A training iteration can produce several observations: a loss, an evaluation return, then a corrected metric. Treating every write as a new iteration misaligns the experiment; discarding repeated steps loses information. EvoRL, the reinforcement learning and evolutionary learning framework, added local JSONL recording to main on November 12, 2026. A filename follow-up arrived on November 13, Beijing time. Together, they provide a local record of what was written and a separate, step-aligned view for analysis.
Concept illustration: preserve individual observations, organize their combined view by training step.
A local artifact with the experiment attached
Selecting recorders=[log,jsonl] saves machine-readable metrics alongside ordinary logs without adding a tracking SDK. The versioned quickstart documents the same recorder setup used by the training scripts, so the JSONL backend receives the project, run name, resolved configuration, tags, output path and optional group.
Initialization writes that context as a metadata header. Each subsequent write appends one JSON object containing the supplied integer step and nested data, then flushes the file. The raw file preserves arrival order, including out-of-order steps and repeated writes. This makes it useful when moving an experiment’s output directory to a separate analysis machine: configuration and observations travel together in a readable local artifact.
Merge at reading time, not during recording
The loader load_jsonl_records(path), exported from evorl.recorders, reads the complete dataset and returns a list sorted by step. It recursively merges dictionaries for the same step; later writes replace conflicting values while leaving unmentioned metrics intact. Arrays are replaced as whole values, and a later null also replaces an earlier value. The loader skips the optional first-line metadata header rather than including it in the metric list.
Consider training loss and evaluation return written separately at step 10, followed by a corrected loss for that step. The raw file retains all three writes. The loaded record contains the latest loss and the evaluation return. A step 8 record arriving afterward sorts before step 10. This distinction matters when plotting learning curves: iteration identity comes from step, not from line position or the number of logging calls.
Know what survives serialization
The implementation and regression tests cover NumPy, JAX and pandas values. Arrays and Series become lists. DataFrames retain columns and row data, but omit their index. NaN, infinity and pandas missing values become JSON null. If a DataFrame index identifies samples, promote it to an explicit column before recording; that follows directly from the stored table representation.
Malformed or incomplete lines raise an error identifying the file and line number. The loader does not silently discard a truncated final record. It loads the complete dataset into memory, so plan analysis memory around the accumulated file. An application metric named step remains inside data, distinct from the recorder’s outer iteration field.
Give each run a distinct destination
The follow-up shares sanitize_filename between JSONL and script-created log files. It replaces / with _: MetaWorld/Hammer-v3 becomes MetaWorld_Hammer-v3. Other characters remain unchanged, and metadata keeps the original run name. The files use separate .jsonl and .log extensions. This replaces the initial commit’s percent-encoding rule; filename normalization is not a uniqueness guarantee.
Reinitializing the same JSONL file path overwrites its contents. Use a distinct output directory or filename for each run, especially where names differ only by slash versus underscore. Start with the configuration above on a source checkout containing both commits, then load a small experiment and inspect its header, repeated-step merge and destination before collecting a longer run. The earlier EvoRL development report explains the gradient, device and metric-shape context; this update makes those observations easier to retain and revisit locally.