mirror of
https://github.com/zenlm/zen-trainer.git
synced 2026-07-27 03:10:00 +00:00
Training framework for Zen Coder models supporting 5 architectures: - Zen Coder 4B (Qwen3-4B-Instruct) - Zen Coder 24B (Devstral-Small-2-24B) - Zen Coder 123B (Devstral-2-123B) - Zen Coder MAX (GLM-4.7 358B MoE) - Zen Coder ULTRA (Kimi K2 1T MoE) Features: - Unified ZenTrainer class with auto backend selection (MLX/Unsloth/DeepSpeed) - 12 ARC benchmark suite (TAU-Bench, SWE-bench, etc.) - Training cost estimation for 8xH200 @ $35/hr - CLI tools: zen-train, zen-benchmark Dataset: hanzoai/zen-agentic-dataset (8.47B tokens, 3.35M samples)