Claude Skills by VectorSpaceLab

Build paired low/high-resolution items and cumulative diffusion schedules that match SR3 equations. This reusable skill supports bounded SR3 recovery experiments.

Apply the SR3 reverse-chain update from Gaussian noise to a refined high-resolution image. This reusable skill supports bounded SR3 recovery experiments.

Evaluate SR3-style proxy runs with consistency metrics, source-boundary records, and mechanism checks. This reusable skill supports bounded SR3 recovery experiments.

Integrate reduced InterFlow velocity fields and summarize executable mechanism-faithful recovery diagnostics.

Compute and test the simulation-free quadratic velocity objective used by stochastic interpolant flows.

Construct endpoint-respecting stochastic interpolant samples and time derivatives for InterFlow recovery experiments.

Compute stochastic-interpolant velocity, denoiser, and score objective diagnostics from sampled tuples without evaluating densities.

Construct endpoint-valid stochastic interpolant samples with latent schedules and derivative targets for reduced or full generative recovery experiments.

Run a deterministic reduced Gaussian-mixture recovery that exercises stochastic-interpolant construction, objectives, denoising, and sampling evidence.

Integrate bounded stochastic-interpolant probability-flow ODE and simple diffusion sampler updates from learned velocity and score fields.

Run a bounded CMA-ES proxy experiment that validates sampling, selection, CSA, and covariance adaptation evidence.

Compute CMA-ES default strategy parameters from the tutorial formulas for bounded recovery experiments.

Update CMA-ES cumulative step-size paths and covariance matrix from selected normalized steps.

Sample and select CMA-ES Gaussian offspring for derivative-free black-box optimization recovery.

Compute ImageNet-C CE, mCE, relative CE, and relative mCE from clean and corrupted top-1 error tables.

Apply ImageNet-C-style common image corruptions with five severity levels and reproducible metadata for robustness experiments.

Run a bounded mechanism-faithful ImageNet-C proxy recovery using generated corruption and metric skills without reading the original repository.

Compute ImageNet-P-style prediction flip probability over ordered perturbation sequences.

Use this skill to calibrate post-training quantization scales for a layer or matrix multiplication using deterministic candidate search.

Use this skill to rank quantization candidates with the PTQ4ViT Hessian-guided squared-gradient reconstruction metric.

Use this skill to run a bounded PTQ4ViT proxy recovery that invokes generated scale-search, twin-quantization, and Hessian-metric skills.

Use this skill to apply PTQ4ViT twin uniform quantization to post-softmax or post-GELU activations with range flags.

Use this skill when tracking SAR entropy EMA and deciding whether a model or optimizer reset is required after collapse.

Use this skill when selecting trainable parameters for SAR-style test-time adaptation while freezing non-normalization model weights.

Use this skill when implementing SAR reliable entropy minimization, entropy-margin filtering, or proxy checks for selected samples.

Use this skill when implementing the SAM-like sharpness-aware two-step update used by SAR recovery experiments.

Use this skill when a recovery or experiment needs to construct SAR-style dynamic wild-world test streams with mixed domains, small batches, and label imbalance.

Run and validate Tent-style entropy minimization updates on unlabeled target batches with trainable modulation parameters.

Configure normalization layers for Tent by exposing only affine modulation parameters and target batch statistics.

Evaluate Tent recovery artifacts for numeric metrics, source-boundary compliance, and mechanism-faithful proxy evidence.

Validate fully test-time adaptation experiments so unlabeled target inputs are the only data used for adaptation updates.

Select and report Visual Prompt Tuning classification results with validation accuracy and parameter efficiency.

Configure and verify Visual Prompt Tuning trainability with frozen backbone and trainable prompt/head parameters.

Run a bounded soft-mode Visual Prompt Tuning proxy that proves prompt insertion, frozen training, and evaluation mechanisms.

Build shallow and deep Visual Prompt Tuning token sequences for Transformer-style image embeddings.

Create task-level seen and unseen splits with leakage diagnostics for Natural Instructions style generalization experiments.

Render Natural Instructions task records into paper-faithful instruction encoding variants for text-to-text models.

Encode Natural-Instructions-style task instructions and inputs into leakage-safe model prompts for cross-task generalization recovery.

Normalize crowdsourcing task instructions into validated Natural Instructions records for cross-task generalization experiments.

Run a bounded soft-mode proxy recovery harness for BART0 cross-task generalization using generated skills and validator-compatible logs.

Execute a bounded instruction-conditioned training proxy and ROUGE-L evaluation for BART0-style recovery.

Compute deterministic ROUGE-L style generation metrics for Natural Instructions recovery outputs and ablations.

Generate and validate fixed-cardinality alpha-subset membership matrices for datamodel training and recovery experiments.

Estimate training-data removal effects by summing linear datamodel weights and evaluating counterfactual correlations.

Run a bounded synthetic proxy recovery that validates linear datamodel prediction and counterfactual mechanisms with executable evidence.

Fit target-specific linear datamodel surrogates from subset membership vectors to observed target outcomes.

Evaluate mechanism-faithful continual-learning recovery metrics and gate proxy evidence.

Maintain frozen-key episodic memory with random write, random replay sampling, and Euclidean KNN retrieval.

Build one-pass lifelong language-learning streams without exposing dataset identifiers to the model.

Retrieve memory neighbors and adapt temporary parameters for a single prediction while preserving base parameters.