Claude Skills by VectorSpaceLab

Build GeDi true/false control-code inputs for multi-class and zero-shot topic control with exactly two class-conditional passes.

Compute GeDi generative-discriminator class posteriors from control and anti-control log probabilities using stable Bayes-rule normalization.

Run and validate a bounded GeDi proxy recovery that invokes generated GeDi modules and records mechanism-faithful evidence.

Use this skill when you need to compose task formatting, model adapter calls, filters, and metrics into an auditable evaluation run.

Use this skill when you need to compute deterministic accuracy, exact match, token F1, and aggregate evaluation metrics.

Use this skill when you need to wrap a model backend behind loglikelihood, rolling loglikelihood, and generation request methods.

Use this skill when you need to post-process raw model outputs with regex extraction, take-first, take-k, or majority-vote pipelines.

Use this skill when you need to define reproducible lm-evaluation-harness style task configurations and format documents into model requests.

Construct prompted multitask training mixtures while enforcing held-out dataset boundaries for zero-shot evaluation.

Render structured dataset examples into PromptSource-style natural-language source and target pairs for prompted multitask recovery.

Execute a deterministic minimal optimizer step on prompted multitask text examples and log loss/parameter changes.

Compute held-out prompted-task accuracy and prompt wording consistency for T0/P3-style recovery experiments.

Use when a recovery or implementation needs differentiable PPLM attribute scores for BoW or simple discriminator-style control.

Use when generated candidates must be ranked by attribute likelihood with Dist-n diversity filtering as in PPLM BCR.

Use when a PPLM-style controller must perturb latent/logit state at inference time with attribute and KL losses while freezing the base model.

Use when validating a reduced or full PPLM recovery experiment with numeric metrics and mechanism checks.

Build and validate DDPM forward-process schedules and closed-form noisy samples for reduced recovery experiments.

Compute DDPM epsilon-prediction losses and weighted denoising objective terms for recovery training checks.

Run a bounded DDPM proxy experiment that exercises generated schedule, epsilon-loss, and reverse-step skills.

Convert predicted DDPM epsilon values into reverse-process means and deterministic or stochastic denoising samples.

Compute Conditional Flow Matching losses and bounded optimizer updates for a small parametric vector field.

Integrate learned Flow Matching vector fields with fixed-step ODE solvers and report trajectory/error diagnostics.

Compute Conditional Flow Matching squared vector-field losses and deterministic one-parameter proxy updates.

Construct and validate Flow Matching conditional optimal-transport paths and target vector fields for paired noise/data samples.

Build Gaussian and optimal-transport conditional Flow Matching path samples and target vector fields from noise/data pairs.

Run a bounded soft-mode Flow Matching proxy recovery that invokes generated OT-path, CFM-loss, and CNF-ODE skills.

Integrate simple Flow Matching vector fields and verify ODE trajectory, endpoint, and NFE evidence.

Evaluate whether a soft-mode reduced Flow Matching recovery is executable, metric-bearing, and mechanism-faithful.

Compute the analytic Gaussian bridge posterior q(Xt | X0, X1) for I2SB paired clean/degraded endpoints.

Build conditional flow matching interpolation states, target velocities, and squared regression losses for bounded CNF recovery experiments.

Compute deterministic squared-cost minibatch optimal transport pairings for OT conditional flow matching recovery runs.

Run bounded executable OT-CFM proxy recoveries with command logs, mechanism checks, and validator-compatible artifacts.

Build small entropy-regularized coupling checks and bridge variance schedules for SB conditional flow matching.

Construct uniform and exact small-batch BatchOT couplings for Multisample Flow Matching recovery and tests.

Compute Joint Conditional Flow Matching interpolation, target velocity, loss, and tiny optimizer updates for MSFM recovery.

Evaluate whether an MSFM recovery result exercised BatchOT coupling, Joint CFM loss, optimizer evidence, and soft-mode proxy constraints.

Run a bounded synthetic MSFM proxy experiment that invokes generated coupling and Joint CFM loss skills.

Build rectified-flow coupling interpolation records with endpoint displacement targets for velocity regression.

Simulate rectified-flow Euler transport and compute straightness, reflow, and convex transport diagnostics.

Compose rectified-flow module skills into an executable reduced recovery experiment with auditable JSON logs.

Fit and audit a small rectified-flow velocity model with least-squares regression and optimizer evidence.

Build continuous denoising score-matching targets and deterministic optimizer checks for score-SDE recovery.

Execute bounded reverse predictor-corrector sampling checks for score-SDE mechanism-faithful recovery.

Assemble executable soft-mode recovery artifacts for reduced score-SDE training and mechanism validation.

Define continuous-time score-SDE schedules, perturbation kernels, reverse drifts, and probability-flow checks for recovery experiments.

Build and validate SR3-style low-resolution conditioning paired with high-resolution targets for reduced or full recovery experiments.

Compute SR3-style conditional diffusion corruption, noise-prediction loss, and a deterministic scalar optimizer update.

Run a deterministic SR3-inspired reverse denoising trajectory conditioned on low-resolution input for proxy recovery.

Evaluate whether an SR3 soft-mode proxy recovery exercised conditioning, diffusion loss, optimizer update, and iterative refinement evidence.

Compute and optimize the SR3 conditional denoising objective for a lightweight trainable denoiser. This reusable skill supports bounded SR3 recovery experiments.