Machine Learning Skills

Profiler-driven hill-climbing to close the inference throughput gap between TorchTitan's unified model (running inside vLLM) and vLLM's native model. Benchmark with generate.py --benchmark, climb optimization rungs (compile / cudagraph / fused kernels), profile torchtitan vs the native target, then patch the single biggest gap at a time and re-measure. Use when the user wants to benchmark or optimize RL inference generation speed, reproduce previous hill climbing study, or invokes /inference_perf_hillclimb.

> This skill should be used when the user wants to "run an evaluation", "evaluate my ADK agent", "write an eval dataset", "analyze eval failures", "compare eval results", "optimize agent", or needs guidance on the Agent Platform eval methodology and the Quality Flywheel. Covers eval metrics, dataset schema, LLM-as-judge scoring, and common failure causes. Do NOT use for API code patterns (use google-agents-cli-adk-code), deployment (use google-agents-cli-deploy), or project scaffolding (use google-agents-cli-scaffold).

Deep-crawl any website from start URLs, return per-page LLM-ready text/markdown/HTML plus metadata (title, description, author, language, canonical URL, OG) and in-scope outbound links. Use when user mentions deep crawl website, recursive crawl, crawl a whole site, scrape entire website, scrape docs site, scrape documentation, scrape knowledge base, scrape blog, build RAG corpus, build vector database from website, knowledge base for chatbot, GPT knowledge files, llms.txt, sitemap crawl, BFS crawl, scrape with depth or page limit, include exclude URL globs, remove boilerplate, strip navigation header footer, website to markdown, website to text, multi-page extraction, bulk page scraping, clean markdown from URL, docs site to markdown corpus, site to clean corpus. Also applies to building RAG pipelines, indexing a customer site, syncing docs into a vector store, generating training corpora from any docs hub, or expanding a single start URL into a clean corpus of every reachable in-scope page.

Trustpilot customer reviews scraper for any company listed on trustpilot.com — given a company domain (e.g. shopify.com, apple.com, shopwagandtail.com) plus optional filters (page number, single star rating 1-5, single language ISO code, verified-only flag, has-company-reply flag, date period last30days/last3months/last6months/last12months, free-text keyword search, and reviewer countryCode client-side filter), extracts paginated reviews with 20+ fields per review: reviewId, reviewUrl, reviewTitle, reviewDescription, reviewRatingScore (1-5), reviewDate (ISO publishedDate), reviewDateOfExperience, reviewLabel, isReviewVerified, reviewer (display name), reviewerId, reviewersCountry (ISO Alpha-2), reviewLanguage, reviewCompanyResponse (the company reply text), plus pagination metadata (currentPage, totalPages, totalCount). Optional flags add reviewer profile metadata (reviewerNumberOfReviews, reviewerProfileIsVerified), reply analysis dates (companyReplyPublishedDate, companyReplyUpdatedDate), extended metadata (reviewSource, reviewVerificationSource, reviewLikes), and review photos. Use when user mentions Trustpilot reviews, scrape Trustpilot, get reviews from Trustpilot, Trustpilot review scraper, Trustpilot customer reviews, extract Trustpilot reviews, Trustpilot review data, download Trustpilot reviews, paginate Trustpilot, paginated Trustpilot reviews, filter Trustpilot reviews by star, filter Trustpilot by language, verified reviews Trustpilot, has reply Trustpilot, company response Trustpilot, Trustpilot reply analysis, brand monitoring Trustpilot, sentiment analysis Trustpilot, competitor reviews Trustpilot, bulk reviews from a Trustpilot page, batch download reviews, trustpilot.com reviews scraping, fake review research Trustpilot, Trustpilot dataset, persona research from Trustpilot reviews, Trustpilot keyword search reviews, Trustpilot date range reviews, review export Trustpilot. Also applies to comparing review trends across competitors, training data for LLM sentiment classifiers, monitoring competitor response time, exporting reviews for BI dashboards, lead generation from active reviewers, ML / fake-review detection, and aggregator sites that display third-party reviews.

Hugging Face Hub CLI (`hf`) for downloading, uploading, and managing repositories, models, datasets, and Spaces on the Hugging Face Hub. Replaces now deprecated `huggingface-cli` command.

Use this skill for Hugging Face Dataset Viewer API workflows that fetch subset/split metadata, paginate rows, search text, apply filters, download parquet URLs, and read size or statistics.

