Skip to main content
Each training or evaluation is a Run. Every Run originates in exactly one Experiment and is a row in its Measurement’s Ledger. This page covers the row’s whole life: launch, record, chain, read.

​
Runs

A Run is one of two kinds. A training Run (train@N) produces a Checkpoint. An eval Run (eval@N) scores one, with eval@0 scoring the base model. N is the Checkpoint’s absolute step count, never a run index. The training Run that ends at step 50 is train@50. eval@50 scores that Checkpoint. Address a Run as <measurement>/<run> anywhere the CLI takes one (logs, show, stop, promote). Use the run ID or an alias like eval@0. When an alias matches several rows, the latest wins.

​
Checkpoints

A Checkpoint is the model’s saved state at one step. Two shapes matter. The full resume set (weights, optimizer, and dataloader state) continues training. The servable folder (an HF-layout weights and tokenizer directory) is what an inference server loads. A load-check proves a servable folder actually serves: one prompt through a real serving stack. It is not a score and not a Run. The final claim’s Checkpoint selection is loop discipline: see The loop.

​
Lineages

A Lineage is the group of Runs sharing a base model and comparability config. You read deltas within a Lineage, never across two. Each Lineage needs its own Baseline. Monte derives a Lineage from row fields at read time and never stores it.

​
The Ledger

The Ledger is append-only: rows are never rewritten. Each Run writes its own event file, one writer per file. Every surface (the CLI, the Console) reads through monte.api, never the files. The file format is internal. Every row records full provenance: git SHA, config hash, Split hash, eval-config hash, comparability hash, cost, and an image digest. A training Run’s digest is the true container image digest. An eval Run serves from a venv, so its digest is a venv: hash over the serving stack’s resolved versions. Any score is traceable to its exact inputs.

​
Traces

A Trace is one Run’s attempt at one Task: the rollout verbatim plus the grader’s fields (score, expected and extracted answer). Per-task results and Traces have no CLI command. Read them in the Console, or through monte.api from Python.

​
One command at a time

One lock per Measurement serializes mutating commands. A second command refuses instead of waiting. A dead process on the same host reclaims automatically. If a gone box left a lock behind, you remove it by hand: see Troubleshooting.

​
Run artifacts

Each run directory on the durable tier holds the launched plan (plan.json), the event file (events.jsonl), and for training Runs the training log and tensorboard files.

The loop

What the rows mean: baselines, deltas, and the final claim.

CLI reference

The commands that write and read these rows.