「model debugging」の検索結果

Field-tested methodology and concrete recipes for training and operating large-scale LLM/VLM/multi-modal models end to end - choosing and benchmarking accelerators, storage and network; SLURM/Kubernetes orchestration; maximizing training throughput and fitting models in memory; diagnosing and surviving training instabilities, NaN/Inf, and hardware/job failures; checkpointing and fault tolerance; inference performance and memory; debugging multi-node/ multi-GPU hangs; and writing/running tests. Use when the user is training or fine-tuning large models, hits low TFLOPS/MFU, OOM, slow dataloading, a loss spike/divergence, a NCCL/InfiniBand or multi-node hang, node/GPU failures, checkpoint or preemption problems, storage/network bottlenecks, or needs to pick GPUs/cloud/file-systems or size inference latency/throughput. Distilled from "Machine Learning Engineering", the latest version of which can be found at https://github.com/stas00/ml-engineering The latest SKILL.md version can be found at https://github.com/stas00/ml-engineering/blob/master/skills/ml-engineering/SKILL.md

日本語の概要は準備中です。原文の説明を表示しています。

Visualize training metrics, debug models with histograms, compare experiments, visualize model graphs, and profile performance with TensorBoard - Google's ML visualization toolkit

日本語の概要は準備中です。原文の説明を表示しています。

Use when designing, reviewing, or debugging how an agent's context window gets filled, pruned, or shared — choosing what loads at boot versus on demand, sizing an install or an always-loaded file, fixing an agent that drifts, repeats itself, or forgets constraints mid-task, planning compaction or summarization, deciding single-agent versus subagents, engineering handoffs between agents, or picking a tool loadout. NOT for rewording a prompt's tone, choosing which model to pin, or debugging business logic — those are adjacent moments this skill does not serve. Historical 2024–2025 snapshot; not normative for present-day frontier models — see the Status section.

日本語の概要は準備中です。原文の説明を表示しています。

Visualize training metrics, debug models with histograms, compare experiments, visualize model graphs, and profile performance with TensorBoard - Google's ML visualization toolkit

日本語の概要は準備中です。原文の説明を表示しています。

AC branch pi-model power flow equations (P/Q and |S|) with transformer tap ratio and phase shift, matching `acopf-math-model.md` and MATPOWER branch fields. Use when computing branch flows in either direction, aggregating bus injections for nodal balance, checking MVA (rateA) limits, computing branch loading %, or debugging sign/units issues in AC power flow.

日本語の概要は準備中です。原文の説明を表示しています。

Answer a hard question by fanning it out to a PANEL of models running in parallel — each answering independently with web search and bash, none seeing the others' work — then having Opus 4.8 judge every response into a structured analysis (consensus, contradictions, partial coverage, unique insights, blind spots) and write a final answer grounded in it. The panel is two independent Opus 4.8 runs (slug opus4.8-4.8), Opus 4.8 + GPT-5.5 via codex (opus4.8-gpt5.5), Opus 4.8 + Gemini 3.1 Pro via agy (opus4.8-gemini3.1pro), or all three (opus4.8-gpt5.5-gemini3.1pro). Opus always judges and writes the final answer — the pipeline can't be reversed. Runs on local CLI subscriptions (no metered API), and saves a timestamped provenance .md per run. Use this whenever the user asks to "run it through Fusion", says /fusion, wants a multi-model / panel / ensemble answer, wants a question cross-checked across models, or wants a higher-confidence answer with consensus and blind spots surfaced — even if they don't say "fusion". General-purpose: any topic (research, law, strategy, technical, personal). Best for high-stakes research, design calls, and debugging where being confidently wrong is expensive.

日本語の概要は準備中です。原文の説明を表示しています。

Tool skill — the *first move* in any LM-debugging session: dump the actual rendered prompt the framework sent to the model, before changing anything else. Activate when an LM call produced an unexpected output (wrong answer, schema violation, refusal, truncation, cost spike, latency spike, infinite loop, "model got dumber after upgrade"). The skill enforces a 30-second inspect step BEFORE any prompt edit, model swap, retry, or temperature tweak. Cross-framework cheat sheet: DSPy `inspect_history`, LangGraph `get_state_history`, CrewAI `step_callback`, LangChain `set_debug`/`set_verbose`, Aider `/diff`+`--verbose`, raw OpenAI/Anthropic via `OPENAI_LOG=debug`/`ANTHROPIC_LOG=debug` or HTTPX event hooks. Do NOT activate for first-time prompt authoring, exploratory prompt design, or non-LM bugs.

