Output Format

Trainer archives data preparation, model checkpoints, metrics, and console logs under artifacts/. The primary downstream output is the trained checkpoint path recorded back into config.yaml.

Artifact layout

The main Trainer outputs are organized as follows:

artifacts/
├── data/                               # downloaded or converted training data
│   ├── hf_data/                        # exact files fetched from Hugging Face
│   └── lf_data/                        # LLaMA-Factory datasets + dataset_info.json
├── training_config/                    # generated LLaMA-Factory YAML snapshots
├── model/                              # model checkpoints and training metrics
│   └── <run>/
│       ├── checkpoint-<step>/          # saved model checkpoint
│       ├── trainer_log.jsonl           # live per-step metrics
│       ├── trainer_state.json          # complete trainer state and log history
│       ├── train_results.json          # final loss and runtime summary
│       ├── all_results.json            # throughput and performance summary
│       └── training_loss.png           # rendered loss curve
├── logs/                               # console logs for training runs
└── index.yaml                          # append-only run history

Most users should inspect artifacts/model/<run>/ first, then use config.yaml -> runtime_info.output to find the latest checkpoint and metrics. Run timestamps recorded in artifacts/index.yaml use UTC.

Data preparation outputs

Prepared LF data is written under:

artifacts/data/lf_data/
├── <dataset>.json
└── dataset_info.json

For Hugging Face exact-file inputs, the downloaded source is cached under artifacts/data/hf_data/ before registration. Ready-made local_lf inputs are registered directly without conversion.

Generated training config

Before training starts, Trainer renders the effective LLaMA-Factory YAML from config.yaml and saves it under:

artifacts/training_config/

This snapshot helps reproduce a run without maintaining a separate hand-written training YAML.

Model and metric outputs

Each training run writes to:

artifacts/model/<run>/
├── checkpoint-<step>/
├── trainer_log.jsonl
├── trainer_state.json
├── train_results.json
├── all_results.json
└── training_loss.png

The checkpoint directory is the main model artifact. trainer_log.jsonl is the live dashboard source, while train_results.json and trainer_state.json summarize final loss, runtime, total steps, and training history.

Recorded config outputs

After a successful run, scripts/train.sh updates only config.yaml -> runtime_info.output:

Run history

artifacts/index.yaml stores archived run entries. Use it for run history, and use artifacts/model/<run>/ for the actual checkpoint and metrics payload. During an active run, the dashboard reads artifacts/model/<run>/trainer_log.jsonl directly; there is no separate status file to refresh.