# SmolVLA base policy checkpoint

LeRobot's SmolVLA learned vision-language-action base checkpoint, intended for task-specific fine-tuning. This reference describes provider-pinned safetensors weights; it does not supply a ready robot policy or execute inference.

ID: `lerobot-smolvla-base-policy-weights`

Record version: `source-7cd549ac2351fb069c0ddb3c34ad2d09cfc92b56a15dccdfc2e41467aaca01eb`

Hub commit: `bee2df6672c145beeff064d2a203cebdbd3c5bf3`

## Access

Inspect the exact pinned upstream artifact and setup instructions. This record has no Hub runtime adapter.

Inert file retrieval is unavailable for this reference (`source_too_large`): This single-file acquisition exceeds the fixed 8 MiB limit. Use the pinned upstream source and its acquisition instructions. The Hub CLI does not download this artifact. See [source-file retrieval limits](https://hub.superfastrobotics.com/source-files.md).

Hub preparation adapter: **false**. Runtime verification: **not_run**.

Upstream setup: https://huggingface.co/lerobot/smolvla_base/blob/5e8d12a6e2975b0e5e5fce7c8caf47c371d257b6/README.md

## Source and license

- Repository: https://huggingface.co/lerobot/smolvla_base
- Pinned source: https://huggingface.co/lerobot/smolvla_base/resolve/5e8d12a6e2975b0e5e5fce7c8caf47c371d257b6/model.safetensors
- Revision: `sha256:7cd549ac2351fb069c0ddb3c34ad2d09cfc92b56a15dccdfc2e41467aaca01eb`
- Path: `model.safetensors`
- Source SHA256: `7cd549ac2351fb069c0ddb3c34ad2d09cfc92b56a15dccdfc2e41467aaca01eb`
- Source bytes: 906712520
- Original authors: lerobot
- License: [Apache-2.0](https://huggingface.co/lerobot/smolvla_base/blob/5e8d12a6e2975b0e5e5fce7c8caf47c371d257b6/README.md)
- License scope: The pinned model repository card declares apache-2.0. No separate LICENSE file is present; config.license is null. This retains the model-card declaration without independently clearing tensor-payload rights. The external SmolVLM2 base, training/evaluation datasets, dependencies, hardware and task-specific derivatives require separate terms and provenance review.

## Artifact

```json
{
  "format": "safetensors",
  "kind": "policy"
}
```

## Requirements

```json
{
  "hardware": [],
  "notes": "Model-card declarations only; no Python/package/GPU-memory/OS combination was pinned or tested. Intended for fine-tuning on a user-selected dataset. Config expects six-state/six-action features, image inputs, preprocessing/postprocessing and separate normalization-state safetensors. Config/card reference unpinned HuggingFaceTB/SmolVLM2-500M-Video-Instruct; tokenizer/base weights and dataset terms remain separate. Card inference:false is provider metadata, not proof that inference is impossible. SO100 and lego-box tasks belong to an example using a user's fine-tuned checkpoint, not a compatibility claim for these base weights.",
  "packages": [
    "lerobot[smolvla]",
    "torch"
  ],
  "python": null,
  "system": [
    "Pinned config defaults to CUDA; model-card example selects CPU if CUDA is unavailable"
  ]
}
```

## Invocation

```json
{
  "entrypoint": "SmolVLAPolicy.from_pretrained(model_id) with model_id = 'lerobot/smolvla_base' (pinned card example; not executed)",
  "notes": "Use the upstream pinned model-card setup and compatible reviewed LeRobot code after independent authorization. The card's loading snippet names the moving model ID and does not pass a revision, so reproduce with an explicitly selected provider revision and all required config/processor files rather than assuming this example pins them. This 906712520-byte artifact exceeds the Hub single-file 8 MiB acquisition limit. Hugging Face resolve issues a signed CDN redirect; no redirect, tensor transfer, package install, model/dataset load, GPU/cloud operation or robot contact was performed. No Hub adapter or runtime success is claimed."
}
```

## Compatibility

```json
[]
```

Missing environment fields remain unknown. Source discovery labels do not establish robot compatibility.

## Evidence

Level: **contributor_reported**. Runtime verification: **not_run**.

Data-only intake structure checked. Source content, checksum, license and runtime claims require maintainer review; no submitted code was executed.

Source review and contributor reports do not establish task completion. Preserve upstream attribution and licenses.