wiruka/sdf-contrastive-grader-comp-users-loops

LoRA adapters (rank 32) trained on openai/gpt-oss-120b with Tinker as part of the measuring entities in SDF project. Source run: runs/sdf-contrastive-grader-comp-users-loops.

Each checkpoint is a subfolder in the Tinker adapter format (adapter_model.safetensors + adapter_config.json). To get a PEFT adapter for vLLM/SGLang, convert with tinker_cookbook.weights.build_lora_adapter(base_model=..., adapter_path=<subfolder>, output_path=...).

checkpoint	training step
000309	309
000618	618
000928	928
final	1237

Training config: "data/contrastive_grader_comp_users_loops/train.jsonl" (see config.json if present).