195 skills · sorted by relevance
Adapt large language models to specific tasks through fine-tuning, or additional training. Prepares and validates datasets, configures LoRA or QLoRA to train a small subset of parameters, and supplies training settings, evaluation scripts, and deployment optimization guidance.
Submit groups of AI requests as batch jobs with the OmniRoute command line. Manage input files, monitor or cancel jobs, save results and errors, and incorporate batch processing into automated build and deployment workflows.
Recover context from local coding-agent transcripts and produce a checkpoint covering the objective, completed work, blockers, and next step. Explicit import requests support read-only recovery, with continuation governed by the current workspace rules and the user's request.
JSONL比例切分工具。读取JSONL数据集，按给定比例拆成多个子集，支持随机切分和按字段分层切分。 当用户明确提到按比例拆分JSONL、训练集/验证集/测试集切分、分层保持类别分布等需求时使用此skill。 即使用户没有明确说出"切分"，只要任务是把一份JSONL按比例分成多个子集，就应该使用此skill。 不负责按固定条数均匀分块，不负责按份数切块，也不负责处理非JSONL输入。
The English summary is being prepared. Showing the original description.
Prepare datasets for language model training by formatting examples, checking conversation templates, and packing examples into training sequences. Document data sources, synthetic data ratios, and validation results in a dataset card before starting a run.
Investigate difficult bugs by adding temporary logs to the actual execution path and reproducing the reported behavior. Analyze JSONL logs, which store one structured event per line, then fix the cause, verify regression tests, and remove debugging artifacts.
Analyze gh-aw execution traces, which record individual processing steps, to investigate slow or failing workflows. Using local JSONL files or telemetry backends, the skill checks timing, errors, missing attributes, and trace connections to support focused improvements.
JSONL 随机采样工具。读取 JSONL 数据集，按抽样数量或抽样比例无放回抽取样本，输出保留原字段结构的采样结果。 当用户提到随机采样、按数量抽样、按比例抽样、JSONL 抽样筛查等需求时使用此skill。 即使用户没有明确说出"random_sampler"，只要任务涉及对 JSONL 记录做无放回随机抽取，就应该使用此skill。 不负责分层采样、领域均衡采样、表格采样或其他格式转换。
The English summary is being prepared. Showing the original description.
Dry-run one cycle of a sales agent workflow without calling language models. Validate the template and agent types, check budget reservations, generate agent prompts, and record a structured status summary for downstream automation.
Turn local Q&A session files into an English weekly report with topic rankings, unresolved issues, activity metrics, and comparisons with the previous week. Include supporting message excerpts and structured data for later document and dashboard rendering.
Reproduce Mastra Code bugs with a real model and compare timestamped logs across network, streaming, execution, and terminal interface layers. Use the evidence to locate timing problems, incorrect token rates, flicker, or stuck status displays.
Prepare to fine-tune open models or Gemini models on Agent Platform. Check the environment, convert and validate training data, choose supported models and settings, and estimate tuning costs for open models.
Preserve the context of long Claude Code sessions across compaction, which shortens conversation history. Create a restore map linking decisions, files, and next steps, then use ranked readings and targeted transcript lookups to rebuild working context.
onboarding workflow for creating datasets and applications in Viking AI Search. Supports one-time import from local files (JSON, JSONL, CSV) and MySQL databases, plus scheduled incremental sync for append-only JSONL files and MySQL. All sources are first exported to a bootstrap JSONL file; backend-driven schema inference handles detection, and optional background sync keeps the dataset up to date as new data arrives.
The English summary is being prepared. Showing the original description.
Analyze pi (coding-agent) session JSONL and ego-benchmark-harness run artifacts to explain slow or failed individual tasks, compare execution-path differences between agents or browser tools, and count tool calls and error types. Use this skill whenever users ask to analyze a pi session, explain why a task was slow or failed, compare two runs, investigate an error, locate the session corresponding to tasks.jsonl, inspect an agent execution path, count tool calls, analyze a run ID, or perform any forensic or retrospective analysis of JSONL under ~/.pi/agent/sessions/ or agent-home/.pi/agent/sessions/.
The English summary is being prepared. Showing the original description.
Create and run automated tests that score LLM answers against expected outputs or judging criteria. Monitor progress, inspect individual results, compare models on the same tests, and use score thresholds in CI, an automated checking workflow.
Capture token usage from Claude Code session logs, calculate USD costs using the supplied pricing reference, and save structured records to AgentDB. Review totals by model and tier, and provide data for cost reports.
Delegate a clearly scoped repository analysis or implementation task to Pi without an interactive session. The skill limits the working directory and available tools, then checks structured output for errors before using the result.
Delegate repository analysis and coding tasks to GitHub Copilot CLI without an interactive session. The skill reads structured output, checks errors, and summarizes the final result along with any denied tool requests.
Delegate a clearly scoped repository task to Kimi Code without opening its interactive interface. The skill reads structured execution events, extracts the final response, and checks for errors, with separate handling for read-only work and authorized changes.
Delegate code analysis and implementation to Codex CLI without an interactive session. The skill reads ordered output events, extracts the final agent response, and reports errors, with a read-only sandbox used by default.
Search the agent’s conversation history by topic or date to retrieve relevant past entries. Inspect neighboring records when needed to understand the surrounding discussion, using the history log specified for the agent’s workspace.
Turn documents, notes, research, or code into a knowledge graph that maps identifiable things and their relationships. Create or update entity files and typed connections, with source evidence attached for inspection and later visualization.
Guide reinforcement learning, training models through rewards, with slime. Prepare data, configure Megatron-LM training and SGLang response generation, and set up asynchronous runs or custom multi-turn workflows for supported language models.