Home

Gym

AI Model Training Platform

Democratizing AI training and fine-tuning for researchers, educators, and developers worldwide

By Zoo Labs Foundation Inc - A 501(c)(3) Non-Profit Organization

Get Started → Learn CL-GRPO

99.8%

Cost Reduction

100+

Models Supported

10+

Training Methods

$18

Training Cost

🦁 About Zoo Labs Foundation¶

Zoo Labs Foundation Inc is a 501(c)(3) non-profit organization dedicated to democratizing artificial intelligence. We believe that advanced AI training should be accessible to everyone - from individual researchers and educators to small teams and large organizations. Gym is our flagship open-source platform, embodying our mission to break down the financial and technical barriers that have kept AI innovation in the hands of a few.

Our Mission: Make AI training 99.8% cheaper, 1000× faster to deploy, and completely transparent.

🚀 Revolutionary: Continuous Learning GRPO¶

 Learn from experience, not parameters - Our breakthrough Training-Free GRPO achieves comparable or better performance than traditional fine-tuning by operating in the context space instead of parameter space. 

💰

99.8% Cost Reduction

Train for $18 instead of $10,000+. Use 50-100 examples instead of thousands. Complete in minutes instead of hours/days.

📈

Better Performance

+2-5% improvement over traditional fine-tuning on AIME math benchmarks. Achieves 82.7% on AIME24, 73.3% on AIME25.

🔍

Human-Readable

Every learned experience is natural language - transparent, auditable, and explainable. No black box parameters.

🧩

Composable

Experiences are modular - add, remove, or modify without retraining. Share knowledge across models and domains.

🌐

Decentralized

Contribute to global semantic memory via DSO. 31.7× BitDelta compression, Byzantine-robust aggregation.

⚡

Lightning Fast

Minutes to adapt, not days. No GPU clusters required. Runs on CPU or single GPU. Instant deployment.

📊 Performance Comparison¶

Metric	Traditional Fine-Tuning	LoRA/QLoRA	Gym (CL-GRPO)
💰 Cost	$10,000+	$1,000-5,000	$18 ✨
📚 Data Required	10,000+ examples	1,000-5,000	50-100 ✨
⏱️ Training Time	Hours/Days	Hours	Minutes ✨
🔍 Interpretability	Black box	Black box	Human-readable ✨
🧩 Modularity	Monolithic	Single adapter	Composable ✨
📈 Performance	Baseline	Baseline - 2%	Baseline + 2-5% ✨
🖥️ GPU Required	Multiple	1-2	Optional ✨
🚀 Deployment	Complex	Moderate	Instant ✨

🎯 Use Cases¶

🏥 Domain Adaptation

Specialize in Medical, Legal, Finance with just 50-100 domain examples. Cost-effective compliance and safety updates.

💬 Chat Agent Learning

Agents learn continuously from every conversation. No expensive retraining. Improve in production automatically.

🔄 Cross-Domain Transfer

Single frozen model + domain experiences > multiple fine-tuned models. Save costs and maintenance.

🛡️ Safety & Compliance

Update model behavior with auditable experiences. Human-readable safety constraints. DAO governance.

🛠️ Quick Start¶

Command LinePython APIWeb Interface

# Install Gym
pip install zoo-gym

# Train with Continuous Learning GRPO
gym train \
  --model_name_or_path Qwen/Qwen3-4B-Instruct \
  --template qwen3 \
  --dataset alpaca_en_demo \
  --finetuning_type lora \
  --output_dir ./output/my-model

from gym.train import run_sft
from gym.hparams import get_train_args

# Configure training
config = {
    "model_name_or_path": "Qwen/Qwen3-4B-Instruct",
    "template": "qwen3",
    "dataset": "alpaca_en_demo",
    "finetuning_type": "lora",
    "output_dir": "./output/my-model"
}

# Run training
model_args, data_args, training_args, finetuning_args, generating_args = get_train_args(config)
run_sft(model_args, data_args, training_args, finetuning_args, generating_args)

# Launch web UI
gym webui

# Open browser at http://localhost:7860
# Select model, dataset, and training method
# Click "Start Training"

🌟 Key Features¶

100+ Models

Qwen3, LLaMA 3.3, DeepSeek V3, Mistral, Mixtral, Gemma, ChatGLM, Phi, and more

10+ Methods

LoRA, QLoRA, PPO, DPO, GRPO, GSPO, Continuous Learning GRPO, Full Fine-tuning

Multimodal Support

Text, Vision (LLaVA, Qwen-VL), Audio (Qwen2-Audio), Video - unified training pipeline

Production Ready

Web UI, CLI, API server, Docker, Kubernetes - deploy anywhere, scale easily

🌐 Decentralized Semantic Optimization (DSO)¶

 Federated Active Inference at Token-Level - Share compressed semantic experiences across nodes with 31.7× BitDelta compression and Byzantine-robust aggregation. 

10,000× communication efficiency vs federated learning
Byzantine-tolerant - handles 33% malicious nodes
Privacy-preserving - natural language experiences
On-chain governance - DAO voting for experience quality

Learn about DSO →

Ready to Democratize AI?

Start training models in minutes, not days. Join the revolution.

Get Started → View on GitHub

🤝 Community & Support¶

💬

Discord Community

Join 1,000+ developers, researchers, and AI enthusiasts. Get help, share ideas, collaborate.

Join Discord →

📚

Documentation

Comprehensive guides, tutorials, API references. 30+ pages, 100+ code examples.

Read Docs →

🐛

GitHub

Open source, Apache 2.0. Report bugs, request features, contribute code.

View Repository →

🎓

Hugging Face

Pre-trained models, datasets, spaces. Try demos, download models.

Explore Models →

📄 License & Contributing¶

License: Apache 2.0 - Free for commercial and non-commercial use

Contributing: As a 501©(3) non-profit, we welcome: - Code contributions (bug fixes, features) - Documentation improvements - Dataset contributions - Bug reports and feature requests - Tax-deductible donations

Contributing Guide →