SCBE Colab Training Ops

Use this skill when Colab should become part of the normal SCBE operating surface instead of an ad hoc notebook tab.

Entry Surface

• Start with issac-help and look at the COLAB section.
• Use colab-catalog first to see what notebooks already exist.
• Use colab-show <name> or colab-url <name> before guessing which notebook to run.
• Use colab-bridge-set only when the user already has a Colab local-connection URL.

Core Lanes

1. Notebook Selection

• colab-catalog
• colab-show pivot
• colab-show finetune
• colab-show webtoon

2. Local Bridge / Session Reuse

• colab-bridge-set -BackendUrl <url> -Name pivot
• colab-bridge-status -Name pivot
• colab-bridge-env -Name pivot

3. Training Prep

• hf-generate-sft
• hf-train-wave
• hf-agent-loop

4. Guided Multi-Step Lane

• buildflow-colab <topic>

What Exists Already

• notebook catalog script: scripts/system/colab_workflow_catalog.py
• bridge script: C:\Users\issda\.codex\skills\scbe-n8n-colab-bridge\scripts\colab_n8n_bridge.py
• command-center surface: scripts/hydra_command_center.ps1
• free T4 fine-tune notebook: notebooks/scbe_finetune_colab.ipynb
• pivot notebook: notebooks/scbe_pivot_training_v2.ipynb

Read references/notebook-map.md when you need the notebook-by-notebook breakdown.

Safety Gates

1. Do not invent a new notebook if an existing one already matches the job.
2. Do not store raw Colab tokens in repo files.
3. Prefer bridge profiles and env exports over pasting secrets into notebooks.
4. Keep notebook selection explicit: pivot, finetune, qlora, webtoon, or cloud workspace.
5. Treat Colab as compute, not as the source of truth. The source of truth stays in repo files and datasets.

Output Contract

Every Colab operation should leave:

• the chosen notebook name and path
• the Colab URL or local bridge profile name
• any generated SFT/training artifacts
• a note about whether compute stayed local, Colab-only, or crossed into HF push