RasoSynthTune: Text-to-Dataset & Fine-Tuning

ML teams fine-tuning models spend immense time hand-curating data or blindly trusting API-generated data that silently poisons training runs with hallucinations. RasoSynthTune solves this bottleneck as an autonomous multi-agent data refinery that converts raw, unstructured documents into verified fine-tuning datasets. The platform ensures quality through a rigorous three-layer verification shield: consensus-based LLM judge scoring, iterative self-refinement where low-scoring samples are dynamically revised using the judge's own critiques, and a Human-in-the-Loop (HITL) manual review API. Furthermore, the system actively truncates hallucinated claims in generated multi-turn dialogues. Crucially, RasoSynthTune provides genuine hardware utilization rather than just standard API routing. Heavy multi-agent workflows and dynamic inference routing natively prioritize Fireworks AI, executing Google Gemma models (gemma2-9b-it and gemma2-27b-it) directly on AMD Instinct GPUs. The platform takes this further with an on-metal LoRA distillation loop, utilizing ROCm to fine-tune a smaller model directly on AMD hardware using the pipeline's own rigorously verified output data. Built for enterprise scale, the entire architecture—featuring a LangGraph multi-agent engine, FastAPI router, PostgreSQL, Redis, Qdrant vector storage, and a React/Next.js visual workflow studio—is fully containerized and features OpenTelemetry and Prometheus observability. It is engineered to deploy instantly on the AMD Developer Cloud, delivering transparent datasets (JSONL, Parquet) complete with hardware and model provenance cards.