exp020-simpo-merged

SFT + CPO/SimPO merged model. Full 16-bit weights, no adapter loading required.

Training Pipeline

1. SFT: tomofusa/exp015-blend-h-lora
2. CPO/SimPO: u-10bei/dpo-dataset-qwen-cot (1 epoch, lr=5e-07, beta=2.5)

CPO/SimPO Configuration

• Trainer: CPOTrainer (reference-free)
• Loss type: simpo
• Learning rate: 5e-07
• Beta: 2.5 (SimPO scale, NOT DPO scale)
• SimPO gamma: 1.375
• CPO alpha: 0.0
• LoRA: r=64, alpha=128
• Max length: 1024