mirror of
https://github.com/zenlm/gym.git
synced 2026-11-04 17:11:44 +00:00
Major updates: - Implemented BitDelta 1-bit quantization (10-25× memory reduction) - Added DeltaQuant flexible quantization (INT2/4/8, Binary, Ternary) - Created DeltaSoup for Byzantine-robust community aggregation - Configured accurate Qwen3 models for Zen family (Nano, Eco, Coder, Omni, Next) - Added GRPO/GSPO training algorithms with memory optimizations - Simplified CLI from gym-cli to gym - Updated WebUI with Zoo Labs black monochromatic theme - Integrated Unsloth optimizations for 2-3× speedup - Added comprehensive test suite for quantization methods Zoo Labs Foundation Inc. - Copyright 2025
1.0 KiB
1.0 KiB
Zen Eco Model (4B)
Balanced performance for production deployments
Overview
Zen Eco provides the optimal balance between performance and resource usage, ideal for:
- Production applications
- Small-to-medium enterprises
- Real-time services
- Multi-instance serving
Model Details
- Base Model: Qwen/Qwen3-4B-Instruct
- Parameters: 4B
- Context Length: 1024 tokens
- Memory Required: ~8GB VRAM
Training
Quick Start
# Train with GSPO algorithm
gym train models/eco/configs/gspo_training.yaml
# Multi-GPU training
CUDA_VISIBLE_DEVICES=0,1 gym train models/eco/configs/gspo_training.yaml
Configuration
- Algorithm: GSPO (Group Sequence Policy Optimization)
- LoRA Rank: 8
- Learning Rate: 3e-5
- Batch Size: 4
Performance
- MMLU: 62.5
- HumanEval: 48.7
- MT-Bench: 7.2
- Speed: 85 tokens/sec
Use Cases
✓ API services ✓ Chatbots ✓ Content generation ✓ Code assistance ✓ Data analysis
Part of the Zen Model Family by Zoo Labs Foundation