The Agent Skills Directory

[COMMAND_EXECUTION]: The skill facilitates the execution of training and benchmarking pipelines via the tools/launcher/launch.py script and various shell wrappers (e.g., dflash_online_training.sh, vllm_smoke_test.sh). These are standard operations for machine learning development within the provided project structure.- [INDIRECT_PROMPT_INJECTION]: The skill defines a pipeline that ingests external datasets for model training, creating a potential surface for indirect prompt injection.
Ingestion points: Training data is loaded from .jsonl files (e.g., train.jsonl) generated during data synthesis or provided by the user.
Boundary markers: The skill utilizes Jinja2 chat templates (e.g., chat_template_train.jinja) which include specific tokens like {% generation %} to delineate assistant responses and focus loss calculation.
Capability inventory: The skill possesses extensive capabilities, including executing shell commands via uv run, submitting Slurm jobs, writing artifacts to shared storage (/scratchspace), and initiating network services for model serving (vLLM).
Sanitization: The skill relies on structured input formats and template-based masking (answer_only_loss) rather than explicit string sanitization.- [DYNAMIC_EXECUTION]: The documentation references the trust_remote_code configuration parameter, which is a standard requirement for loading custom model architectures in the Hugging Face and vLLM ecosystems. The skill correctly identifies this as a per-model adjustment rather than a default behavior, providing context for its necessity in specific architectural scenarios.

modelopt-speculative-decoding