Claude Skills by VectorSpaceLab

Run a reduced executable QDagger recovery experiment with generated-skill invocations, training traces, and mechanism checks.

Compute the QDagger objective for policy-to-value reincarnating RL using n-step TD loss plus teacher-policy distillation.

Compute QDagger teacher-distillation coefficient schedules for weaning a value student off a suboptimal teacher.

Compute reduced APPO update evidence with PPO clipping and V-trace-style value correction for asynchronous policy lag.

Design and validate Sample Factory-style asynchronous rollout, policy, and learner topology with policy-lag estimates.

Model Sample Factory double-buffered environment sampling and compare idle-time estimates against synchronous sampling.

Specify Sample Factory-style multi-policy self-play assignments and population-based exploit/mutate decisions.

Compute influence-style hypergradients and projected unlabeled weight updates from validation gradients, per-example unsupervised gradients, and a damped Hessian approximation.

Apply masked Adam-style sparse updates so only selected per-example unlabeled weights and their optimizer moments change during SSL hyperparameter optimization.

Run a bounded reduced semi-supervised recovery experiment that invokes generated protocol, influence, and masked optimizer skills and emits validator-compatible evidence.

Construct deterministic semi-supervised splits with labeled, validation, unlabeled, pseudo-label, and per-example weight fields for SSL recovery experiments.

Run a bounded CCS proxy experiment and emit validator-compatible recovery evidence.

Select CCS-style coresets with mislabel filtering and deterministic score-stratified coverage.

Convert per-example training dynamics into CCS-compatible importance scores for coreset selection.

Run a bounded soft-mode recovery experiment for EL2N/GraNd data pruning with executable mechanism evidence.

Compute EL2N and GraNd example-importance scores for supervised classification pruning experiments.

Select retained and pruned supervised-training examples from EL2N or GraNd scores using deterministic pruning rules.

Select removable data subsets by constraining aggregate influence vector norm instead of ranking independent scalar scores.

Validate paper-target metadata, aggregate influence tolerance, and observed pruning metric gaps for generalization-influence pruning.

Estimate per-example parameter influence vectors for dataset pruning using gradients and damped inverse-Hessian approximations.

Run a bounded end-to-end proxy recovery for generalization-influence dataset pruning using generated module skills.

Compute nested historical lexi-targets from objective vectors, goals, and tolerances for prioritized minimization objectives.

Run a bounded randomized direct-search loop that samples paired directions, updates by targeted lexicographic comparisons, and tracks the best historical point.

Evaluate whether a recovery run exercised LexiFlow mechanisms and achieved a declared proxy or benchmark target with auditable metrics.

Compare two objective vectors under target-aware equality and strict preference while preserving vanilla lexicographic tie breaking.

Select Moderate-DS coresets by keeping examples whose scalar scores are closest to the score median.

Run a bounded Moderate-DS proxy recovery harness with generated skill invocations and optimizer evidence.

Compute class-center representation distances for Moderate-DS style data selection from labeled feature records.

Compare Moderate-DS median selection against center-close, far-from-center, and two-end score policies.

Project coreset probability vectors onto the box-constrained simplex budget used by probabilistic bilevel coreset selection.

Apply the score-function outer update for probabilistic bilevel coreset selection without implicit differentiation.

Build Bernoulli coreset mask probabilities, sample masks, and compute score-function gradients for probabilistic bilevel coreset selection.

Run a bounded mechanism-faithful proxy experiment for probabilistic bilevel coreset selection with executable evidence.

Validate paired ID/OOD accuracy records before fitting accuracy-on-the-line calibration experiments.

Fit and inspect linear calibration between in-distribution and out-of-distribution accuracies.

Decide whether a declared soft-mode proxy exercises the Accuracy on the Line mechanism.

Fit the probit-space ID-to-OOD agreement line and report agreement-on-the-line diagnostics without OOD labels.

Compute ID accuracy, pairwise ID/OOD agreement, and clipped probit statistics for agreement-on-the-line experiments.

Solve the ALine-D pairwise linear system to predict OOD accuracies from ID accuracy and unlabeled agreement statistics.

Validate agreement-on-the-line prediction tables while separating estimator inputs from evaluation-only OOD labels.

Run and evaluate a bounded agreement-on-the-line proxy recovery with source-boundary and mechanism checks.

Compute and test CLIP-style symmetric contrastive loss with normalized image-text similarities.

Validate CLIP-style paired image-text batches and emit diagonal contrastive labels for recovery experiments.

Coordinate bounded proxy recovery evidence for CLIP contrastive pretraining and zero-shot transfer.

Build natural-language prompt classifiers and evaluate zero-shot predictions from embeddings.

Validate and normalize ImageNetV2-style candidate pools with class labels, MTurk selection frequencies, and ranked model predictions.

Compute ImageNetV2-style top-k accuracy and drops from sampled candidate records with labels and ranked predictions.

Fit and inspect original-vs-new ImageNetV2 accuracy relationships for rank preservation and linear generalization gaps.

Sample class-balanced ImageNetV2 proxy datasets using MatchedFrequency, Threshold0.7, or TopImages selection-frequency strategies.

Create deterministic semantic-preserving texture and style perturbations for bounded DeepAugment-style recovery tests.