菜单
Fine-tuning pipeline for small open models using LoRA/QLoRA. Trains domain-specific adapters from conversation logs, codebase files, or custom datasets. Pre-built recipes for writing style, codebase conventions, and domain vocabulary. Exports to LiteLLM for immediate use.
用 Codex 或 Claude 帮你安装 复制这段 Prompt，粘贴到 Codex、Claude 或其他助手里，让它检查 Skill 页面并帮你完成安装。
基于 SOC 职业分类
Sends messages between agents, broadcasts to channels, and retrieves message history for inter-agent communication.
Spawns, manages lifecycle, and terminates autonomous AI agents from the 40+ built-in agent catalog.
Optimize prompts and context windows to reduce token usage while preserving quality.
Animated desktop character state machine for Sven's Tauri companion app. Manages character form (ORB, ARIA, REX, ORION), state transitions (idle→thinking→speaking→celebrating), walk cycles, thought bubbles, sound effects, and real-time agent event sync.
Multi-model deliberation system. Sends queries to multiple LLMs simultaneously, has them peer-review each other's responses anonymously, then a chairman model synthesizes the best answer. Supports configurable council composition, voting strategies, and cost tracking.
Interactive educational autograd engine — port of Karpathy's micrograd. Build, train, and visualise tiny neural networks step-by-step to learn how backpropagation works.
| name | model-trainer |
| description | Fine-tuning pipeline for small open models using LoRA/QLoRA. Trains domain-specific adapters from conversation logs, codebase files, or custom datasets. Pre-built recipes for writing style, codebase conventions, and domain vocabulary. Exports to LiteLLM for immediate use. |
| version | 1.0.0 |
| publisher | acmecorp |
| handler_language | typescript |
| handler_file | handler.ts |
| inputs_schema | {"action":{"type":"string","required":true,"enum":["create_job","get_job","list_jobs","cancel_job","list_recipes","list_exports","get_stats"]},"job_id":{"type":"string","description":"Training job ID"},"org_id":{"type":"string","description":"Organisation ID"},"user_id":{"type":"string","description":"User ID"},"recipe":{"type":"string","enum":["writing_style","codebase_conventions","domain_vocabulary","task_specific","custom"],"description":"Pre-built fine-tuning recipe domain"},"config":{"type":"object","description":"Partial training configuration override"},"data_sources":{"type":"array","description":"Array of data source configurations"},"samples":{"type":"array","description":"Inline training samples (input/output pairs)"},"status_filter":{"type":"string","description":"Filter jobs by status"}} |
| outputs_schema | {"result":{"type":"object","description":"Action-specific result (job, recipes, stats)"}} |
| tags | ["ai-agency","fine-tuning","lora","qlora","training","model"] |
| scope | {"orgs":"all","channels":["admin-ui","canvas-ui"]} |
Fine-tunes small open models (Qwen3-4B, Gemma-2B, etc.) on domain-specific data using LoRA/QLoRA adapters.
Training Data Sources → Data Prep → LoRA Fine-Tune → Evaluation → Export to LiteLLM
↓ ↓ ↓
Conversation Logs HuggingFace PEFT LiteLLM Model Registry
RAG Documents bitsandbytes QLoRA Local Model Fleet
File Uploads Transformers Trainer
Inline Samples
| Recipe | Base Model | Purpose | Epochs |
|---|---|---|---|
| writing_style | Qwen2.5-4B | Match user's tone, vocabulary, structure | 3 |
| codebase_conventions | Qwen2.5-4B | Learn project patterns and naming | 2 |
| domain_vocabulary | Qwen2.5-4B | Domain-specific terms and knowledge | 5 |
[{role, content}]{instruction, output}{prompt, chosen, rejected}