This skill should be used when users want to train or fine-tune language models using TRL (Transformer Reinforcement Learning) on Hugging Face Jobs infrastructure. Covers SFT, DPO, GRPO and reward modeling training methods, plus GGUF conversion for local deployment. Includes guidance on the TRL Jobs package, UV scripts with PEP 723 format, dataset preparation and validation, hardware selection, cost estimation, Trackio monitoring, Hub authentication, and model persistence. Should be invoked for tasks involving cloud GPU training, GGUF conversion, or when users mention training on Hugging Face Jobs without local GPU setup.

This skill should be used when users want to run any workload on Hugging Face Jobs infrastructure. Covers UV scripts, Docker-based jobs, hardware selection, cost estimation, authentication with tokens, secrets management, timeout configuration, and result persistence. Designed for general-purpose compute workloads including data processing, inference, experiments, batch jobs, and any Python-based tasks. Should be invoked for tasks involving cloud compute, GPU workloads, or when users mention running jobs on Hugging Face infrastructure without local setup.

Use Transformers.js to run state-of-the-art machine learning models directly in JavaScript/TypeScript. Supports NLP (text classification, translation, summarization), computer vision (image classification, object detection), audio (speech recognition, audio classification), and multimodal tasks. Works in Node.js and browsers (with WebGPU/WASM) using pre-trained models from Hugging Face Hub.

Submit compact CELLxGENE Discover API requests for public collection and dataset metadata. Use when a user wants concise single-cell collection summaries

Submit compact NCBI Datasets v2 requests for assembly, genome, taxonomy, and related metadata endpoints. Use when a user wants concise NCBI Datasets summaries; save raw JSON or text only on request.

Submit compact ProteomeXchange PROXI requests for datasets, libraries, peptidoforms, proteins, PSMs, spectra, and USI examples. Use when a user wants concise PROXI summaries

Create new skills, modify and improve existing skills, and measure skill performance. Enhanced version with quick commands. Use when users want to create a skill from scratch, update or optimize an existing skill, run evals to test a skill, benchmark skill performance with variance analysis, or optimize a skill's description for better triggering accuracy. Triggers on phrases like "make a skill", "create a new skill", "build a skill for", "improve this skill", "optimize my skill", "test my skill", "turn this into a skill", "skill description optimization", or "help me create a skill".

Guides technology selection and implementation of AI and ML features in .NET 8+ applications using ML.NET, Microsoft.Extensions.AI (MEAI), Microsoft Agent Framework (MAF), GitHub Copilot SDK, ONNX Runtime, and OllamaSharp. Covers the full spectrum from classic ML through modern LLM orchestration to local inference. Use when adding classification, regression, clustering, anomaly detection, recommendation, LLM integration (text generation, summarization, reasoning), RAG pipelines with vector search, agentic workflows with tool calling, Copilot extensions, or custom model inference via ONNX Runtime to a .NET project. DO NOT USE FOR projects targeting .NET Framework (requires .NET 8+), the task is pure data engineering or ETL with no ML/AI component, or the project needs a custom deep learning training loop (use Python with PyTorch/TensorFlow, then export to ONNX for .NET inference).

> Activate this skill when BenchmarkDotNet (BDN) is involved in the task — creating, running, configuring, or reviewing BDN benchmarks. Also activate when microbenchmarking .NET code would be useful and BenchmarkDotNet is the likely tool. Consider activating when answering a .NET performance question requires measurement and BenchmarkDotNet may be needed. Covers microbenchmark design, BDN configuration and project setup, how to run BDN microbenchmarks efficiently and effectively, and using BDN for side-by-side performance comparisons. Do NOT use for profiling/tracing .NET code (dotnet-trace, PerfView), production telemetry, or load/stress testing (Crank, k6).

Export a PyTorch model to .pte format for ExecuTorch. Use when converting models, lowering to edge, or generating .pte files.

Search the ExecuTorch tribal knowledge base covering QNN, XNNPACK, Vulkan, CoreML, Arm, and Cadence backends, quantization recipes, export pitfalls, runtime errors, and SoC compatibility. Use when debugging ExecuTorch errors, choosing quantization configs, checking backend op support, or answering questions about Qualcomm HTP / Snapdragon / Apple Neural Engine behavior.

Run comprehensive quality checks on tabular data (CSV/Excel/TSV/JSON), detecting missing values, duplicates, outliers, format issues, and type inconsistencies to produce an overall score, grade, and actionable suggestions. Triggered when users ask to check data quality, find missing or duplicate values, detect outliers, validate formats, profile data, or clean data.

