Shivam311/rocmpilot-agent-sft

ROCmPilot Agent SFT

This dataset contains seed supervised fine-tuning examples for ROCmPilot, a multi-agent tool that helps developers migrate PyTorch and vLLM workloads from CUDA/NVIDIA assumptions to AMD ROCm readiness.

The examples teach ROCmPilot's production-facing agent behaviors:

• —repo_doctor: scan repository evidence for CUDA/NVIDIA assumptions
• —migration_planner: identify CUDA/NVIDIA migration blockers and recommend ROCm-safe fixes
• —patch_planner: convert findings into scoped patch recommendations
• —benchmark_agent: prepare AMD validation plans without overstating proof
• —report_agent: write credible migration report sections with clear fallback/live-evidence boundaries
• —memory_agent: turn agent discussions into reusable long-context memories
• —cuda_rocm_coach: answer user questions about CUDA, ROCm, PyTorch, vLLM, and proof boundaries
• —endpoint_troubleshooter: diagnose AMD ROCm/vLLM endpoint states and fallback behavior
• —migration_kit: plan downloadable migration kits and validation files
• —agent_discussion: model useful lead/helper agent conversations with memory writes

Current seed size: 297 synthetic examples.

This is still a synthetic seed dataset, but it now covers the full ROCmPilot MVP workflow. Production training should continue adding real repository scans, human-reviewed patch previews, endpoint logs, and post-demo user feedback.