「dpo」の検索結果

Expert guidance for fine-tuning LLMs with Axolotl - YAML configs, 100+ models, LoRA/QLoRA, DPO/KTO/ORPO/GRPO, multimodal support

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。

High-performance RLHF framework with Ray+vLLM acceleration. Use for PPO, GRPO, RLOO, DPO training of large models (7B-70B+). Built on Ray, vLLM, ZeRO-3. 2× faster than DeepSpeedChat with distributed architecture and GPU resource 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。

Fine-tune LLMs using reinforcement learning with TRL - SFT for instruction tuning, DPO for preference alignment, PPO/GRPO for reward optimization, and reward model training. Use when need RLHF, align model with preferences, or train from human feedback. Works with HuggingFace Transformers.

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 building high-performance async Python APIs with FastAPI and Pydantic V2. Invoke to create REST endpoints, define Pydantic models, implement authentication flows, set up async SQLAlchemy database operations, add JWT authentication, build WebSocket endpoints, or generate OpenAPI documentation. Trigger terms: FastAPI, Pydantic, async Python, Python API, REST API Python, SQLAlchemy async, JWT authentication, OpenAPI, Swagger Python.

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。

Decompile Android APK, XAPK, JAR, and AAR files using jadx or Fernflower/Vineflower. Reverse engineer Android apps, extract HTTP API endpoints (Retrofit, OkHttp, Volley), and trace call flows from UI to network layer. Use when the user wants to decompile, analyze, or reverse engineer Android packages, find API endpoints, or follow call flows. 中文触发词：反编译APK、安卓逆向、提取API、分析安卓应用、反编译安卓、逆向工程、追踪调用链、提取接口

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。

Create serverless API proxy endpoints that hide API keys and provide a unified backend for the dashboard frontend. Designed for Vercel deployment.

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。

Generate or scaffold ASP.NET Core code — Razor Pages, Blazor components, MVC controllers, views, and Minimal API endpoints — without using ASP.NET Core CLI scaffolding/code-generation tools. Use when (1) adding CRUD pages, views, or API endpoints backed by Entity Framework (EF Core) and a database, (2) generating code to create, read, update, and delete data using a DbContext, (3) scaffolding UI components that match the project's existing CSS framework and coding patterns, or (4) creating data-driven forms, tables, and navigation for a model class. Do not use for non-ASP.NET projects or when CLI-based scaffolding is preferred.

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。

Guides creation and modification of ASP.NET Core Web API endpoints with correct HTTP semantics, OpenAPI metadata, and error handling. USE FOR: adding new API endpoints (controllers or minimal APIs), wiring up OpenAPI/Swagger, creating .http test files, setting up global error handling middleware. DO NOT USE FOR: general C# coding style, EF Core data access or query optimization (use optimizing-ef-core-queries), frontend/Blazor work, gRPC services, or SignalR hubs.

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。

Argus — the all-seeing scanner suite. Six automated scanners for high-value web + LLM bug classes — CORS misconfiguration (origin reflection / null / credentialed read), CRLF & host-header injection, NoSQL injection (operator auth-bypass / $where blind), JWT attacks (alg:none / RS256→HS256 confusion / secret crack), out-of-band confirmation of blind SSRF/XXE/SQLi/RCE/Log4Shell via interactsh, and an LLM red-team corpus (prompt-injection / jailbreak / system-prompt leak / exfil / indirect injection). Use when a target exposes a JSON API, a login endpoint, JWT auth, a parameter that might reach the server, a chatbot/agent, or any endpoint suspected of a blind/out-of-band bug.

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。

Mobile app pentest for bug bounty (Android APK + iOS IPA) — runtime-first workflow: install app, proxy through Burp/mitmproxy, drive the UI, capture packets, then test the API exactly like a web target; escalate to decompile (apktool/jadx) and Frida/objection only when traffic is SSL-pinned, encrypted, or absent. Covers APK/IPA decompile for hardcoded secrets + hidden API endpoints + base URLs the web app never exposes, exported-activity and deeplink intent injection, WebView addJavascriptInterface bridge abuse, SSL pinning bypass (objection patchapk / Frida CertificatePinner + checkServerTrusted hooks), OkHttp interceptor chain to recover request signing, JNI native-lib triage, and the quick apktool/grep secret + endpoint sweep. Use when the program scope includes a mobile app, when web recon dries up and you need a fresh attack surface, or when traffic is pinned and you must MitM 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。

