aiwf-local-ai-training

AIWF local AI training planning and implementation skill. Use for QLoRA, LoRA, fine-tune, adapter training, dataset readiness, VRAM budget, trainer choice, config review, checkpoint and resume plans,

SKILL.md

GitHub

AIWF local AI training

Use this skill when the hard question is how to train or fine-tune locally. It owns the path from dataset readiness to training plan, checkpoint policy, export, and validation handoff.

Safety gate

Do not download large models, launch training, reserve GPU time, or overwrite checkpoints unless Shawn explicitly asks. Planning, config review, small file inspection, and dry-run validation are allowed.

Workflow

  1. Read project guidance first: AGENTS.md, PROJECT_SKILLS.md, training docs, dataset manifests, and existing trainer configs.
  2. Classify the run:
    • LLM LoRA or QLoRA.
    • Image LoRA, DreamBooth-style adapter, embedding, or checkpoint fine-tune.
    • Video or multimodal adapter.
    • Dataset preparation only.
  3. Read references/training-readiness.md before recommending a trainer, hyperparameters, or launch command.
  4. Create a training plan when the work is non-trivial:
python <this-skill>\scripts\create_training_plan.py --model <model-id-or-path> --method qlora --dataset <dataset-path> --out <plan.json>
  1. Verify dataset readiness before trainer choice:
    • provenance and license status
    • schema and file format
    • train/validation/test split
    • deduplication and contamination risk
    • token, image, video, or sample count
    • safety and privacy exclusions
  2. Build the resource budget:
    • GPU model and VRAM
    • RAM, disk, and cache paths
    • precision and quantization
    • sequence length, image resolution, frame count, batch size, gradient accumulation
    • expected checkpoint size and cadence
  3. Choose the trainer only after the dataset and hardware fit. Use source docs for trainer-specific flags. Treat Axolotl, Unsloth, PEFT/Transformers, kohya, EveryDream, Diffusers scripts, and project-local trainers as different contracts.
  4. Define checkpoint and resume behavior:
    • non-overwriting output folder
    • save cadence
    • resume source
    • adapter-only versus merged export
    • tokenizer, config, and metadata preservation
  5. Define post-training gates:
    • no-GPU config parse
    • tiny dry run only when cheap and requested
    • adapter load smoke through aiwf-model-loader
    • eval plan through aiwf-ai-evals
    • serving plan through aiwf-inference-serving when the model will be hosted

Output

Report:

  • training objective and method
  • dataset readiness status
  • hardware and VRAM fit
  • trainer choice and rejected alternatives
  • checkpoint, resume, and export policy
  • commands or configs changed
  • validation run and remaining blockers

Use aiwf-agent-mok when the plan needs source-backed research or findings capture.