Run stepped HTTP load tests with ab/wrk, ramping concurrency levels to collect p50/p90/p99 latency, detect performance inflection points, and recommend optimal concurrency. Triggered by requests like 'load test this URL', 'benchmark my API', 'find the max concurrency', or mentions of p99 latency, throughput saturation, or capacity planning.

Retrieve information from simulator manual and example DATA files. Use when answering keyword format questions, syntax queries, or when looking up official documentation and working examples. Essential for understanding keyword definitions, parameter tables, and concrete usage patterns.

LLM Tuning Patterns

Generate Hugging Face Hub (huggingface_hub) release notes from cached PR JSON files. Use when asked to draft release notes from PR files.

Use when adding or migrating non-thumbnail images for a Hugging Face Blog post. Uploads body images to the Hugging Face documentation-images dataset, updates the Markdown, and verifies the new links.

Establish baseline-vs-candidate evaluation plans, delegate missing evaluations, compare validated results, and decide quantization feasibility. Use when the user asks to compare baseline vs quantized runs, explain an accuracy drop/regression, verify whether a quantized checkpoint is acceptable, or compare NEL/MLflow evaluation outputs. Do NOT use for generic single-model evaluation without comparison intent (use evaluation), live NEL status/debugging (use launching-evals), or generic MLflow browsing without a comparison goal (use accessing-mlflow).

> Add a new model to the EAGLE3 offline pipeline. Generates an hf_offline_eagle3.yaml launcher config for a new model checkpoint, choosing the right hidden state dump backend (TRT-LLM / HF / vLLM) and GPU configuration. Use when user wants to run EAGLE3 on a model that does not yet have a YAML in tools/launcher/examples/ or asks how to configure the pipeline for a new checkpoint.

Serve a quantized or unquantized LLM checkpoint as an OpenAI-compatible API endpoint using vLLM, SGLang, or TRT-LLM. Use when user says "deploy model", "serve model", "start vLLM server", "launch SGLang", "TRT-LLM deploy", "AutoDeploy", "benchmark throughput", "serve checkpoint", or needs an inference endpoint from a HuggingFace or ModelOpt-quantized checkpoint. Do NOT use for quantizing models (use ptq) or evaluating accuracy (use evaluation).

>- Inspect Hugging Face decoder layers on meta tensors and plan or run per-rank BF16, FP8, and NVFP4 GEMM or fused-MoE microbenchmarks with the bundled scripts and a local FlashInfer checkout. Use when choosing a model, GPU, TP, EP, or M/token-concurrency sweep; deriving common fused QKV and gate/up shapes without loading checkpoint weights; or using FlashInfer benchmark utilities. Do not use for end-to-end server throughput or request latency.

> Validate that an EAGLE3 pipeline run completed successfully end-to-end. Checks all 4 steps produced expected artifacts, verifies acceptance rate meets threshold (>= 2.1), and produces a summary report. Use when user wants to verify a pipeline run or check benchmark results.

> Triage a failed EAGLE3 pipeline run. Identifies which step failed (data synthesis, hidden state dump, training, or benchmark), diagnoses root cause from logs, and suggests fixes. Use when user reports an EAGLE3 pipeline failure or asks why a specific step failed. Also helps debug new model support issues.

>- Use when the user asks to "quantize a model", "run PTQ", "post-training quantization", "NVFP4 quantization", "FP8 quantization", "INT8 quantization", "INT4 AWQ", "quantize LLM", "quantize MoE", "quantize VLM", or needs to produce a quantized HuggingFace checkpoint from a pretrained model using ModelOpt. Do NOT use for multi-candidate recipe exploration or optimization (use quant-recipe-search).

Hunt LLM/AI feature bugs — prompt injection, indirect injection, exfiltration via tool-use/markdown, ASCII smuggling, agentic AI security (OWASP Agentic Apps 2026, ASI01-ASI10). Patterns: direct injection ('ignore previous instructions'), indirect injection via documents/web pages/email the model reads, ASCII smuggling (Unicode Tags block U+E0000-U+E007F, invisible to humans, decoded by the model), tool-use exfiltration (model has fetch/browse tool, attacker injects OOB URL, model exfils chat history/secrets), markdown-image zero-click exfil, system-prompt extraction, IDOR-via-AI (cross-tenant data). Targets: chatbots, RAG, summarizers, agentic copilots, MCP tools. Detection: any LLM-backed endpoint, doc upload triggering AI processing, autonomous agent with tools. Validate: OOB/Collaborator callback for exfil, verbatim-reproducible system-prompt leak (run twice), verifiable cross-tenant leak or RCE. Confabulation is NOT a finding. Use when hunting AI features, chatbots, RAG, agentic systems, MCP.

