gpt-oss-120b · SDPO++ Exp 10c · step-120 (~1 epoch) — RAW Tinker LoRA adapter

• Checkpoint: global step 120 (= 118 epoch-1 steps + 2 epoch-2 steps ≈ 1 epoch)
• LoRA: r=32, alpha=32, target_modules="all-linear", dropout=0
• Base: openai/gpt-oss-120b

⚠️ Format note

1. On a machine that has the base model, run:
   python

   1from tinker_cookbook import weights 2weights.build_lora_adapter( 3 base_model="openai/gpt-oss-120b", 4 adapter_path="<this repo>", 5 output_path="<peft_out>", # PEFT-format, vLLM --lora-modules ready 6)
   (or build_hf_model(...) to produce a fully merged model)
2. Serve the converted adapter/model with vLLM.