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
1.1 KiB
1.1 KiB
Zen Coder Model (7B MoE)
Specialized for code generation and technical tasks
Overview
Zen Coder is optimized for programming and technical documentation:
- Code generation and completion
- Technical documentation
- API development
- Code review and analysis
- Developer tools
Model Details
- Base Model: Qwen/Qwen3-Coder-7B-MoE-Instruct
- Parameters: 7B (2B active per token via MoE)
- Context Length: 2048 tokens
- Memory Required: ~16GB VRAM
Training
Quick Start
# Train with GSPO algorithm
gym train models/coder/configs/gspo_training.yaml
# Train with code-specific dataset
gym train models/coder/configs/gspo_training.yaml \
--dataset code_alpaca
Configuration
- Algorithm: GSPO (optimized for MoE stability)
- LoRA Rank: 16
- Learning Rate: 2e-5
- Batch Size: 2
Performance
- MMLU: 68.3
- HumanEval: 72.4 (Best-in-class)
- MT-Bench: 7.8
- Speed: 65 tokens/sec
Use Cases
✓ Code generation ✓ Code review ✓ Documentation ✓ Debugging ✓ Refactoring
Part of the Zen Model Family by Zoo Labs Foundation