slime-rl-training
Pass
Audited by Gen Agent Trust Hub on Sep 17, 2026
Risk Level: SAFEEXTERNAL_DOWNLOADSDYNAMIC_EXECUTIONINDIRECT_PROMPT_INJECTION
Full Analysis
- [SAFE]: The skill documents the 'slime' framework, a legitimate tool for LLM post-training developed by the THUDM research team. All referenced URLs (GitHub, Documentation, and Docker Hub) are official and consistent with the project's identity.
- [EXTERNAL_DOWNLOADS]: The skill instructs users to clone the framework from the THUDM GitHub repository and pull the associated Docker image from Docker Hub. These are standard operations for utilizing this type of software.
- [DYNAMIC_EXECUTION]: The framework supports loading user-defined Python scripts for custom reward functions and multi-turn generation logic via command-line arguments. While this involves dynamic loading, it is a core and intended feature of the tool's RL research purpose.
- [INDIRECT_PROMPT_INJECTION]: The framework ingests training data in JSONL format. 1. Ingestion Points: Data provided via the --prompt-data flag. 2. Boundary Markers: Structure is enforced by the JSONL schema. 3. Capability Inventory: The framework executes training tasks using Megatron-LM and SGLang. 4. Sanitization: Standard data loading is used without specific mention of prompt sanitization. This is a common surface for training tools and is evaluated as safe for this use case.
Audit Metadata