model-trainer

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"]}

Model Trainer Skill

Fine-tunes small open models (Qwen3-4B, Gemma-2B, etc.) on domain-specific data using LoRA/QLoRA adapters.

Architecture

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

Pre-Built Recipes

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

Data Formats

• conversation: Chat messages [{role, content}]
• instruction: Alpaca-style {instruction, output}
• completion: Raw text continuation
• preference: DPO-style {prompt, chosen, rejected}