name: vpt_frozen_prompt_training
description: Configure and verify Visual Prompt Tuning trainability with frozen backbone and trainable prompt/head parameters.
VPT Frozen Prompt Training
Use this skill when a VPT implementation or recovery harness needs to decide which parameters are trainable, compute tunable-parameter ratios, or verify that frozen backbone parameters did not change. Do not use this skill to insert prompt tokens or compute classification accuracy.
Inputs
- Parameter records with
name, count, and optional before/after scalar values.
- Role hints or naming conventions identifying prompt, classifier head, and backbone parameters.
- Optional optimizer settings for prompt/head groups.
Outputs
- A trainability mask where prompt and head parameters are trainable and all backbone parameters are frozen.
- Optimizer parameter-group names containing no frozen backbone parameters.
- Tunable-count and tunable-percentage summaries.
- Freeze verification showing unchanged frozen parameters after a training step.
Workflow
- Classify each parameter by explicit role or by conservative name matching.
- Mark only prompt and classifier-head roles as trainable.
- Treat ambiguous Transformer, patch, attention, MLP, norm, and position parameters as frozen.
- Build optimizer groups from the trainable subset.
- Compute tunable percentage against total and frozen-backbone counts.
- Compare before/after values to prove frozen parameters did not mutate.
Validation
Run python tests/test_training_config.py, or use the Distiller validator with --run-tests.
Limitations
This skill validates trainability semantics and parameter accounting. It does not implement a full deep-learning framework optimizer.