数据与 AI技能

Implement and debug large-model training with PyTorch FSDP, which splits training state across devices. Consult documentation and code patterns for sharding choices, mixed precision, and checkpoints when building or reviewing distributed training code.

中文简介正在准备中。当前显示英语简介。

• Implement distributed training for large models
• Compare FSDP wrapping and sharding patterns
• Review mixed-precision and checkpoint examples

Explore concepts learned by language models using sparse autoencoders, which separate internal activity into a small set of features. Analyze pretrained models, train custom autoencoders, assess reconstruction quality, and experiment with feature based output adjustments.

中文简介正在准备中。当前显示英语简介。

• Exploring features with a pretrained SAE
• Training an SAE for a chosen model layer
• Comparing feature responses across input texts

Integrate Flash Attention into Transformer models to speed up attention, the calculation that relates input elements, while reducing GPU memory use. Covers PyTorch integration, performance measurement, output comparison, and a separate H100 optimization path.

中文简介正在准备中。当前显示英语简介。

• Optimize training with long input sequences
• Reduce memory use during long-context inference
• Measure speed and compare model outputs

Control local language model output with patterns and grammar rules to generate structured JSON, formatted dates, and fixed classification labels. The skill also covers reusable Python functions and workflows that combine multiple generation steps.

中文简介正在准备中。当前显示英语简介。

• Generate JSON that follows a defined schema
• Keep dates and email addresses in a specified format
• Classify text using a fixed set of labels

Build search and retrieval systems with Pinecone, a managed database for numerical representations of content. Covers data uploads, filtered similarity searches, combined semantic and keyword search, and connections to LangChain and LlamaIndex.

中文简介正在准备中。当前显示英语简介。

• Building retrieval for AI answers grounded in documents
• Combining semantic and keyword search
• Filtering results by category or price

Helps organize PyTorch model training with LightningModule and Trainer, including validation, logging, model saving, and early stopping. It also guides distributed training across multiple GPUs and settings for reduced-precision computation.

中文简介正在准备中。当前显示英语简介。

• Converting existing PyTorch training code to Lightning
• Adding validation metrics and model checkpoints
• Setting up training across multiple GPUs

Prepare LLM training datasets with NeMo Curator by removing duplicates, filtering low-quality content, and redacting personal information. The skill also covers image, video, and audio curation, including GPU processing and distributed workflows.

中文简介正在准备中。当前显示英语简介。

• Clean web-scraped text for LLM training
• Remove exact, near, and meaning-based duplicates
• Remove or replace personal information in text

Helps you build similarity search with FAISS to find nearby vectors, numerical representations of data. Covers index selection, saving and loading, GPU use, and tuning search speed, accuracy, and memory consumption for large datasets.

中文简介正在准备中。当前显示英语简介。

• Building similarity search over large vector datasets
• Choosing indexes for speed, accuracy, and memory use
• Running batch searches on GPUs

Use LLaVA, a model that understands images and text, to answer questions about pictures, write image descriptions, and build image chat applications. Includes guidance for setup, ongoing conversations, and model memory options.

中文简介正在准备中。当前显示英语简介。

• Writing descriptions for photographs
• Asking questions about objects and scenes
• Building a chatbot that discusses images

Set up reinforcement learning, training guided by response rewards, for large language models. Use Megatron-LM and SGLang to prepare data, configure training and response generation, and build asynchronous or multi-turn experiments.

中文简介正在准备中。当前显示英语简介。

• Setting up GRPO training for a language model
• Overlapping response generation and training
• Training agents that use tools across multiple turns

Build similarity search with Qdrant, a database for numerical representations of content. Store and retrieve vectors, filter results by metadata, connect retrieval to AI answers, and configure memory usage and deployment for larger workloads.

中文简介正在准备中。当前显示英语简介。

• Retrieving reference documents for AI answers
• Filtering similarity search by category or date
• Building real-time recommendations

Analyze and modify refusal behavior in open-weight language models without retraining. Use OBLITERATUS to select methods suited to available hardware, compare approaches, and evaluate changes in refusal rates and response quality.