Hunt vector-store / embedding-layer weaknesses in RAG pipelines (OWASP LLM08 Vector and Embedding Weaknesses) — persistent corpus poisoning that survives across sessions and users (distinct from one-shot indirect prompt injection, which is owned by hunt-llm-ai), cross-tenant vector-database IDOR (unauthenticated or unscoped queries against Pinecone/Weaviate/Chroma/Milvus/Qdrant/pgvector), source-text/metadata leakage in similarity-search results, and retrieval-hijack via adversarial embedding proximity ('SEO poisoning' for RAG). Targets: any app with a shared knowledge base, document upload feeding a chatbot, or a directly reachable vector-DB port. Validate: a second, clean session/account must inherit a poisoned result, or a cross-tenant artifact must be independently verifiable — confabulation is not a finding, same bar as hunt-llm-ai. Use when target is RAG-backed, exposes a vector-DB port, or lets users upload documents that other users' queries later retrieve.

Search 2500+ curated ChatGPT and LLM open-source repositories. Use when the user asks to find tools, libraries, or repos related to ChatGPT, LLMs, RAG, agents, langchain, NLP, AI development, or any open-source AI tooling.

Use when creating benchmark scenarios for new openzl codec nodes in unitBench - adding kernel-level encode/decode benchmarks or graph-level compress/decompress benchmarks for codecs like bitsplit, delta, transpose, entropy, etc.

Official NVIDIA-authored guidance for navigating PhysicsNeMo — pick the model, datapipe, or example for a SciML/AI4Science task (surrogates, forecasting, downscaling, physics-informed, inverse, generative). Points at existing files via live repo search; never writes code. Do NOT use for installation or environment setup, training-loop or other code authoring/scaffolding, contributor/CI/packaging questions, repo-specific questions in physicsnemo-sym/-cfd/-curator, or general (non-physics) ML/PyTorch.

> Analyze SaaS company valuation compression between funding rounds. Use this skill between rounds, why the ARR multiple compressed or expanded, comparing a company's compression to macro benchmarks, or explaining what drove valuation changes for any VC-backed software company. Trigger on phrases like "valuation compression", "ARR multiple", "round-to-round valuation", "multiple change", or when the user asks to compare a company's funding rounds. Always use this skill for any multi-round SaaS valuation analysis — do not try to answer from memory alone.

> Look up Y Combinator companies, batches, and startup ecosystem data using the yc-oss API (read-only). Use this skill whenever the user wants to research YC-backed startups, find companies in a specific batch or industry, check which YC companies are hiring, explore top YC companies, or analyze startup trends by sector or tag. "top Y Combinator companies", "find YC companies tagged AI", "W25 batch", "S24 companies", "YC stats", "Y Combinator portfolio", "startup research", "which YC companies do X", "venture research on YC", any mention of Y Combinator, YC batch, or YC-backed companies in the context of startup research, venture analysis, or market intelligence. This is a read-only data source — the API is a static JSON dataset updated daily.

Update and maintain the llms.txt file. Use when the user wants to add sections, add links, correct information, or review the llms.txt file.

> Build and maintain a topic graph over your agent session history. Reads every Claude session transcript plus the pruned sessions that survive only in history.jsonl, clusters them by topic using local TF-IDF (no API calls, no embeddings), and writes an interactive graph you can open in a browser. Use when the user says "/session-brain", "build my session map", "cluster my claude sessions", "map my session history", "rebuild the session graph", "show me my session graph", "what have I been working on lately", "what topics have gone stale". Different from over the raw sessions and never writes to the vault.

Build torch-tensorrt locally — install/pin the matching PyTorch nightly, drive Bazel through setup.py, do a clean rebuild after libtorch ABI changes, and recover from common build failures (undefined symbol, stale _C.so, libtorchtrt.so missing). Invoke whenever the user asks to build, rebuild, install editable, or upgrade the torch nightly; or when an import fails with an undefined-symbol error tying torch_tensorrt to libtorch.

Provides AI and machine learning techniques for CTF challenges. Use when attacking ML models, crafting adversarial examples, performing model extraction, prompt injection, membership inference, training data poisoning, fine-tuning manipulation, neural network analysis, LoRA adapter exploitation, LLM jailbreaking, or solving AI-related puzzles.

