Add HumanEval benchmark results (pass@1: 3.0%)

Upload Devstral Small 2507 fine-tuned LoRA adapters (QLoRA r32, trained on 2100 AI coding conversations)

initial commit