The Agent Skills Directory

[DYNAMIC_EXECUTION]: The skill uses trust_remote_code=True when loading the meta-llama/Llama-2-7b-hf model in the fine-tuning and merging patterns.
This setting allows the transformers library to execute model-specific code provided by the repository. In this context, it is used with a well-known model from a recognized provider.
[EXTERNAL_DOWNLOADS]: The skill includes instructions to install the lm-eval package for model evaluation.
This refers to the EleutherAI Evaluation Harness, a standard tool for LLM benchmarking downloaded from official package registries.
[COMMAND_EXECUTION]: The skill includes a bash block to run the lm_eval command-line tool.
This is used to perform standardized benchmarks on the resulting fine-tuned models and is part of the intended evaluation workflow.
[INDIRECT_PROMPT_INJECTION]: The skill implements logic to ingest and format instruction datasets, creating an attack surface for indirect prompt injection if the source data is malicious.
Ingestion points: Data is read from train_instructions.jsonl and raw list inputs in SKILL.md.
Boundary markers: The formatting functions (format_instruction) use fixed markers like ### Instruction: but do not include explicit instructions for the model to ignore embedded control sequences.
Capability inventory: The skill includes capabilities to write files to the local system (save_model, save_pretrained), merge weights, and execute evaluation commands.
Sanitization: The patterns do not show sanitization of input fields (instruction, input, output) before they are interpolated into the training prompt.

llm-tuning-patterns