing terms identify CC BY-NC-SA 3.0 US for the CoFL contributors' original contributions, the independent Matterport3D source-data agreement, and the Apache-2.0 terms of the SigLIP2 components.

See third-party notices for source attribution and the remaining limits of the historical annotation-provenance record. This is a research model release with use restrictions; no commercial authorization is provided by this model card.

Artifact and loading

• File: best.ckpt (5,336,379,647 bytes).
• SHA-256: cce748914a88167a2566ffd724ed43f0f651841dcab25afcbef706f5ccd62b12.
• Format: complete PyTorch Lightning checkpoint, CoFL schema 2.
• Saved training position: epoch 0, global step 85,000.
• Inputs: one RGB-D observation and a navigation instruction.
• Outputs: continuous ground-plane field and STOP / MOVE_FORWARD / TURN_LEFT / TURN_RIGHT logits.
• Backbone: SigLIP2 base, patch 16, 224-pixel image preprocessing.

Install the matching CoFL source release and its locked dependencies. Once best.ckpt is downloaded, use the existing loader:

from cofl.training.policy import load_policy

policy, config = load_policy("best.ckpt", device="cpu")

Model configuration, tokenizer/processor files and weights are embedded. Loading does not require a pretrained-model download or the original training paths. For deployment, select an available device supported by the installed PyTorch.

Download this release from lhk66666/CoFL-S using the Hugging Face CLI:

hf download lhk66666/CoFL-S --local-dir cofl-s-release
cd cofl-s-release
sha256sum -c SHA256SUMS
tar -xzf source/cofl-source.tar.gz
cd CoFL
uv sync --locked

The matching code snapshot is included in source/cofl-source.tar.gz; it contains both CoFL and CoFL-S code under Apache-2.0, including configuration, dependency locks and reproduction instructions. Its local tests and maintainer tools are excluded. The wheel in source/ also includes the built Studio UI. From this extracted checkout, the checkpoint path is ../best.ckpt.

For an immutable download, add --revision with this repository's full commit hash from its commit history. Verify SHA256SUMS after downloading. The Hub supports downloads pinned to a revision.

Sanitization and verification

Seven metadata locations were changed: the saved dataset directory and checkpoint callback paths, including keys in best_k_models. Paths now use datasets/cofl-s and checkpoints/cofl-s. No local usernames, mount directories or run identifiers from the identified private paths remain in checkpoint metadata or embedded processor assets. No tensor or processor asset was edited.

The complete optimizer, scheduler, loops, data identity and random states were retained. This does not add a fine-tuning API. Relocating a Lightning run may change which checkpoint callback history Lightning restores, depending on the configured output directory.

• All 2,274 tensor references, including 575 state_dict entries, are byte-identical.
• All six processor assets and all other retained metadata are unchanged.
• Original and sanitized checkpoints load with networking blocked.
• On two deterministic synthetic RGB-D inputs, encoded context, field predictions and action logits are exactly equal. This is a compatibility check, not a new navigation benchmark evaluation.
• Small CPU CoFL-S tests passed at 16 interrupted-training boundaries across num_workers=0 and 2, with identical training suffixes and final states. These tests do not claim a rerun of full-size CUDA training.

Reports: sanitization, inference, resume tests, and checksum.

The source checkpoint SHA-256 is 015acb99eaf3fa8a47833e2c6cc6166eb6d18b274344789c7ed850a12763ca42. Sanitization changes file bytes and the checkpoint hash. Historical evaluations retain their original checkpoint hash; use a new evaluation output directory when evaluating this release file.

License and attribution

Read LICENSE and THIRD_PARTY_NOTICES.md with the checkpoint. Retain them when redistributing it, as required by the applicable licenses. Metadata sanitization does not change source-data obligations.