An autonomous training and benchmarking ecosystem for LLM-based Site Reliability Engineering (SRE) agents. Built with a custom Gymnasium environment, it simulates real-world infrastructure failures modeled after major 2025–2026 production post-mortems, including the tj-actions supply-chain attack and cascading Kafka broker OOM failures.

We fine-tuned Qwen 2.5 (2B and 1.5B) using Supervised Fine-Tuning (SFT with LoRA rank 16) on verified expert recovery trajectories, followed by Group Relative Policy Optimization (GRPO) driven by a multi-component reward function. The pipeline includes distributed Docker orchestration, TRL 1.2.0 compatibility fixes for mixed bf16/fp16 precision on P100/T4 GPUs, and a live Gradio observability dashboard.