animal2vec workshop · MPIABDay 2 · 11:15–12:15

Adapt the recipe—leave as much untouched as you can

What to edit	Relevant fields	What to preserve / coordinate
Data and task	`task.data`, `unique_labels` `dataset.train_subset / valid_subset`	Label order, complete annotations, held-out groups
Starting model	`model.w2v_path` `task.sample_rate / conv_feature_layers`	Rate and frontend must match the pretrained checkpoint
Compute and output	`distributed_world_size`, `max_tokens` `update_freq`, `checkpoint.save_dir`	Match the actual launch; use a new run directory
Short-run timing	`max_update`, `freeze_finetune_updates` warmup, validation and saving intervals	A 200-update run must not inherit a 10,000-update freeze schedule

Use eight minutes to orient the room, then work directly in a copy of configs/MeerKAT/finetune_mixup_100.yaml. Do not edit the reproduction baseline in place. The checked-in recipe has four ranks, nine accumulated microbatches, 30,000 updates, a 10,000-update classifier-only phase, 2,000 warmup updates, and late validation. A short teaching run needs coordinated timings and checkpoints, not only a smaller max_update. model.feature_grad_mult=0.0 keeps the local frontend frozen even after Transformer unfreezing. CPU-smoke configurations use tiny architectures and are not compatible shortcuts for large official weights. checkpoint.save_dir may be supplied through the config/CLI workflow; confirm the actual command in the live clinic. The table is guidance, not a complete runnable YAML file.

Suggested time: 8 min. Slide ID: d2-config.