Claude Skills by VectorSpaceLab

Train an online text classifier with sparse replay updates from episodic memory.

Evaluate held-out direct and CoT instruction examples, compute accuracy delta, and emit FLAN recovery mechanism checks.

Build FLAN-style multi-source instruction-finetuning mixtures while excluding held-out evaluation tasks and auditing source/CoT coverage.

Render FLAN-style direct, few-shot, and chain-of-thought instruction prompt/completion pairs without leaking hidden answers into prompts.

Execute a bounded optimizer-based proxy for FLAN instruction finetuning and record loss plus parameter-change evidence.

Render FLAN-style instruction examples with direct, exemplar, and chain-of-thought formatting controls.

Apply LoRA parameter freezing and compact checkpoint filtering for reduced adaptation experiments and audits.

Assemble and validate a soft-mode reduced LoRA recovery with executable mechanism evidence and source-boundary logs.

Run a deterministic reduced LoRA training step that updates only low-rank factors and records loss evidence.

Construct LoRA low-rank linear adapters and verify merged inference equivalence without relying on the original implementation repository.

Score latent replay candidates with KL drift, entropy confidence penalties, and diversity filtering for MIR.

Run bounded MIR recovery experiments and emit accuracy, forgetting, traces, and mechanism evidence.

Build bounded online continual-learning streams, replay memories, and forgetting ledgers for MIR-style experiments.

Rank replay candidates by loss increase under a virtual incoming-batch update for MIR selection.

Compute MMLU option-label accuracy and confidence calibration gap from model predictions.

Build MMLU zero-shot and five-shot prompts while hiding the test answer from the model query.

Validate and canonicalize MMLU-style four-option subject examples for evaluation workflows.

Run a bounded MMLU proxy recovery using generated item, prompting, and scoring skills.

Materialize PromptSource/P3-style template collections into prompted dataset records with template identity, metadata propagation, coverage counts, and rendering diagnostics.

Apply PromptSource-style templates across multiple examples and summarize browse, sourcing, and review diagnostics for prompt iteration.

Validate PromptSource-style prompt metadata and community quality constraints for natural-language prompts, metrics, answer choices, and target format.

Render PromptSource-style Jinja prompt templates into natural-language input/target pairs with answer choices, deterministic choices, skip handling, and structured diagnostics.

Fast, deterministic feedback for PassNet work: pre-flight pattern/match verification WITHOUT burning a GPU evaluation, per-node bottleneck analysis of a sample's graphs, and eval-log parsing into per-variant status + estimated score + failure classification. Use BEFORE every GPU evaluation (check_pattern), at the START of a sample (analyze_graph), and AFTER every evaluation (parse_eval_log).

Round/batch planning and decision support for PassNet graph optimization. Use for multi-sample triage, worker allocation, eval-budget policy, and post-eval keep/revert decisions. For solving one concrete sample end to end, use passnet-solve as the entry skill; consult this skill only when you need broader planning or a specific decision gate.

Author PassNet pass files: write patterns that actually MATCH the FX graph, pick fusion regions (which ops to absorb into one Triton kernel), and structure multi-pass submissions with the mandatory shared-dispatch architecture. Use when creating or fixing pattern()/replacement_args()/replacement_func() files, when a pass "failed to match", or when deciding how to fuse multiple kernels/ops.

PassNet GPU kernel optimization via compiler passes. Design and implement Triton-based optimization passes, create pass files under ./pass_dir/, self-evaluate with pass_evaluator, and iterate to maximize GPU speedup.

END-TO-END playbook for solving one PassNet sample: analyze the computation graphs, decide the optimization strategy (what to fuse, what to replace, what to leave alone), drive the iteration loop, and maximize the sample score. This is the ENTRY skill — invoke it first for any PassNet optimization task; it tells you when to use passnet-pattern-fusion, passnet-triton-opt and passnet-feedback.

Make a single PassNet Triton kernel fast AND numerically faithful: performance model (when a replacement can win at all), block/grid/warp tuning, autotune policy, launch overhead, and per-op numeric recipes to pass the dtype baseline tolerances. Use when a pass matches and is correct but speedup below expected, or when correctness fails by small numeric margins.