中文简介正在准备中。当前显示英语简介。

• Studying how language models produce refusals
• Testing refusal-related weight changes without retraining
• Selecting models and methods for available hardware

Classify images using written labels without additional training data, or find images that match a text query. This skill provides CLIP workflows for image-text similarity, batch processing, and storing searchable image representations in vector databases.

中文简介正在准备中。当前显示英语简介。

• Classify photos using written category labels
• Find images by describing their content
• Compare images with candidate descriptions

Set up and troubleshoot Honcho memory for Hermes so user preferences and facts carry across conversations. Configure how memory is recalled, manage separate agent identities across profiles, and adjust context size, observation settings, and reasoning frequency.

中文简介正在准备中。当前显示英语简介。

• Setting up cloud or self-hosted memory for Hermes
• Investigating missing memories or profile synchronization issues
• Sharing user context across multiple agent profiles

Retrieve Hyperliquid market data, wallet balances, positions, and recent trading history without an API key. Review trades alongside market trends and fees, or export price candles and funding payments as JSON for later analysis.

中文简介正在准备中。当前显示英语简介。

• Check market prices, funding, and order book depth
• Inspect wallet positions and spot balances
• Review recent trades alongside fees and market trends

Inspect Solana wallet balances, token holdings, transactions, and NFTs with USD pricing. Check portfolio value, token details, recent activity, large SOL transfers, and network statistics through eight commands that require no API key.

中文简介正在准备中。当前显示英语简介。

• Check a wallet's holdings and total value in USD
• Inspect transaction fees and balance changes
• Look up token prices and major holders

Extract information from web pages through Aside using existing signed-in sessions. Collect lists, tables, prices, or links and receive one JSON document, a structured data format, without changing the page.

中文简介正在准备中。当前显示英语简介。

• Extracting product names and prices
• Collecting tables and links as JSON
• Reading information from signed-in pages

Compare Claude, GPT, and Gemini on the same skill or prompt. Get a table of response times, token usage, and costs, with optional AI scoring to help choose a model for the task.

中文简介正在准备中。当前显示英语简介。

• Choosing a model for a specific skill
• Comparing response times and costs
• Comparing output quality with optional AI scoring

Turn a Caveman token-usage report into specific edits to instruction files, repeated context, and memory. Review each proposed change before approval, then measure the result and distinguish confirmed file reductions from less certain session-level effects.

中文简介正在准备中。当前显示英语简介。

• Trimming a large CLAUDE.md file
• Moving repeatedly pasted context into cavemem
• Reviewing duplicate rules and broken memory links

Review Caveman Cloud spending, latency, errors, and potential improvements using current reports and traces, or individual execution records. Produce findings tied to time windows and trace IDs, while keeping measured costs, estimated opportunities, and verified savings separate.

中文简介正在准备中。当前显示英语简介。

• Understanding which workflows drive LLM spending
• Investigating slow responses and errors
• Reviewing measured costs and verified savings separately

View recorded output tokens, cache-read tokens, response counts, and available mode attribution for a Claude Code session. In other agent environments, find the native usage report where supported, without treating observed usage as proof of savings.

中文简介正在准备中。当前显示英语简介。

• Checking recorded token usage for the current session
• Reviewing response counts and mode attribution
• Aggregating recorded output across sessions

Organize the information an AI coding agent needs for each task. This skill guides project rules, relevant file selection, conversation summaries, and session handoffs so the agent can follow existing conventions and retain key decisions.

中文简介正在准备中。当前显示英语简介。

• Set up project rules for AI coding agents
• Select relevant files before starting a task
• Trim stale information from a long conversation

Create HTML dashboards and reports from local data or connected services, with data that can be refreshed without redesigning the page. Record sources, timestamps, and transformations so readers can trace where displayed information came from.

中文简介正在准备中。当前显示英语简介。

• Build a dashboard whose data can be refreshed
• Create a report with traceable sources
• Turn connected service data into a status page

Find, compare, and run AI models through Replicate’s API, a service interface for calling models. Use it to explore image, audio, and video generation options and support workflows that frequently switch between models.

中文简介正在准备中。当前显示英语简介。

• Find models for image generation
• Compare audio generation models
• Try video generation models