The Agent Skills Directory

[EXTERNAL_DOWNLOADS]: The skill provides instructions to download several standard machine learning libraries and datasets.
Fetches official packages from PyPI: trl, transformers, datasets, peft, accelerate, and torch.
Downloads instruction and preference datasets from Hugging Face repositories: trl-lib/Capybara, trl-lib/ultrafeedback_binarized, trl-internal-testing/descriptiveness-sentiment-trl-style, argilla/Capybara-Preferences, trl-lib/tldr, and openai/gsm8k.
Downloads pre-trained models from the official Qwen (Alibaba) and Google Hugging Face organizations.
[COMMAND_EXECUTION]: The documentation includes standard CLI commands for environment setup and model training.
Provides pip install commands for dependency management.
Includes python -m trl.scripts.ppo and trl dpo commands for launching training sessions via the TRL library's built-in scripts.
Mentions accelerate launch for multi-GPU training orchestration.
[INDIRECT_PROMPT_INJECTION]: The skill is designed to process external datasets, creating a surface for indirect prompt injection.
Ingestion points: Reads training data from local JSON files and Hugging Face datasets (e.g., templates/basic_grpo_training.py, SKILL.md).
Boundary markers: Uses structured formats like ChatML and XML tags (<reasoning>, <answer>) to separate model reasoning from final output.
Capability inventory: Limited to file writes (saving models) and standard training operations; no arbitrary shell execution or network exfiltration from within the scripts.
Sanitization: Uses re.search and specific tag extraction logic to isolate and validate model completions against ground truth answers.

fine-tuning-with-trl