🧘 Zen Training Space

🎯 Features

Supported Models

• zen-nano (0.6B) - Edge deployment
• zen-eco (4B) - Balanced performance
• zen-omni (7B) - Multi-task
• zen-coder (14B) - Code generation
• zen-next (32B) - Frontier performance

• zen-vl-4b - Efficient VL with function calling
• zen-vl-8b - Enhanced VL capabilities
• zen-vl-30b - Maximum VL performance

Supported Datasets

• AgentTuning OS/KG/DB (~15k samples)
• Synatra (99k agent trajectories)
• Code Feedback (66k samples)
• Go Browse (27k web interactions)

• xLAM 60k (Salesforce high-quality function calling)

• Alpaca (52k instruction samples)

🚀 How to Use

1. Select Model: Choose from language or vision-language models
2. Select Datasets: Check multiple datasets to combine them
3. Configure Training: Set epochs, batch size, learning rate, max samples
4. Set Output Repo: Specify HuggingFace repo for trained model
5. Start Training: Click the button and monitor logs

⚙️ Training Configuration

Recommended Settings

• Batch Size: 1-2
• Max Samples: 10,000-30,000
• Time: 4-8 hours
• Cost: ~$3-5

• Batch Size: 2-4
• Max Samples: 30,000-50,000
• Time: 8-12 hours
• Cost: ~$15-20

• Batch Size: 1-2
• Max Samples: 50,000-100,000
• Time: 20-30 hours
• Cost: ~$50-80

📊 Dataset Combinations

For Agent Training:

ADP Synatra (80%) + xLAM (20%)
= Strong agent + quality function calling

For Code Models:

Code Feedback (70%) + Alpaca (30%)
= Code expertise + general instruction following

For VL Models:

ADP (all configs) + xLAM
= Complete vision-language agent training

🔒 Requirements

• HuggingFace Pro account (for GPU access)
• Write access to output repository
• HF_TOKEN secret set in Space settings

💡 Tips

1. Start Small: Test with 1,000 samples first
2. Mix Datasets: Combine complementary datasets for best results
3. Monitor Logs: Watch for OOM errors and adjust batch size
4. Save Often: Lower save_steps for longer training runs

📚 Resources

• Website: https://zenlm.org
• GitHub: https://github.com/zenlm
• Models: https://huggingface.co/zenlm
• Datasets:
  • ADP
  • xLAM

📄 License

🙏 Citations