Generates llms.txt and llms-full.txt files for LLM-friendly project documentation following the llms.txt specification. Use when the user wants to create LLM-readable summaries, llms.txt files, or make their wiki accessible to language models.

Golang benchmarking, profiling, and performance measurement. Use when writing, running, or comparing Go benchmarks, profiling hot paths with pprof, interpreting CPU/memory/trace profiles, analyzing results with benchstat, setting up CI benchmark regression detection, or investigating production performance with Prometheus runtime metrics. Also use when the developer needs deep analysis on a specific performance indicator - this skill provides the measurement methodology, while `samber/cc-skills-golang@golang-performance` provides the optimization patterns.

Use when the user wants to create a dataset, generate synthetic data, or build a data generation pipeline.

> Use this skill when the operator is authoring, building, loading, or debugging a custom doca-bench plug-in — a versioned shared library with DOCA_EXPERIMENTAL-marked C entry points that doca-bench loads to measure a workload class its built-in modes do not cover, with doca_bench_cuda as the shipped reference exemplar. Trigger even when the user does not say "doca-bench-extension" or "doca_bench_cuda" — typical implicit phrasings include "no built-in doca-bench mode fits my workload", "how do I benchmark a CUDA GPUNetIO RX/TX kernel", "doca-bench cannot find or load my custom .so", "extension exported symbols do not match what the parent expects", "soversion mismatch after a DOCA upgrade", or "my GPU kernel hangs because stop_flag was never set". Refuse and route elsewhere for questions about which built-in doca-bench mode to pick, DOCA GPUNetIO programming semantics, CUDA toolkit installation, or contributor work on in-tree extensions — those belong to other skills.

> Run `doca_bench` (DOCA 2.7.0 or newer) to measure throughput, bulk latency, precision latency, or maximum bandwidth for RDMA, Compress, AES-GCM, SHA, DMA, EC, Ethernet, Comch, or GPUNetIO on a host or BlueField Arm. Use it to discover enabled benchmark libraries, capture a reproducible command/version/device/environment baseline, compare stable runs against a declared tolerance, or diagnose configuration, device-binding, workload-precondition, and measurement failures. Trigger for requests such as measuring BlueField compression speed, NIC RDMA throughput, crypto latency, or a pre-upgrade baseline. Do not use for application end-to-end timing, custom benchmark code, DOCA installation, or binary patches.

> Use this skill when the user is measuring the host or DPU-CPU control-plane rate of a DOCA Flow pipeline with doca_flow_perf — picking a JSON policy from configs/, choosing the DPDK or DOCA backend, running the single-iteration smoke then the iterative eval loop, interpreting per-iteration CPU cycles and num_pushed / num_failed, or capturing the four-tuple (DOCA version, BlueField/firmware, JSON policy, worker/queue/burst config) that makes a Kops/sec number defensible. Trigger even when the user does not explicitly mention "doca-flow-perf" — typical implicit phrasings include "how many rules per second can my BlueField insert", "5-tuple hairpin rule rate", "Kops/sec for steering", "flow-perf number does not match release notes", "DPDK vs DOCA benchmark", or "rule-install variance too high". Refuse and route elsewhere for optimizing a live Flow app (doca-flow-tune), the DPA-offloaded path (doca-flow-dpa-perf), dataplane throughput or latency, or library-internal pipe semantics — those belong to other skills.

> Use this skill when the user is building, running, or interpreting the doca/tools/gpunetio_ib_write_bw client+server benchmark — a CUDA kernel on the client posts RDMA WRITE work requests through the doca-gpunetio device-side surface to measure sustained GPU-driven WRITE bandwidth on a GPU+IB-device pair. Trigger even when the user does not explicitly mention "doca-gpunetio-ib-write-bw" or "GPUNetIO" — typical implicit phrasings include "measure WRITE BW when the GPU posts the WRs", "BW swings between runs on the same flags", "is the NIC saturated or am I CPU-bound on the CUDA kernel", "meson compile fails for the GPUNetIO bw tool", "nvidia_peermem isn't picking up my GPU buffer", or "GPU-initiated WRITE throughput vs CPU-initiated perftest". Refuse and route elsewhere for general doca-gpunetio library work, DOCA install, the GPU-initiated WRITE latency analog, the CPU-initiated upstream perftest, or application-level end-to-end throughput — those belong to other skills.