exp028b-downproj-dpo-lr1e6-ep2-merged

SFT + DPO merged model. Full 16-bit weights, no adapter loading required.

Training Pipeline

1. SFT: tomofusa/exp021b-blend-h-lora
2. DPO: u-10bei/dpo-dataset-qwen-cot (2 epoch, lr=1e-06, beta=0.1)

DPO Configuration

• Learning rate: 1e-06
• Beta: 0.1
• Loss type: ipo
• LoRA: r=64, alpha=128
• Max length: 1024