together-fine-tuning
Pass
Audited by Gen Agent Trust Hub on Sep 16, 2026
Risk Level: SAFEINDIRECT_PROMPT_INJECTIONEXTERNAL_DOWNLOADSCOMMAND_EXECUTION
Full Analysis
- [INDIRECT_PROMPT_INJECTION]: The skill facilitates the ingestion of user-supplied datasets in JSONL and Parquet formats for model fine-tuning. This inherent functionality creates an attack surface where training data could potentially contain instructions aimed at influencing the behavior of the resulting model.
- Ingestion points: Training and validation files are uploaded via the together SDK in workflow scripts like scripts/finetune_workflow.py and scripts/dpo_workflow.py.
- Boundary markers: The skill utilizes structural validation (local and server-side) to ensure format compliance but does not perform semantic sanitization of training messages.
- Capability inventory: The workflow scripts enable creating fine-tuning jobs and deploying inference endpoints on Together AI managed infrastructure.
- Sanitization: Validation is restricted to file format and schema structural checks.
- [EXTERNAL_DOWNLOADS]: The skill interacts with well-known and trusted services, including the official Together AI API and HuggingFace. It involves installing official packages like the together SDK and requests, and downloading standard datasets or base model weights for fine-tuning.
- [COMMAND_EXECUTION]: The provided Python scripts are administrative tools designed for developers to manage their Together AI resources. They perform API-driven tasks such as job monitoring, file management, and endpoint lifecycle control as their primary intended function.
Audit Metadata