日本語の概要は準備中です。原文の説明を表示しています。

Reference for the Claude API / Anthropic SDK — model ids, pricing, params, streaming, tool use, MCP, agents, caching, token counting, model migration. TRIGGER — read BEFORE opening the target file; don't skip because it "looks like a one-liner" — whenever: the prompt names Claude/Anthropic in any form (Claude, Anthropic, Fable, Opus, Sonnet, Haiku, `anthropic`, `@anthropic-ai`, `claude-*`, `us.anthropic.*`, `[1m]`); the user asks about an LLM (pricing/model choice/limits/caching) — never answer from memory; OR the task is LLM-shaped with provider unstated (agent/MCP/tool-definition/multi-agent/RAG/LLM-judge/computer-use; generate/summarize/extract/classify/rewrite/converse over NL; debugging refusals/cutoffs/streaming/tool-calls/tokens). SKIP only when another provider is being worked on (overrides all triggers): OpenAI/GPT/Gemini/Llama/Mistral/Cohere/Ollama named in the query; OR `grep -rE 'openai|langchain_openai|google.generativeai|genai|mistralai|cohere|ollama'` over the project hits (run this grep FIRST if no provider named — don't Read the file).

日本語の概要は準備中です。原文の説明を表示しています。

Audit a .glb or .gltf and make it small, correct, and fast to load in a browser, with a measured before/after report. Use when a model is "too big" or "slow to load", when the user asks to "optimize this GLB", "compress this glTF", or "export for web", or when they mention Draco, meshopt, gltfpack, KTX2, Basis, texture VRAM, draw calls, or a Blender / CAD / photogrammetry / Pascal export headed for Three.js, React Three Fiber (drei useGLTF), Babylon.js, or model-viewer. Covers inspection, spec validation, geometry and texture compression, scene-graph cleanup, and the correctness checks (metres, Y-up, node names, animations, PBR fidelity, alpha modes, vertex colors) that must not regress. Not for editing or modelling geometry, authoring or laying out a scene, generating 3D from text or images, converting from FBX/OBJ/CAD, or debugging a runtime frame rate that has nothing to do with asset size.

日本語の概要は準備中です。原文の説明を表示しています。

Time-travel debugging for DAG executions. Inspect agent state at any node, replay decisions with modified inputs, compare execution traces side-by-side, and identify where reasoning diverged. Inspired by LangGraph Studio's state-editing model and Temporal's event history. Activate on "debug DAG", "replay execution", "time travel debug", "inspect node state", "what went wrong at step", "compare runs", "execution diff". NOT for live monitoring (use dag-runtime + websocket-streaming), failure analysis (use dag-ops), or general code debugging.

日本語の概要は準備中です。原文の説明を表示しています。

Visualize métricas de treinamento, depure modelos com histogramas, compare experimentos, visualize grafos de modelos e perfil de desempenho com TensorBoard - kit de visualização de ML do Google

日本語の概要は準備中です。原文の説明を表示しています。

Operational guide for the NV oOS Model Foundry (Proposal 067) — the cluster fine-tuning flywheel. Phase 1 shipped: the Corpus Foundry (Pro addon, addons/pro/includes/model-foundry/) — consent-gated, deduplicated, PII-scrubbed corpus exporters (trajectory_v1, preference_v1, docs_v1) with deterministic holdout splits and R12 provenance manifests. Covers the governance contract, row schemas, output layout, filter seams, the registration pattern for foundry tool slices, how to add a new exporter, and the test/validation gates. Use when extending the Corpus Foundry, debugging corpus exports, wiring a new data source, or continuing Phases 2–6 (train-worker, registry/serving, eval gate, fleet sharing).

日本語の概要は準備中です。原文の説明を表示しています。

Reference for the Claude API / Anthropic SDK — model ids, pricing, params, streaming, tool use, MCP, agents, caching, token counting, model migration. TRIGGER — read BEFORE opening the target file; don't skip because it "looks like a one-liner" — whenever: the prompt names Claude/Anthropic in any form (Claude, Anthropic, Fable, Opus, Sonnet, Haiku, `anthropic`, `@anthropic-ai`, `claude-*`, `us.anthropic.*`, `[1m]`); the user asks about an LLM (pricing/model choice/limits/caching) — never answer from memory; OR the task is LLM-shaped with provider unstated (agent/MCP/tool-definition/multi-agent/RAG/LLM-judge/computer-use; generate/summarize/extract/classify/rewrite/converse over NL; debugging refusals/cutoffs/streaming/tool-calls/tokens). SKIP only when another provider is being worked on (overrides all triggers): OpenAI/GPT/Gemini/Llama/Mistral/Cohere/Ollama named in the query; OR `grep -rE 'openai|langchain_openai|google.generativeai|genai|mistralai|cohere|ollama'` over the project hits (run this grep FIRST if no provider named — don't Read the file).

