claw-finetune

name

claw-finetune

description

Collect fine-tuning requirements, validate datasets and API keys, generate backend-specific LoRA SFT runs for Tinker or HPC-AI SDK, launch training, and report progress. Use when the user wants to fine-tune a model on local data or a backend-hosted/common dataset such as gsm8k, knights-and-knaves, no_robots, or tulu-3-sft-mixture, start a LoRA/SFT job, compare Tinker vs HPC-AI, or ask for training status.

ClawFinetune

When To Use

Apply this skill when the user wants to:

• fine-tune a model on local data,
• fine-tune a model on a backend-hosted or common dataset such as gsm8k, knights-and-knaves, no_robots, or tulu-3-sft-mixture,
• choose between Tinker and HPC-AI SDK,
• generate a training script,
• start a training run,
• check training progress or completion status.

First Actions

1. Read docs/agent-workflow.md.
2. Read docs/dataset-format.md.
3. Read the backend note that matches the requested backend:
   • docs/backend-hpcai.md
   • docs/backend-tinker.md
4. Collect any missing required intake fields.
5. Present the compact confirmation block from docs/agent-workflow.md.
6. Only proceed after the user confirms.

Required Intake Fields

Collect and confirm:

• backend
• HPC-AI account origin if backend is hpcai (www.hpc-ai.com or cloud.luchentech.com)
• model name
• dataset path or dataset identifier
• dataset config if using a shared/common dataset
• dataset split if using a shared/common dataset
• output root if the user wants a non-default location
• LoRA rank
• learning rate
• batch size
• max length
• epoch count
• checkpoint cadence
• optional WandB settings
• optional canonical base_url override if the user wants to pass the endpoint directly

Use defaults from docs/agent-workflow.md when the user does not care.

Execution Workflow

After confirmation:

1. Create the run:

python scripts/create_finetune_run.py --spec /abs/path/to/request.json

2. Launch the run:

python scripts/run_finetune.py --run-dir /abs/path/to/run_dir

3. Check status:

python scripts/check_status.py --run-dir /abs/path/to/run_dir

Host Agent Rules

• Never hardcode API keys or write them into generated files.
• If TINKER_API_KEY or HPCAI_API_KEY is missing, stop and tell the user how to set it.
• If backend is hpcai, collect either hpcai_account_origin or a canonical base_url before launch.
• For HPC-AI, map www.hpc-ai.com accounts to https://www.hpc-ai.com/finetunesdk and cloud.luchentech.com accounts to https://cloud.luchentech.com/finetunesdk.
• Explain that HPC-AI API keys from www.hpc-ai.com and cloud.luchentech.com are not interchangeable.
• Reject invalid dataset formats before launch.
• Prefer resumable runs and reuse the latest checkpoint if the run directory already contains one.
• Report the run_id, backend, model, log paths, and current phase after launch.

Output Requirements

When starting a run, tell the user:

• chosen backend,
• HPC-AI account origin and base URL if backend is hpcai,
• model,
• dataset path or dataset identifier,
• run directory,
• generated script path,
• where logs will appear.

When reporting progress, summarize:

• phase,
• step and progress if known,
• latest loss if known,
• latest checkpoint if known,
• failure reason if failed.

Additional Resources

• Shared workflow: docs/agent-workflow.md
• Dataset rules: docs/dataset-format.md
• HPC-AI notes: docs/backend-hpcai.md
• Tinker notes: docs/backend-tinker.md