ContextQMD
Back to Ruflo

Neural Training Skill

.agents/skills/neural-training/SKILL.md

3.6.301.4 KB
Original Source

Neural Training Skill

Purpose

Train and optimize neural patterns using SONA, MoE, and EWC++ systems.

When to Trigger

  • Training new patterns
  • Optimizing agent routing
  • Knowledge consolidation
  • Pattern recognition tasks

Intelligence Pipeline

  1. RETRIEVE — Fetch relevant patterns via HNSW (150x-12,500x faster)
  2. JUDGE — Evaluate with verdicts (success$failure)
  3. DISTILL — Extract key learnings via LoRA
  4. CONSOLIDATE — Prevent catastrophic forgetting via EWC++

Components

ComponentPurposePerformance
SONASelf-optimizing adaptation<0.05ms
MoEExpert routing8 experts
HNSWPattern search150x-12,500x
EWC++Prevent forgettingContinuous
Flash AttentionSpeed2.49x-7.47x

Commands

Train Patterns

bash
npx claude-flow neural train --model-type moe --epochs 10

Check Status

bash
npx claude-flow neural status

View Patterns

bash
npx claude-flow neural patterns --type all

Predict

bash
npx claude-flow neural predict --input "task description"

Optimize

bash
npx claude-flow neural optimize --target latency

Best Practices

  1. Use pretrain hook for batch learning
  2. Store successful patterns after completion
  3. Consolidate regularly to prevent forgetting
  4. Route based on task complexity