---

name: visual_projection_training_step description: Execute a deterministic tiny projection update that mirrors LLaVA's visual-token to language-token alignment step.

Visual Projection Training Step

Use this skill when a recovery needs executable evidence that the trainable projection matrix mechanism ran. It is appropriate for reduced/proxy experiments when full CLIP/Vicuna training is blocked.

Inputs

• Numeric visual feature vector.
• Numeric target language embedding vector.
• Initial scalar or matrix parameters.
• Learning rate and number of steps.

Outputs

A training trace with loss_before, loss_after, params_before, params_after, and optimizer_state_changed.

Workflow

1. Compute projected visual features with a linear map.
2. Compute mean squared error against the target language vector.
3. Apply deterministic gradient descent to trainable parameters.
4. Report parameter and loss changes for recovery validation.

Validation

Run the standard-library tests or the Distiller skill-tree validator.

Limitations

This script is a tiny mathematical proxy. It does not load CLIP, Vicuna, LLaMA, or real image tensors.