Cosmos-Embed1 video-text embedding for text-to-video retrieval, video-to-video search, semantic deduplication, and fine-tuning. Use when the user asks to "fine-tune Cosmos-Embed1", "run cosmos-embed inference", "export Cosmos-Embed1", "embed videos", or "search videos with text".
This NVIDIA TAO skill drives Cosmos-Embed1, a joint video-text embedder, for text-to-video retrieval, video-to-video search, zero-shot/kNN classification, and semantic deduplication. It wraps the packaged `cosmos-embed1` CLI (train, evaluate, inference, export) and gives an agent a reproducible way to fine-tune, evaluate, run inference on, and export the model. The problem it solves is standardizing containerized Cosmos-Embed workflows — resolving the correct published container image, mounting the expected workspace layout, and running each action with consistent, safe Docker options.
The skill resolves the container image through packaged metadata (skill_info.yaml and versions.yaml) rather than building from the private source tree, defines a standard workspace (data, model, specs, results, hf_cache) and compact starting spec templates for train, evaluate, inference, and both ONNX and HuggingFace export. It documents concrete operational details: a protobuf<7 pin to keep Weights & Biases from breaking imports, disabling W&B for smoke/offline runs, and passing a HuggingFace token only as an environment variable (never written into specs, logs, or reports) because the model may download a public BERT Q-Former component. A train-action policy notes AutoML is not packaged for this model and routes all actions through the direct model skill. A bundled benchmark reports it passed NVSkills-Eval, scoring 100% on the security dimension.
Target users are developers and ML engineers who need to fine-tune, evaluate, export, or run inference with Cosmos-Embed1 video-text models. It requires Docker with the NVIDIA container toolkit, GPU access, the published TAO container, and a HuggingFace token for pretrained weights.
It is a joint video-text embedder supporting text-to-video retrieval, video-to-video search, zero-shot and kNN classification, and semantic deduplication. The packaged cosmos-embed1 CLI exposes train, evaluate, inference, and export actions.
Docker plus the nvidia-container-toolkit, GPU access, the published Cosmos-Embed TAO container resolved from versions.yaml, and a HuggingFace token when downloading pretrained nvidia/Cosmos-Embed1-* weights.
It is passed only as the HF_TOKEN environment variable (or via a mounted persistent HuggingFace cache) and must never be written into specs, logs, or reports. The model may still download a public BERT Q-Former component even when the checkpoint is disabled.
The image ships wandb 0.21.0 with protobuf 7.x, which breaks the W&B import before training or evaluation. Running `pip install 'protobuf<7'` as a startup preamble, plus WANDB_DISABLED/WANDB_MODE=disabled for offline runs, avoids the failure.
No. AutoML is not packaged for this model because there are no Cosmos-Embed schemas, so all actions use the direct model skill even when a higher-level request sets automl_policy: on.
Quick Setup:
.claude/skills/Repository
nvidia/skills