The Agent Skills Directory

[EXTERNAL_DOWNLOADS]: The skill instructions require downloading code from an external repository at github.com/THUDM/slime.git and pulling a Docker image from slimerl/slime:latest to set up the environment.
[DYNAMIC_EXECUTION]: The slime framework provides command-line arguments, such as --custom-generate-function-path and --custom-rm-path, that enable the dynamic loading and execution of arbitrary user-provided Python scripts during the training process.
[INDIRECT_PROMPT_INJECTION]: The skill is designed to ingest and process large external datasets, which presents a surface for indirect prompt injection attacks.
Ingestion points: Training and evaluation prompts are loaded from JSONL files via the --prompt-data and --eval-prompt-data arguments as shown in SKILL.md and references/api-reference.md.
Boundary markers: There are no documented instructions for using specific delimiters or markers to distinguish between data content and potential instructions within the ingested datasets.
Capability inventory: The skill environment executes subprocesses for model training, performs file system writes for checkpoints and logging, and initiates network operations for distributed training coordination.
Sanitization: No methods for sanitizing or validating ingested prompts are described in the provided workflows.

slime-rl-training