Save and resume

Training writes logs and configuration to the output directory and overwrites latest/ on each checkpoint save. Copy the complete directory after saving finishes to retain a step. Frozen weights still depend on the original model_path.

Bash

python -m valen.train --config finetune.json --resume output/domain/run/latest

--resume restores optimizer, random state and data cursor; keep the process count, data and learning rates unchanged. Use --initialize to start a new experiment with different data or stages.

Load

Bash

python -m valen.inference --checkpoint output/domain/run/latest \
  --data test.jsonl --output predictions.jsonl

Export

A training checkpoint is not a standalone deployment model. Adapter exports need the matching base; complete exports merge adapters into it, while full SFT exports all trained weights.

Use the existing export tool. It requires completed-epoch records and a verified base manifest; a short smoke does not satisfy the release pipeline. Check native predictions after export, then deploy through ONNX or GGUF.