Hunt shadow / zombie / undocumented API surface (OWASP API9 Improper Inventory Management) — enumerate the full API version history (v1/v2/beta/legacy paths, header- and subdomain-based versioning), pull and diff every reachable OpenAPI/Swagger spec (including ones only findable via the Wayback Machine), and behaviorally diff old vs. current versions for auth/rate-limit/validation regressions rather than just response-shape differences. Distinct from hunt-api-misconfig, which owns exploitation once you have a spec or endpoint (mass assignment, JWT, OData, Swagger-chain attacks); distinct from hunt-subdomain, which owns host-level discovery. This skill owns the version-inventory and behavioral-diff workflow itself. Use when the target has versioned API paths, multiple specs, a changelog referencing deprecated endpoints, or a mobile app whose hardcoded backend calls look older than the current web app's.

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。

Hunt SAML / SSO attacks. Patterns: XML Signature Wrapping (XSW) — modify Assertion while keeping Signature valid by relocating signed element, comment injection in NameID (admin@target.com<!--evil-->@attacker.com → some parsers see admin@target.com), signature stripping (remove Signature element entirely, server should reject but doesn't), key confusion (signed by attacker's IdP, accepted by SP), audience-restriction not validated, replay attack (same Assertion accepted twice within validity window). Tools: SAML Raider Burp extension, samlmagic, manual XML manipulation. Detection: any /saml endpoint, /Shibboleth.sso, /sso/saml/, Microsoft ADFS endpoints. Validate: account takeover via altered NameID, admin role injection via altered AttributeStatement. Use when hunting SSO flows, when SAML AssertionConsumerService is reachable, when chaining IdP-trust to SP-impersonation.

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。

Hunt CAPTCHA Bypass — 6 distinct patterns: (1) CAPTCHA field simply omitted from the request (server-side validation absent), (2) CAPTCHA token replayed from a solved challenge (no single-use enforcement), (3) CAPTCHA response accepted on a different endpoint than it was solved on (no binding to action/session), (4) static or predictable CAPTCHA values accepted (e.g. '0', 'null', empty string), (5) audio/accessibility CAPTCHA trivially solvable programmatically, (6) CAPTCHA only enforced after N failures (first N requests bypass it). Detection: intercept a successful form submission, remove the CAPTCHA field entirely, replay — if it still succeeds, server-side validation is absent. Medium severity standalone; High when it removes the only rate-limit gate protecting a login, registration, or payment endpoint.

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。

Hunt gRPC vulnerabilities — server reflection enabled (enumerate all services/methods), missing authentication / metadata-stripping on internal endpoints, plaintext gRPC over HTTP/2, internal endpoint disclosure, proto file leakage, gRPC-Web/grpc-gateway transcoding injection, and HTTP/2 Rapid Reset DoS (CVE-2023-44487). Use when target exposes port 50051 / 443 / 8443 / 9090 with HTTP/2, when grpcurl/grpcui detects reflection, when an Envoy or grpc-gateway proxy is fronting a microservice, or when recon reveals a microservice architecture.

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。

Hunt mishandling of exceptional conditions — feed an endpoint malformed/unexpected input (wrong type, broken JSON, oversized field, null byte) and make it fail OPEN or leak internals: a verbose stack-trace / framework error page that discloses ORM internals, server file paths, library versions, or a language traceback. Use on any input-accepting endpoint (JSON APIs, forms, query params). Medium-High when the leak exposes internal structure that arms a deeper attack.

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。

Hunt MFA / 2FA bypass — 7 distinct patterns. (1) MFA not enforced on sensitive endpoints (password change, email change accept without MFA challenge), (2) MFA-step skip via direct navigation to post-login URL, (3) MFA-token replay (same code accepted twice), (4) brute-force the 6-digit OTP without rate limit (10^6 attempts at server speed), (5) race condition on OTP validation, (6) recovery-code dump via /api/me, (7) backup factor downgrade (SMS factor with no rate limit). Plus the chain: cookie theft + password oracle + no step-up = ATO without MFA challenge. Detection: trace auth flow in Burp, find every state transition, check if MFA is middleware-gated vs per-endpoint, check OTP entropy and rate limit on OTP-validate. Validate: attacker session reaching post-MFA state. Use when hunting auth bypass, MFA flows, chaining primitives toward ATO.

