Fine-tuned Qwen2.5-7B to 96% of Claude Haiku on a domain-specific task using ~$3 of API calls and zero human labelers

In three linesFine-tuned Qwen2.5-7B reaches 96% of Claude Haiku performance on domain-specific decision-reasoning task using ~$3 API spend and zero human annotators. DV-DPO method: 3-voice council + adversarial cross-examination generates 1,040 training pairs. Latency 11s vs 3s (T4 4-bit). Autonomous loop in production with failure detection and auto red-teaming.

Summary generated by Claude — human-verified