LocoTrainer

About

LocoTrainer addresses a narrow, recurring need: engineers working with the MS-SWIFT training framework need codebase analysis and structured reports, and a frontier-model agent is expensive for that repetitive task. The accompanying model was trained on 361,830 samples (agent trajectories, MS-SWIFT knowledge, and project structure paths) over roughly 25 hours on 8x H100s, and the surrounding framework replicates the Claude Code-style environment the model saw in training, down to absolute paths and system reminders, because that fidelity is what makes small-model tool calling reliable. The agent loop reads, searches, and analyzes the MS-SWIFT repository and emits markdown reports, running locally through GGUF quantization at zero API cost or through OpenAI-compatible providers. ML engineers fine-tuning with MS-SWIFT are the intended users.