Update README: proper training infrastructure model card, remove Gradio Space metadata

README.md

@@ -1,148 +1,81 @@
 ---
-title: Zen Training
-emoji: 🧘
-colorFrom: blue
-colorTo: purple
-sdk: gradio
-sdk_version: 4.0.0
-app_file: app.py
-pinned: true
+language: en
 license: apache-2.0
-hardware: a10g-large
+tags:
+  - training
+  - zen
+  - zenlm
+  - hanzo
 ---
 
-# 🧘 Zen Training Space
+# Zen Training
 
-**Unified Training Platform for All Zen Models**
+Training infrastructure and recipes for the Zen model family.
 
-Train any Zen model with any dataset combination from HuggingFace. Everything runs directly from HF datasets - no local storage needed!
+**Zen LM by Hanzo AI** — Open training configurations for all Zen models.
 
-## 🎯 Features
+## Overview
 
-### Supported Models
+This repository contains the training configurations, scripts, and recipes used to train Zen models using the Zen MoDE (Mixture of Distilled Experts) architecture. All training runs use mixed-precision distributed training with full support for LoRA/QLoRA fine-tuning and alignment techniques.
 
-**Language 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
+## Training Recipes
 
-**Vision-Language Models:**
-- `zen-vl-4b` - Efficient VL with function calling
-- `zen-vl-8b` - Enhanced VL capabilities
-- `zen-vl-30b` - Maximum VL performance
+| Model | Type | Parameters | Context | Hardware |
+|-------|------|-----------|---------|----------|
+| Zen Nano | Dense | 0.6B | 32K | 1x H100 |
+| Zen Eco | Dense | 4B | 64K | 4x H100 |
+| Zen Pro | Dense | 8B | 128K | 8x H100 |
+| Zen MAX | MoE | 235B (22B active) | 128K | 64x H100 |
 
-### Supported Datasets
+## Features
 
-**Agent Training (ADP):**
-- AgentTuning OS/KG/DB (~15k samples)
-- Synatra (99k agent trajectories)
-- Code Feedback (66k samples)
-- Go Browse (27k web interactions)
+- Mixed precision training (BF16)
+- Gradient checkpointing
+- Distributed training with FSDP / DeepSpeed ZeRO-3
+- LoRA / QLoRA fine-tuning support
+- RLHF and DPO alignment pipelines
+- Dataset mixing and curriculum scheduling
+- Evaluation harness integration
 
-**Function Calling:**
-- xLAM 60k (Salesforce high-quality function calling)
+## Supported Training Tasks
 
-**Instruction Tuning:**
-- Alpaca (52k instruction samples)
+- Instruction tuning
+- Function calling
+- Agent trajectory training
+- Vision-language alignment
+- Code generation fine-tuning
+- Reasoning / chain-of-thought distillation
 
-## 🚀 How to Use
+## Dataset Support
 
-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 recipes support direct streaming from HuggingFace datasets:
 
-## ⚙️ Training Configuration
+- Instruction tuning corpora
+- Agent behavior datasets
+- Function calling datasets
+- Code and math reasoning sets
+- Multilingual alignment data
 
-### Recommended Settings
+## Quick Start
 
-**4B Models (A10G - 24GB):**
-- Batch Size: 1-2
-- Max Samples: 10,000-30,000
-- Time: 4-8 hours
-- Cost: ~$3-5
+See [github.com/zenlm/zen-family](https://github.com/zenlm/zen-family) for full documentation, training scripts, and configuration files.
 
-**8B Models (A100 - 40GB):**
-- Batch Size: 2-4
-- Max Samples: 30,000-50,000
-- Time: 8-12 hours
-- Cost: ~$15-20
-
-**32B Models (A100 - 80GB):**
-- 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
+```bash
+git clone https://github.com/zenlm/zen-family
+cd zen-family/training
+pip install -r requirements.txt
+python train.py --config configs/zen-pro-8b.yaml
 ```
 
-## 🔒 Requirements
-
-- HuggingFace Pro account (for GPU access)
-- Write access to output repository
-- HF_TOKEN secret set in Space settings
-
-## 💡 Tips
+## Related Repositories
 
-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
+| Repo | Description |
+|------|-------------|
+| [zenlm/zen-family](https://huggingface.co/zenlm/zen-family) | Model family overview |
+| [zenlm/zen-nano-600m-instruct](https://huggingface.co/zenlm/zen-nano-600m-instruct) | Zen Nano — 0.6B |
+| [zenlm/zen-pro-8b-instruct](https://huggingface.co/zenlm/zen-pro-8b-instruct) | Zen Pro — 8B |
+| [zenlm/zen-max-235b-a22b-instruct](https://huggingface.co/zenlm/zen-max-235b-a22b-instruct) | Zen MAX — 235B MoE |
 
-## 📚 Resources
-
-- **Website**: https://zenlm.org
-- **GitHub**: https://github.com/zenlm
-- **Models**: https://huggingface.co/zenlm
-- **Datasets**:
-  - [ADP](https://huggingface.co/datasets/neulab/agent-data-collection)
-  - [xLAM](https://huggingface.co/datasets/Salesforce/xlam-function-calling-60k)
-
-## 📄 License
+## License
 
 Apache 2.0
-
-## 🙏 Citations
-
-```bibtex
-@software{zen-training-2025,
-  title={Zen Training: Unified Training Platform for Zen Models},
-  author={Zen AI Team},
-  year={2025},
-  url={https://huggingface.co/spaces/zenlm/zen-training}
-}
-
-@article{adp2024,
-  title={Agent Data Protocol},
-  author={NeuLab},
-  journal={arXiv preprint arXiv:2510.24702},
-  year={2024}
-}
-
-@dataset{xlam2024,
-  title={xLAM Function Calling Dataset},
-  author={Salesforce Research},
-  year={2024}
-}
-```