The Agent Skills Directory

[REMOTE_CODE_EXECUTION]: The skill instructs the agent to set the environment variable TORCH_FORCE_NO_WEIGHTS_ONLY_LOAD=1 when using the checkpoint_averaging.py script. This setting disables PyTorch's weights_only security check during torch.load, allowing for the deserialization of arbitrary Python objects. This creates a risk of arbitrary code execution if a user processes malicious or untrusted checkpoints. Evidence: references/training-evaluation.md.
[COMMAND_EXECUTION]: The skill's core functionality involves the execution of numerous shell commands and Python scripts for environment setup, data preparation, training, and evaluation. This creates a broad attack surface typical of machine learning development environments. Evidence: Found throughout SKILL.md and all reference files.
[PROMPT_INJECTION]: The skill processes external manifest files (JSONL) containing audio paths and text transcripts, which represent an ingestion point for untrusted data. 1. Ingestion points: model.train_ds.manifest_filepath and dataset_manifest. 2. Boundary markers: No explicit delimiters or 'ignore-instructions' warnings are specified for the processed transcripts. 3. Capability inventory: Execution of training, evaluation, and data refinement scripts. 4. Sanitization: No sanitization or filtering of transcript content is described. This surface could be exploited via indirect prompt injection if transcripts contain malicious instructions that influence the agent's behavior during data auditing or refinement planning. Evidence: references/data-lhotse.md.

nemo-speech-asr-finetune