Motivation

trainer turns agent trajectories into ready-to-train datasets and fine-tunes SWE coding models on them.

It is the supervised-fine-tuning stage of the SWE-Lego-Live pipeline, sitting between trajectory generation (tracer) and evaluation (evaluator). It can convert good rollouts produced upstream or consume a ready-made Hugging Face/local ShareGPT dataset, then trains a model with the patched SWE-Lego LLaMA-Factory + DeepSpeed ZeRO-3 — all driven by a single config.yaml.

curator ─▶ tracer ─▶ trainer ─▶ evaluator

trainer provides:

• One-config runs — source trajectories, conversion settings, dataset, model, hyperparameters, and infrastructure all live in config.yaml
• Scaffold-aware conversion — Claude Code, OpenCode, OpenHands SDK, and Terminus-2 trajectories convert through the swe_data_process package
• Ready-made LF sources — load a Hub dataset config, select one exact Hub file, or register a local ShareGPT JSON
• Quality scoring — rule-based scoring runs automatically; optional LLM scoring layers on top
• Eval-leak protection — SWE-bench benchmark repos are filtered out of training data by default
• Multi-GPU training — LLaMA-Factory through torchrun with DeepSpeed ZeRO-3 on a single 8-GPU node
• A live web dashboard that reads the training output directly, with multi-run comparison and optional wandb

Where to go next

• Getting Started — set up the environment and run your first training job
• Example Run — inspect measured runtime, curves, metrics, and artifacts from the checked-in profile
• Core Concepts — trajectories, IM/LF formats, datasets, runs, and checkpoints
• Data Pipeline — convert trajectories into LLaMA-Factory datasets
• Training — launch the pipeline, tune the configuration, and read the outputs
• Dashboard — monitor training runs in real time