Claude Skills by VectorSpaceLab

Compare fixed and optimized QoS allocations using delay/loss SLA satisfaction and mechanism checks for reduced recovery.

Generate random-start l-infinity PGD adversarial examples with projection diagnostics for first-order robust optimization experiments.

Run a deterministic PGD adversarial training loop that updates model parameters on adversarial examples and logs loss traces.

Specify and validate the robust min-max objective used for PGD adversarial training from Madry et al. 2017.

Evaluate natural and PGD adversarial metrics and emit validator-compatible recovery evidence for robust optimization experiments.

Run a deterministic APGD-style projected attack with loss-history step adaptation and auditable traces.

Evaluate robust accuracy with a fixed sequential AutoAttack-style attack protocol and per-attack accounting.

Compute DLR-style logit margin losses for APGD-DLR adversarial optimization checks.

Project adversarial candidates into declared norm balls and valid input boxes for robustness evaluation.

Compute CLIP-style normalized similarity logits and symmetric contrastive loss for image-text batches.

Build and validate CLIP-style image-text pair records and prompt templates for bounded recovery experiments.

Build prompt-ensemble zero-shot classifiers from normalized image and text embeddings in the CLIP style.

Run a bounded soft-mode CLIP proxy recovery that logs contrastive training and zero-shot evaluation evidence.

Format LLaVA-style multimodal prompts with image tokens, role boundaries, and answer-leakage checks.

Canonicalize option-only or short-form ScienceQA-style answers and compute deterministic accuracy.

Build LLaVA-style visual instruction records from captions and boxes for mechanism-faithful recovery experiments.

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

Compute CLIP-style normalized image/text similarity and symmetric InfoNCE loss for paired multimodal embeddings.

Fit log-log power laws for CLIP downstream error metrics over compute or scale points.

Evaluate CLIP-style zero-shot classification and cross-modal retrieval with normalized similarity rankings.

Run an auditable reduced OpenCLIP scaling-law recovery by invoking generated skills and producing validator-compatible recovery artifacts.

Build and validate OpenCLIP-style scale tables for CLIP scaling-law experiments, including total compute and error metric derivation.

Normalize POPE yes/no answers and compute object hallucination classification metrics.

Select POPE absent-object negatives using random, popular, and adversarial strategies for object hallucination evaluation.

Build POPE polling-question datasets from image object annotations and prompt templates.

Run a bounded POPE proxy recovery experiment with auditable generated-skill invocations and metrics.

Generate bounded text-conditioned adversarial feature perturbations for reduced TeCoA robustness recovery.

Run an executable reduced TeCoA recovery experiment with generated skill invocations and mechanism checks.

Compute TeCoA image-to-text contrastive logits, loss, margins, and accuracy for robustness recovery.

Build deterministic CLIP-style zero-shot text prompts for TeCoA image-text adversarial robustness experiments.

Run auditable soft-mode visual jailbreak proxy recovery using generated skills and source-boundary logs.

Compute held-out benign-versus-adversarial obedience deltas and mechanism checks for safe jailbreak proxies.

Build and validate harmless corpus splits that mirror visual jailbreak optimization and held-out evaluation protocols.

Optimize continuous visual prompt vectors with projected gradient descent and auditable constraint diagnostics.

Construct masked universal input programs for adversarial reprogramming of frozen neural networks.

Aggregate frozen target model outputs into adversarial task labels using one-to-one or many-to-one mappings.

Optimize only universal program parameters through a frozen model for adversarial reprogramming recovery experiments.

Validate mechanism evidence for adversarial reprogramming proxy or full recovery experiments.

Run a reduced AutoVP prompt optimization step while proving the source classifier remains frozen.

Select AutoVP candidate configurations and record benchmark-style accuracy targets for recovery runs.

Convert frozen source-model logits or predictions to target classes with AutoVP label mapping strategies.

Build AutoVP frame-shaped visual prompt masks and apply bounded universal prompts to resized images.

Produce validator-ready BAR recovery metrics and mechanism checks for full or declared reduced black-box reprogramming experiments.

Train BAR universal-program parameters with query-only one-sided zeroth-order gradient estimates and SGD updates.

Build BAR-style masked target embeddings and universal tanh-bounded programs while preserving protected target-domain data.

Construct and apply BAR multi-label mappings that average groups of source-model probabilities into target-label probabilities.

Build CLIP-style text label prompts and classify frozen image embeddings by cosine similarity.

Optimize only universal visual prompt parameters while frozen model components remain unchanged.

Run a bounded mechanism-faithful proxy recovery experiment for visual prompting papers.

Create and validate task-level pixel visual prompt templates for frozen vision model adaptation.