Parallelogram

Parallelogram is a strict CLI validator for LLM fine-tuning datasets. It catches every structural failure mode that silently kills training runs — malformed role sequences, empty turns, context window violations, duplicates, and encoding corruption — before any compute is spent. No quality scores, no warnings, no partial passes. If it exits 0, your training run will not fail because of data. Free, open source, runs entirely offline. Apache 2.0.​​​​​​​​​​​​​​​​