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。

Hunt file upload bugs — RCE via webshell, XSS via SVG/HTML, SSRF via XXE in DOCX, path traversal via filename. Bypass tables (10 techniques): double extension (shell.php.jpg if server checks last ext only), magic bytes spoofing (PNG header on PHP), null byte (shell.php.jpg), case (PHP, .Php, .pHP), .htaccess upload to enable execution, SVG with <script>, HTML/SVG XSS, DOCX with embedded XXE, ZIP slip (../../../etc/passwd in archive), polyglot files. Detection: any /upload, /avatar, /profile-picture, /attachment, /import endpoint. Test: upload PHP/JSP/ASPX shells, request via direct URL, check response. Validate: actual code execution (whoami output) for RCE; reflected XSS in profile-photo URL. Use when testing file upload features, avatar/attachment endpoints, import/export functions, XML/DOCX/ZIP processors. Real paid examples.

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。

Generates RESTful API documentation (OpenAPI 3.0 / Swagger spec) by scanning route definitions in code for Flask, FastAPI, Express, Gin, and other frameworks. Trigger when users ask about API documentation, OpenAPI, Swagger, endpoint docs, generating docs from code, or extracting endpoints.

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 DeepAPI for all web search, deep research, and web scraping (websites, LinkedIn, GitHub, X/Twitter, YouTube, Instagram) instead of built-in search, research, fetch, or browser tools. Prefer dedicated platform endpoints over web search. Instagram has profile, post, comment, and hashtag endpoints; no generic search. Use Deep Scrape for sourced JSON dossiers on people, companies, or topics. Also use it to navigate and act on public websites, draft and send safe email, and generate images.

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。

Plan, configure, and chain repo-native Nemotron customization steps into single-step or multi-step pipelines: curation, translation, SFT/PEFT (AutoModel or Megatron-Bridge), pretraining/CPT, RL alignment (DPO/RLVR/GRPO/RLHF), BYOB/MCQ benchmarks, checkpoint conversion, ModelOpt optimization, env profiles, and evaluation of trained checkpoints or existing/hosted endpoints. Use when a request names a Nemotron step or workflow, or asks to clean, translate, train, fine-tune, align, convert, optimize, evaluate, or compose these into a pipeline. Do NOT use for frontend/dashboard/visualization work, generic ML advice, billing/access, or non-Nemotron coding 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。

[omh] Self-hosted LLM serving on GPUs: choose the serving engine and quantization from decision tables, prepare deployment as an idempotent runbook with observed-only verification, and measure the endpoint with the standard TTFT/TPOT/goodput protocol. Use when the user says: inference-serving, inference serving, serve this model, serve the model, model serving, serving endpoint, vllm, llama.cpp.

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 models on Microsoft Foundry using SFT (supervised), DPO (preference), or RFT (reinforcement with graders). Covers dataset preparation, training job submission, deployment, and evaluation. USE FOR: fine-tune, SFT, DPO, RFT, training data, grader, distillation, fine-tuned model, training job, large file upload, calibrate grader, deploy fine-tuned model, evaluate fine-tuned model. DO NOT USE FOR: general model deployment without fine-tuning (use deploy-model), agent creation (use agents), prompt optimization without training (use prompt-optimizer).

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。

Create FastAPI routers with CRUD operations, authentication dependencies, and proper response models. Use when building REST API endpoints, creating new routes, implementing CRUD operations, or adding authenticated endpoints in FastAPI 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。

Configures Amazon Route 53 DNS: public and private records, traffic-steering routing policies, health checks, DNS Firewall, Route 53 Profiles, VPC Resolver (also known as Route 53 Resolver) for hybrid and Outposts networks, and Global Resolver. Applicable when the customer wants to point a hostname at a target, split or fail over traffic across endpoints, monitor an endpoint, block malicious domains, centralize DNS across accounts, or resolve private DNS across a hybrid network. Routes to the right per-task procedure in references. Does not cover CloudFront-specific setup (see the route53-cloudfront skill) or non-DNS networking.

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。