LLM Parameter-Efficient Fine-Tuning (PEFT/LoRA)

Purpose

Fine-tune open-weight LLMs using LoRA, QLoRA, Hugging Face TRL, and custom instruction datasets.

This canonical skill provides deterministic, production-grade operational directives for AI agents executing autonomous engineering and verification tasks within the ai domain.

When to Use

• When designing, developing, optimizing, or debugging within ai tasks.
• When an autonomous agent requires deterministic execution patterns for fine-tuning.
• When verifying compliance, architecture, or performance standards in this domain.

When NOT to Use

• When operating outside the scope of ai engineering.
• When a more specialized sub-skill or external toolchain is explicitly requested by the user.

Capabilities

• Structural design and implementation for LLM Parameter-Efficient Fine-Tuning (PEFT/LoRA).
• Automated verification, schema compliance, and diagnostic troubleshooting.
• Performance profiling, security posture hardening, and error recovery.

Inputs

• Project source files, configuration manifests, or task instructions.
• Target runtime environment parameters and toolchain dependencies.

Workflow

1. Pre-flight Assessment: Inspect existing configuration and environment preconditions.
2. Implementation & Transformation: Apply focused, AST-aware structural modifications.
3. Verification & Audit: Execute domain-specific test suites, syntax checks, or simulation passes.
4. Resolution: Resolve any detected regressions or failure modes prior to completion.

Tools

• Project-approved terminal tools, file editors, and verification test harnesses.

Examples

• Standard operational execution pattern for fine-tuning:

  # Verify environment readiness and execute task workflow
  allskills route "fine-tuning"
  

Safety

• Maintain strict workspace boundary sandboxing.
• Prevent unvetted credential exposure and destructive disk operations.
• Ensure all modifications are verified before concluding execution.

Source

All-Skills Canonical Engineering Framework.

Notes

• Pairs with relevant testing, linting, and architecture verification skills across the platform.