Shared text training execution

text train-lora uses TextLoRATrainingOptions, TextLoRATrainingPlan, and TextLoRATrainingOperation in MereRunCore. Library callers can resolve and run the same operation without importing ArgumentParser or the dashboard server.

Resolve a training plan

Options own numeric and resume validation, model-family selection, and default LoRA targets. Plan resolution reads the training and optional evaluation JSONL into typed dataset snapshots. Gemma vision training retains its single-image policy and image digests. Text families reject media. Resolution does not load weights or create output directories.

The CLI translates option errors to ArgumentParser validation errors and checks MLX availability before reading datasets for an actual training run. Viewer flags, stdout formatting, and diagnostic progress remain in the CLI.

Execute and publish

The operation creates the output directory, invokes the selected native pipeline, and writes the existing adapter manifest after successful completion. A dry run writes a prepared manifest without invoking a trainer or replacing an existing adapter. The default model, training settings, target ordering, and manifest format remain unchanged.

Native pipelines own model loading, tokenization, image integrity checks, optimizer state, metrics, and tensor artifacts. The shared operation preserves the training and evaluation examples, image digests, metadata, resume inputs, and progress callbacks passed to those pipelines. Failure or cancellation does not publish a new trained manifest. Existing checkpoints remain available for resume; cancellation does not roll back tensor files written by a trainer.

The CLI acquires its machine reservation at process startup. The operation does not acquire another permit. Library callers own admission around their training work. The optional CLI dashboard starts after its initial event is written and its server task is cancelled and awaited when the command exits.

Resume and validation boundaries

Training continues to use optimizer-bearing checkpoints and existing LoRA manifests. It does not use the image and transcription run inspect or run retry record formats. --resume-step retains its global-step meaning for legacy checkpoints; native trainers validate actual optimizer compatibility.

TextTrainingCompatibilityTests covers CLI validation order, empty datasets, dry-run artifacts, and target handling. TextLoRATrainingOperationTests covers family resolution, dataset snapshots, publication, failure, cancellation, and resume inputs with injected trainers. Existing vision and optimizer tests cover their respective contracts. These tests do not qualify real model checkpoints or measure training quality.