cyberneticlibrary
← library

Run interactive fine-tune pipeline

finetunecommandsetup L2★0
GIS-DHSIT/DocWain ↗
What it does

Run 5-stage evolving fine-tune pipeline with human checkpoints

Best for

Iteratively improve fine-tuned models with human approval gates at each stage, avoiding costly training mistakes.

Inputs
  • · Pipeline config (YAML)
  • · Interaction signals (JSONL)
  • · Eval prompts
Outputs
  • · SFT pairs (JSONL)
  • · DPO pairs (JSONL)
  • · Training metrics
  • · Model registry (YAML)
Requires
  • · Ollama (DocWain endpoint)
  • · Unsloth (LoRA trainer)
  • · Azure GPT-4.1 (fallback)
Preconditions
  • · Config file exists
  • · Signals dir populated
  • · Ollama running
  • · Unsloth installed
Failure modes
  • · Config validation fails
  • · Insufficient signal data
  • · Training diverges (NaN loss)
  • · Quality gate fails (score < 80%)
  • · Ollama offline
Trust signals
  • · 5 explicit stages with checkpoints
  • · Weighted composite scoring (0.30/0.25/0.20/0.15/0.10)
  • · Tournament ranking of all student models