「rl」の検索結果

Search and preview Irasutoya illustrations with the installed `irasutoya` command-line tool. Use this skill whenever the user wants to find, recommend, preview, randomly pull up, or get the page/image URL of an Irasutoya (いらすとや) image — from a topic, scene, Korean/Japanese/English phrase, visual description, emotion, character, object, or action — even if they don't say "Irasutoya" by name but clearly want a free, cute illustration for a slide, 발표자료, blog, doc, or chat. Also use it to optimize vague requests into fast Japanese keyword candidates and to return page/image URLs. Do NOT use it to generate brand-new AI art (e.g. DALL·E, Midjourney), draw charts or diagrams, design logos or icons, cut out/edit existing images, or recommend other stock-image sites — it only searches Irasutoya's existing catalog.

Framework RLHF de alta performance com aceleração Ray+vLLM. Use para treinamento PPO, GRPO, RLOO, DPO de modelos grandes (7B-70B+). Construído em Ray, vLLM, ZeRO-3. 2× mais rápido que DeepSpeedChat com arquitetura distribuída e compartilhamento de recursos GPU.

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。

[Utilities] Use when a project needs to add, change, list, or remove its OWN protocol rules layered on top of a framework skill — a named overlay of extra project rules that applies whenever the targeted skill runs, stored with the project rather than in the portable framework. `/project-skill-protocol list` prints every overlay with its target and description; `add` drafts a new overlay and confirms it before writing; `update` applies a surgical change; `delete` removes it. Overlays are ADDITIVE ONLY — they never replace, override, or waive anything the skill already does. Triggers: skill protocol, project protocol, custom protocol for skill, extend a skill, overlay rules, add rules to a skill, project rules for skills, skill protocol overlay.

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。

Set up and run the full RLHF pipeline (SFT, reward model training, RL from reward model) using the Tinker API. Use when the user wants to do RLHF, train a reward model, or run the full preference-based RL pipeline.

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。

Gets a custom map into CARLA on UE 5.8 — the four distinct routes and which one to pick — via Util/Tools/Import.py for FBX+OpenDRIVE assets into a source build, CARLA_MAPS_TO_COOK to ship a map inside a package, or client.generate_opendrive_world() for an OpenDRIVE-only map at runtime with no build at all. Covers where imported assets land, why an imported map cannot be loaded from a package by name, and the four capabilities the UE4 skill has that ue58 lacks — standalone asset packages, their installer, semantic-material tagging, and a pedestrian navmesh that is generated but silently skipped. Use when the user asks to "import a map", "add a custom map", "load my .xodr", "package my own town", or "create a standalone map package".

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。

Runs the Autoware autonomous-driving stack against CARLA on UE 5.8 over CARLA's native ROS 2 interface — no bridge process. Reports the DDS contract (rt/vehicle/status/* out, rt/control/command/* in), the Autoware-only sensor blueprints, generates the lanelet2 + point-cloud map artifacts Autoware needs instead of an .xodr, and drives the two shipped modes (classical NDT stack, or the camera-only VAD end-to-end model). Use when the user asks to "run Autoware on CARLA", "connect an AV stack", "drive the ego from Autoware", "generate a lanelet2 map", or asks which topics CARLA exchanges with Autoware.

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。

Installs the CARLA Autonomous Driving Leaderboard (leaderboard.carla.org) at a chosen version — 1.0, 2.0 or 2.1 — together with the exact scenario_runner branch and CARLA build each one requires, then wires the four-root PYTHONPATH and verifies the whole stack imports. Detects which version an existing checkout is by reading its code, not its branch name. Use when the user asks to "install the leaderboard", "set up the CARLA challenge", "evaluate my driving agent", "reproduce my leaderboard score", or names a leaderboard version.

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。

Installs CARLA's ScenarioRunner (the official scenario engine) and picks the branch that matches the CARLA in use — master for 0.9.14-0.9.16 (UE4), ue5-master for 0.10.0 (UE5), leaderboard-1.0/2.0/2.1 when driving the Leaderboard. Clones or checks out the branch, installs requirements, wires PYTHONPATH (including the `agents` package the carla wheel does not ship), and verifies the import. Use when the user asks to "install scenario runner", "set up SR", "run CARLA scenarios", "use OpenSCENARIO", or has a scenario_runner that fails to import.

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。

Cooks and packages CarlaUE4 into distributable tarballs under Dist/ — either the full simulator or a standalone asset package of selected maps and props for later import. ROS2=1 keeps the native ROS 2 interface in the cooked package (a plain cook silently drops it). Use when the user asks to "package CARLA", "make a Dist build", "cook the project", "export a map as a package", "package CARLA with ROS2", or needs a build with working camera and lidar sensors.

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。

Queries the live CARLA world and resolves an entity description into a concrete actor id — find actors by type, role, or colour, or the single one nearest a point/actor; read the world snapshot; get static level bounding boxes by label; cast rays for semantic points; project a point to the ground; or (ros-topics) list the native ROS 2 topics each actor should be publishing and why any is silent. Identifies by stable attributes, not by any rank/order. Use when the user asks "what actors are here", "find the red prius", "the vehicle nearest the ego", "list the pedestrians", or needs a specific actor id for another skill.

