ls, no guessing the path.
Quick Start
1
List the checkpoints
Point at the run’s output directory and see every saved step, newest first.
2
Get JSON for scripts
Add
--json to emit a machine-readable array your shell can parse.How It Works
The command scansoutput_dir for checkpoint-<n> directories, sums each one’s file sizes, and sorts by step — newest first.
Only directories named checkpoint-<n> count — checkpoint-final is skipped, not read as step 0. Sorting is numeric, so checkpoint-1000 comes before checkpoint-200.
Configuration Options
An empty or missing directory exits with code
1 and prints a remediation suggesting save_steps in config or --model-dir pointing at the run’s output_dir.JSON Output
Each entry has three fields — the step number, the checkpoint path, and its size in bytes.Common Patterns
Auto-pick the newest step for export
Pipe--json through jq to grab the newest checkpoint path, then hand it to export.
.[0] is always the latest step.
Confirm a run actually saved
Run it right after an interrupted job to see what survived before deciding whether to resume.Best Practices
Only checkpoint-<n> directories count
Only checkpoint-<n> directories count
The scan matches
checkpoint-<n> exactly. A directory called checkpoint-final is ignored — it is never listed as step 0. Rename or move stray directories if you expect them in the listing.Run against output_dir, not final_model_dir
Run against output_dir, not final_model_dir
Checkpoints live under the run’s
output_dir (default outputs), not final_model_dir (default lora_model). Point -d at the same output_dir the run used.Use save_total_limit to control disk
Use save_total_limit to control disk
Each checkpoint is the full adapter size. Set
save_total_limit in config.yaml to keep only the newest few — the listing then stays short and disk stays bounded.Combine --json with jq in shell scripts
Combine --json with jq in shell scripts
--json gives a stable {step, path, bytes} schema. Parse it with jq to auto-select a checkpoint instead of hardcoding a step number that changes every run.Related
Checkpointing
Config keys —
save_strategy, save_steps, save_total_limit, resume_from_checkpoint.Export a trained model
Publish a chosen checkpoint to HF, GGUF, or Ollama.
Infer CLI
Prompt a model you just trained and watch it stream.
Train CLI
Full flag reference for every subcommand.