Zen Gym

Unified AI Model Training Platform

---

Zen Gym is the training infrastructure for the Zen model family. It provides a single interface for supervised fine-tuning, reinforcement learning from human feedback, and model export across all Zen architectures — from 0.6B parameter edge models to 397B MoE frontier systems.

Supported Models

Model	Parameters	Type
zen-nano	0.6B	Language
zen-eco	4B	Language
zen-eco-instruct	4B	Instruct
zen-eco-thinking	4B	Reasoning
zen-eco-coder	4B	Code
zen-eco-agent	4B	Agent
zen-voyager	32B	Language
zen4-mini	8B	Language
zen4-pro	80B MoE	Language
zen4-thinking	80B MoE	Reasoning
zen4-coder-pro	80B MoE	Code
zen-max	235B MoE	Language
zen4-max	397B MoE	Language
zen-vl-*	4B-30B	Vision-Language

Quick Start

pip install zen-gym

# LoRA fine-tune zen-eco on Alpaca
zen-gym train --model zen-eco --method lora --dataset alpaca

# Launch web UI
zen-gym ui

Training Methods

Method	Stage	Description
Full	SFT	16-bit full parameter fine-tuning
LoRA	SFT	Low-Rank Adaptation, ~30% memory of full
QLoRA	SFT	Quantized LoRA (4/8-bit), ~10% memory of full
DoRA	SFT	Weight-Decomposed Low-Rank Adaptation
DPO	RLHF	Direct Preference Optimization
PPO	RLHF	Proximal Policy Optimization
GRPO	RLHF	Group Relative Policy Optimization
GSPO	RLHF	Group Sampled Policy Optimization (MoE-optimized)
KTO	RLHF	Kahneman-Tversky Optimization
ORPO	RLHF	Odds Ratio Preference Optimization
SimPO	RLHF	Simple Preference Optimization

Hardware Requirements

Model Size	Full	LoRA	QLoRA (4-bit)
0.6B	8 GB	4 GB	2 GB
4B	32 GB	16 GB	8 GB
8B	48 GB	24 GB	12 GB
32B	128 GB	48 GB	24 GB
80B MoE	256 GB	80 GB	48 GB
235B MoE	512 GB	160 GB	80 GB
397B MoE	768 GB	256 GB	128 GB

VRAM listed per-GPU. Multi-GPU setups supported via DeepSpeed ZeRO and FSDP.

Installation

git clone https://github.com/zenlm/gym
cd gym
pip install -e ".[torch,metrics]"

Optional accelerators:

# FlashAttention-2
pip install flash-attn --no-build-isolation

# Unsloth (2-5x speedup)
pip install unsloth

Usage

CLI

# Supervised fine-tuning with LoRA
zen-gym train \
  --model zen-eco \
  --method lora \
  --dataset alpaca \
  --lora-rank 128 \
  --batch-size 4 \
  --epochs 3 \
  --lr 2e-5

# GRPO reinforcement learning
zen-gym train \
  --model zen-eco \
  --method grpo \
  --dataset preference_data \
  --lr 1e-5

# QLoRA 4-bit training
zen-gym train \
  --model zen-voyager \
  --method qlora \
  --quant-bits 4 \
  --batch-size 1 \
  --gradient-accumulation 16

# Export to GGUF
zen-gym export \
  --model ./output/zen-eco-lora \
  --format gguf \
  --quant Q4_K_M

Config File

# configs/zen-eco-lora.yaml
model: zen-eco
method: lora
dataset: alpaca
lora_rank: 128
lora_alpha: 64
batch_size: 4
gradient_accumulation: 4
learning_rate: 2e-5
epochs: 3
flash_attn: true
output_dir: ./output/zen-eco-lora

zen-gym train --config configs/zen-eco-lora.yaml

Web Interface

zen-gym ui
# Open http://localhost:7860

Monitoring

Zen Gym integrates with standard experiment tracking:

• TensorBoard -- built-in, zero config
• Weights and Biases -- --report-to wandb
• MLflow -- --report-to mlflow

Related

• Zen Models -- Pre-trained model weights on Hugging Face
• Zen Engine -- High-performance inference runtime
• Zen Docs -- Documentation and guides
• Zen LM -- GitHub organization

---

Apache 2.0 -- Zen LM -- Hanzo AI