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。

Overlay controller for dialogs, sheets, modals, windows, and toasts opened by dispatch. Use when opening, adding, registering, or debugging an overlay, or touching features/overlays/**, app/DeferredSingletonCore.tsx, lib/redux/slices/overlaySlice.ts, dispatch(openOverlay(...)), useOpenX hooks, or XController components. NOT for the WindowPanel component itself (use window-panels).

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。

REST APIのURL命名、HTTPステータス、一覧の分割取得、エラー応答を設計・レビューし、認証や利用回数制限、バージョン管理の方針も整理するスキル。

• 新しいAPIのURLと応答を設計したいとき
• 既存のAPI仕様をレビューしたいとき
• 一覧の分割取得と絞り込み方式の選定

为 RL 训练设计奖励函数或排查奖励黑客时使用——决定奖励来自规则、人类偏好还是模型评判，结果奖励还是过程奖励，标量、向量还是生成式诊断，以及如何用路径约束与 RLVP 惩罚违规动作；也用于多轮 Agent 的信用分配、隐藏测试与早停判定设计。触发词：奖励设计、reward hacking、reward seeking、RLVR、RLHF、GRM、ORM、PRM、RLVP、信用分配、隐藏测试、过程奖励、奖励塑形、组内差异。

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。

为 Agent 模型做后训练或选路线时使用——判断该先补 Mid-training、用 SFT 立协议还是直接上 RL；构造 Mid-training 语料与 SFT 数据、搭建 RL 环境、在 GRPO/PPO/DPO 与蒸馏之间取舍；排查 pass@k 近零、格式不稳就训 RL、组内无奖励差异、训练-推理数值失配、过度训练泛化下降。触发词：后训练、Mid-training、继续预训练、SFT、RL、GRPO、PPO、DPO、拒绝采样、RFT、蒸馏、on-policy、RLVR、LoRA。

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。

URLScan.io is a free service for scanning and analyzing suspicious URLs. It captures screenshots, DOM content, HTTP transactions, JavaScript behavior, and network connections of web pages in an isolat

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。

Expert guidance for GRPO/RL fine-tuning with TRL for reasoning and task-specific model training

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。

Post-train LLMs with TRL (Transformers Reinforcement Learning) — SFT, DPO, GRPO, KTO, and reward-model training. Use when writing or debugging training code with the TRL Python API or the trl CLI.

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。

Starts a CARLA UE 5.8 server from a ue58-dev tree — the editor binary in -game mode (the working headless server), the packaged shipping build, or the full editor — with offscreen/nullrhi rendering, custom RPC and streaming ports, and native ROS 2 via the single-dash -ros2 flag. Probes a running server, and stops one cleanly despite the 15-character process-name truncation. Use when the user asks to "run/start CARLA UE5", "launch a ue58 server", "run headless", "publish ROS 2 topics from UE5", or a client cannot reach the simulator.

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。

Imports FBX meshes into CARLA on UE 5.8 as spawnable props — drives the editor headlessly to import the mesh, derives the prop size from its measured bounds, and registers it in PropParameters.json so it appears in the blueprint library as static.prop.<name>. Handles the FBX unit-scale problem, verifies the result against a running server, and can revert cleanly. Use when the user asks to "import a prop", "add a custom mesh/object to CARLA", "make my model spawnable", or a prop imports but never shows up in the blueprint library.

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。

Imports a pedestrian into CARLA on UE 5.8 as a spawnable, animating walker — imports the skinned FBX bound to CARLA's shared pedestrian skeleton, duplicates a donor walker blueprint and repoints it at the new mesh, and registers it in WalkerParameters.json as walker.pedestrian.<id>. Can also export a shipped walker to FBX, which is how you obtain a rig-conforming mesh to start from. Use when the user asks to "import a walker/pedestrian", "add a custom character", "clone a walker", or has a walker that spawns but stands still.

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。

Reads a specific actor's live telemetry from a running CARLA server — location and rotation, velocity and speed, acceleration, angular velocity, and for vehicles the control input (throttle/steer/brake), front-wheel steer angle and mass. All kinematics come from one world snapshot so they are frame-consistent. Use when the user asks "how fast is it going", "where/what heading is the car", "show the ego's telemetry", or "stream a vehicle's state". Resolve ambiguous targets with the world-data skill.

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。

Inspects and visualises a ROS-2-enabled CARLA server from Docker containers, with no local ROS 2 install — list/echo/rate-check topics (the "are the topics actually on the wire" check), run the bundled map-and-lidar demo stack (hero vehicle with camera/lidar/GNSS/IMU on autopilot), and open RViz2 with the lane-network preset. Use when the user asks to "check the ROS topics", "echo /carla/map", "is CARLA publishing to ROS", "visualise the lidar in RViz", "open rviz", or "run the ROS 2 demo".

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。