Skip to content

YakStacks/moonshiner-recipes

Folders and files

NameName
Last commit message
Last commit date

Latest commit

 

History

3 Commits
 
 
 
 
 
 
 
 
 
 
 
 
 
 

Repository files navigation

🥃 Moonshiner Recipes

Community-maintained recipe registry for Moonshiner - the drop-dead simple toolkit for distilling large language models.

What are Recipes?

Recipes are pre-configured .moon files that provide ready-to-use pipelines for common distillation and fine-tuning tasks. Just install and run!

Quick Start

# List available recipes
moonshiner recipes

# Search for recipes
moonshiner recipes --search coding

# Download a recipe
moonshiner pull distillation/code-assistant

# Run the downloaded recipe
moonshiner run code-assistant.moon

Available Categories

Category Description
📚 distillation/ Teacher-to-student knowledge transfer
🏥 domain/ Domain-specific fine-tuning (medical, legal, etc.)
🎯 alignment/ DPO/RLHF preference alignment
☁️ cloud/ Cloud-optimized pipelines (RunPod, Lambda, etc.)
🧪 experimental/ Cutting-edge techniques

Featured Recipes

📦 distillation/code-assistant

Distill GPT-4o into DeepSeek-Coder-7B for Python coding tasks.

  • Cost: ~$15-30 for 10k samples
  • Time: ~4-6 hours on 1x A100
  • Tags: coding, python, deepseek

📦 distillation/creative-writer

Distill Claude 3.5 Sonnet into Mistral-7B for creative writing.

  • Cost: ~$20-40 for 10k samples
  • Time: ~6-8 hours on 1x A100
  • Tags: creative, writing, mistral

📦 domain/medical-rag

Medical Q&A with strict hallucination controls (GPT-4 → Llama-3-8B).

  • Cost: ~$25-50 for 5k samples
  • Time: ~8-12 hours on 1x A100
  • Tags: medical, rag, safety-critical

📦 alignment/self-dpo

Self-alignment via DPO using synthetic preference pairs.

  • Cost: ~$10-20 for 5k pairs
  • Time: ~4-6 hours on 1x A100
  • Tags: dpo, self-alignment, preferences

Contributing

We welcome community contributions! To submit a recipe:

  1. Fork this repository
  2. Add your recipe file in the appropriate category folder
  3. Update index.json with your recipe metadata
  4. Submit a pull request

Recipe Requirements

  • Clear documentation in YAML comments
  • Tested on at least one GPU configuration
  • Cost estimates included
  • No hardcoded API keys (use environment variables)

Recipe Template

# Recipe: my-awesome-recipe
# Description: What this recipe does
# Author: your-github-username
# Version: 1.0.0
# Cost Estimate: $X-Y for Nk samples
# Hardware: Tested on 1x A100 40GB
# Tags: tag1, tag2, tag3

model: "your-student-model"
teacher: "your-teacher-api"

# ... rest of config

Index Format

The index.json file contains metadata for all recipes:

{
  "version": "1.0.0",
  "recipes": [
    {
      "name": "recipe-name",
      "category": "distillation",
      "description": "What the recipe does",
      "author": "github-username",
      "version": "1.0.0",
      "tags": ["tag1", "tag2"],
      "config_url": "https://raw.githubusercontent.com/junkyard22/moonshiner-recipes/main/distillation/recipe-name.moon",
      "readme_url": "https://github.com/junkyard22/moonshiner-recipes/blob/main/distillation/README.md"
    }
  ]
}

License

MIT License - see LICENSE

Related

About

No description, website, or topics provided.

Resources

License

Stars

Watchers

Forks

Releases

No releases published

Packages

 
 
 

Contributors