Hugging Face transformers: 'trust_remote_code=True' prompt or error when loading a model. What does it do, is it safe, and how do I load models safely?

Cause: Some Hub models ship their own modeling and configuration Python files instead of using code built into transformers. To load them, from_pretrained needs trust_remote_code=True, which downloads and executes that repository code on your machine. Transformers' docs say to take extra precaution: the Hub scans repositories for malware, but you should still be careful not to execute malicious code. Separately, PyTorch .bin weights use pickle, which can execute code on load, so the Hub recommends trusted sources and the safetensors format. Fix (best first): 1. First check whether you need it. A recent transformers release may already include the architecture natively; upgrade (pip install -U transformers) and retry without trust_remote_code. 2. If remote code is needed, read the repository's modeling and configuration files before enabling it, and prefer models from publishers you trust. 3. Pin the exact commit so the code cannot change under you: from_pretrained("org/model", trust_remote_code=True, revision="<commit-hash>"), with the hash from the repo's commit history. Docs recommend this as an extra layer of security. 4. Prefer safetensors weights (from_pretrained loads them when available) over pickle-based pytorch_model.bin; the Hub's pickle scanning is best-effort, not a guarantee. 5. Run unfamiliar models in a sandbox: a container or VM with no credentials, no mounted secrets and restricted network, not on a machine holding production keys. 6. Never hard-code trust_remote_code=True in shared libraries or services that accept arbitrary model names from users. Verify: The model loads with a pinned revision, only reviewed code was executed, and weights come from safetensors files (check the repository file list). Sources: https://huggingface.co/docs/transformers/en/models https://huggingface.co/docs/transformers/en/custom_models https://huggingface.co/docs/hub/security-pickle