The Agent Skills Directory

[EXTERNAL_DOWNLOADS]: The skill documentation provides instructions to clone the official DeepSeek-OCR repository from GitHub and install pre-release wheels from vLLM's nightly package registry.
[REMOTE_CODE_EXECUTION]: The inference examples utilize the trust_remote_code=True parameter in the HuggingFace Transformers library. This allows the execution of arbitrary Python code defined within the remote model repository during initialization.
[DYNAMIC_EXECUTION]: The use of trust_remote_code=True triggers the dynamic loading and execution of custom model classes and processing logic at runtime from an external source.
[INDIRECT_PROMPT_INJECTION]: The skill's primary function is to process untrusted data sources including images and PDF files.
Ingestion points: Processes external files via parameters like image_file, pdf_path, and image_paths in the vLLM and Transformers inference scripts.
Boundary markers: While instructions include prompt prefixes like <image>\n<|grounding|>, there are no specific safeguards or warnings regarding the processing of adversarial text embedded within the analyzed documents.
Capability inventory: The skill allows writing processed OCR results directly to local files and printing them to the console.
Sanitization: There is no evidence of sanitization, validation, or filtering of the text extracted from the external multi-modal inputs before it is used in subsequent operations.

deepseek-ocr