Choosing profiles

A profile describes a algorithm's lifecycle. A repository may declare several profiles and implement them in one package. Choose them with Copier's profiles multiselect, or supply a data file:

profiles: [finetune, probe, zero_shot_direct]

The canonical lifecycle table lists the exact steps.

Only preprocessing and pretraining share command names. They may share artifacts when input identities, configuration, payload format and training scope match. Selecting two profiles does not automatically make their artifacts compatible.

Profile	Use when
supervised	Learning a task-specific algorithm directly from the dataset's training labels.
finetune	Updating a pretrained encoder together with a task-specific head.
probe	Fitting a task-specific head on frozen encoder representations.
zero_shot_direct	Translating an ACES task into direct algorithm queries and decoding probabilities without labels.
zero_shot_materialized	Generating task-independent futures, then evaluating the task on history plus those futures.
predict_only	Preprocessing then applying a supplied fixed algorithm or an in-context learner, without a training command.

predict_only is the renamed packaged profile. It does not mean probe training. In-context learners may read explicitly selected train/tuning examples; fixed scores need no labels. Neither may read evaluated labels before its probabilities are final.

Each descriptor lives at algorithms/<profile>/model.yaml and is registered as <algorithm-name>__<profile> in MEDS-DEV. Each invocation selects exactly one declared profile. Examples:

meds-algorithm profiles
meds-algorithm commands --profile probe
meds-algorithm-add-to-meds-dev --meds-dev /path/to/MEDS-DEV --name example --profile probe
PROFILE=probe ./slurm/submit.sh --dry-run

The probe's infer_embeddings stage materializes keyed representations without labels. train_probe fits a head on training representations, and predict_probe applies it to explicit evaluated embeddings. Supervised/fine-tuned prediction instead takes input_data_dir and the fitted algorithm. Neither dependency is assumed to disappear just because training recorded its path.

The software contract gives every command's arguments, artifact type, label access, class responsibilities, compatibility checks, and acceptance requirements. See migration for an existing repository.