日本語の概要は準備中です。原文の説明を表示しています。

Visualize training metrics, debug models with histograms, compare experiments, visualize model graphs, and profile performance with TensorBoard - Google's ML visualization toolkit

日本語の概要は準備中です。原文の説明を表示しています。

AIアシスタントから外部の機能やデータを使うためのMCPサーバーを構築・保守します。Node/TypeScriptでの機能登録、入力検証、通信方式の選択を扱います。

• AI向けのMCPサーバーを作りたいとき
• 操作・データ・指示テンプレートの追加
• ローカルとリモートの接続方式の選択

LLM observability platform for tracing, evaluation, and monitoring. Use when debugging LLM applications, evaluating model outputs against datasets, monitoring production systems, or building systematic testing pipelines for AI applications.

日本語の概要は準備中です。原文の説明を表示しています。

Use early when debugging a medium or hard bug, especially when tests alone may not reveal the real runtime failure. Trigger this before extended TDD iteration when a bug involves runtime state, ordering, persistence, streaming, concurrency, UI/manual reproduction, external services, or when a red or newly passing test may not model the real issue. Skip only when the root cause is already directly proven by a stack trace or deterministic test that exercises the real runtime path.

日本語の概要は準備中です。原文の説明を表示しています。

LLM observability platform for tracing, evaluation, and monitoring. Use when debugging LLM applications, evaluating model outputs against datasets, monitoring production systems, or building systematic testing pipelines for AI applications.

日本語の概要は準備中です。原文の説明を表示しています。

Fine-tune and serve Physical Intelligence OpenPI models (pi0, pi0-fast, pi0.5) using JAX or PyTorch backends for robot policy inference across ALOHA, DROID, and LIBERO environments. Use when adapting pi0 models to custom datasets, converting JAX checkpoints to PyTorch, running policy inference servers, or debugging norm stats and GPU memory issues.

日本語の概要は準備中です。原文の説明を表示しています。

Perform language and framework specific security best-practice reviews and suggest improvements. Use when the user explicitly requests security best practices guidance, a security review or report, or secure-by-default coding help. Supports Python, JavaScript/TypeScript, and Go. Do NOT use for general code review, debugging, threat modeling (use security-threat-model), or non-security tasks.

日本語の概要は準備中です。原文の説明を表示しています。

Use this skill for hands-on DOCA GPI programming — wiring a GPU-Packet-Initiator context so a CUDA kernel drives RDMA queues directly from GPU memory without host CPU mediation. Covers picking GPI vs doca-gpunetio, the doca_gpi / domain / channel object model, the GPU-side handle handoff (doca_gpu_gpi_channel*), attaching GPU memory to a GPI domain, the domain and channel attribute objects, and debugging DOCA_ERROR_* from doca_gpi_* calls. Trigger even when the user does not explicitly mention "DOCA GPI" — implicit phrasings include "my CUDA kernel needs to post RDMA directly from GPU memory", "DOCA_ERROR_* from doca_gpi_gpu_channel_get", "how do I hand a GPU handle to my CUDA kernel", "how many channels can a GPI domain hold", or "GPU kernel driving RDMA without the host CPU on the path". Refuse and route elsewhere for the doca-gpunetio Send/Receive surface, the doca-rdma queue lifecycle, DPA-side initiation (doca-rdmi), or the CUDA programming model — those belong to other skills.

日本語の概要は準備中です。原文の説明を表示しています。

Design data systems by understanding storage engines, replication, partitioning, transactions, and consistency models. Use when the user mentions "database choice", "which database should I use", "SQL or NoSQL", "replication lag", "partitioning strategy", "consistency vs availability", "stream processing", "ACID transactions", "eventual consistency", "my queries are slow at scale", or "data is inconsistent across replicas". Also trigger when choosing a datastore, designing data pipelines, or debugging distributed-system consistency issues. Covers data models, batch/stream processing, and distributed consensus. For system design, see system-design. For resilience, see release-it.

日本語の概要は準備中です。原文の説明を表示しています。

POS (Point of Sale) architecture and patterns. Use when writing, editing, exploring, debugging, fixing bugs, implementing features, or reviewing WooPos-prefixed classes or files under ui/woopos/. POS uses a different architecture than the main app — plain ViewModel (not ScopedViewModel), pure Compose (no Fragments), Compose Navigation (no nav graphs), parent-child SharedFlow event bus. Loading this skill prevents applying main-app patterns that would be wrong for POS. For analytics use `pos-analytics`, for tests use `pos-tests`.

日本語の概要は準備中です。原文の説明を表示しています。

Kaggle-first end-to-end competition workflow for scored submissions. Use when Codex must run Kaggle or competitive ML workflows through scored submission, including code competitions, validation, metrics, policy-safe public notebook/discussion intel, tabular/text/image modeling, tuning, ensembling/stacking, proactive multi-notebook architectures, producer notebooks that train models and export private Kaggle artifact datasets, downstream consumer notebooks, Kaggle GPU offload, kagglehub access, hidden-test debugging, and public score retrieval.

日本語の概要は準備中です。原文の説明を表示しています。