---

name: model-internals-and-extension description: "Guides OFA task, model, and criterion registration plus prompt tuning, adapters, bitfit, and encouraging-loss extensions." disable-model-invocation: true metadata: disco-role: operating license: Apache 2.0

model-internals-and-extension

Use this sub-skill when a user wants to inspect OFA internals, add a new task or criterion, or reason about prompt tuning and other extension flags.

Trigger phrases

• "How is OFA registered?"
• "Add a new task / criterion / model"
• "What does ofa_module do?"
• "How do prompt tuning, adapters, or bitfit work?"
• "How do I use encouraging loss?"

What this sub-skill owns

• OFA task/model/criterion registration,
• architecture names and shared model configuration,
• classification-head and generator wiring,
• prompt tuning, adapters, and bitfit guidance,
• encouraging-loss usage and extension planning.

What it excludes

• end-to-end caption/VQA/RefCOCO/OCR/ImageNet commands -> vision-language-tasks,
• Gigaword / GLUE -> language-tasks,
• MMSpeech data prep or stage commands -> mmspeech,
• generic launch or environment setup -> setup-and-command-building.

Read these files

• references/api-reference.md for the main OFA classes, registries, and architecture fields.
• references/criteria-and-tuning.md for prompt tuning, adapters, bitfit, and encouraging loss.
• references/troubleshooting.md for registry and checkpoint mismatch failures.
• scripts/inspect_ofa_registration.py to print the live task/model/criterion registry.

Typical workflow

1. Inspect the registry to confirm the task or criterion already exists.
2. Read the API reference before adding a new extension point.
3. Decide whether the change belongs in prompt tuning, adapters, bitfit, or a new criterion.
4. Check the troubleshooting notes for checkpoint and shell-command pitfalls.

Notes

• The repo uses Fairseq-style registration and architecture decorators.
• Prompt tuning and adapter-style changes are usually command flags, not a separate package.
• A registry check is much safer than assuming the model was imported correctly.