Overview

Trainer turns agent trajectories into training-ready datasets and fine-tunes SWE coding models on them.

It is the supervised-fine-tuning stage of the LegoFlow 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 train a model with the patched LegoX LLaMA-Factory and DeepSpeed ZeRO-3. A single config.yaml drives the run.

Training here is a measurement, not a one-off deliverable: it is how the pipeline finds out whether the upstream data carries signal. Both the data selection and the training hyperparameters are config fields, so an operator, human or agent, can change one variable, rerun, and compare against the previous run rather than rebuilding a training setup by hand.

Trainer provides:

• One-config runs where source trajectories, conversion settings, dataset, model, hyperparameters, and infrastructure live in config.yaml.
• Flexible data intake from Tracer LF files, Hugging Face datasets, local ShareGPT JSON, or raw Harbor jobs.
• A clean Tracer handoff through lf.json, with raw-trajectory conversion available as a fallback.
• Eval-leak protection that filters SWE-bench benchmark repos out of training data by default.
• Multi-GPU training through LLaMA-Factory, torchrun, and DeepSpeed ZeRO-3 on a single 8-GPU node.
• A web dashboard that reads training output directly, supports multi-run comparison, and can include wandb.

Where to go next

• Getting Started: set up the environment and run your first training job.
• Design: see how Trainer converts data, registers datasets, launches training, and publishes checkpoints.
• Output Format: find checkpoints, logs, generated configs, and the Evaluator handoff.
• Dashboard: monitor training runs in real time.