Model Scaling Requirements
Purpose: decide when a larger GRU or Transformer-style model is justified.
Current Scale Checks
Run:
./scripts/estimate-model-scale.sh
Current artifacts:
data/derived/models/model_scale_synthetic_20000.jsondata/derived/models/model_scale_synthetic_balanced_20000.jsondata/derived/models/model_scale_all_italy_real_vlf.jsondata/derived/models/model_scale_central_italy_real_vlf.json
Current Status
| dataset | labeled rows | positives | negatives | sequence features | recommendation |
|---|---|---|---|---|---|
| synthetic 20000-step | 1005 | 364 | 641 | 168 | tiny synthetic-only patch Transformer is plausible |
| synthetic post-burn-in balanced | 804 | 360 | 444 | 168 | GRU/tabular smoke only |
| all-Italy real VLF-aligned | 23 | 23 | 0 | 50 | do not train supervised model |
| central-Italy real VLF-aligned | 23 | 0 | 23 | 50 | do not train supervised model |
Gates
Minimum gates before scaling:
- tabular or GRU smoke: at least 500 labeled rows and 50 rows per class
- larger GRU or tiny synthetic Transformer: at least 1000 labeled rows, 100 rows per class, 3 groups, and sequence features
- small synthetic Transformer: at least 5000 labeled rows, 500 rows per class, 5 groups, and sequence features
- real Transformer training: at least 5000 real labeled rows and 500 rows per class
The current real VLF-aligned rows fail the first gate because each scope has only one class.
CPU-Only Starting Point
If testing a larger synthetic-only model, keep it tiny:
- patch/channel encoder before cross-modality attention
d_model=32layers=2heads=2batch_size=32- dropout around
0.1 - strict ablations: seismic-only, VLF-only, direct avalanche only, full
Avoid large attention models and GPU-only dependencies on the current system.
Current Larger-Model Check
The first larger-model scaffold is implemented as a tiny CPU-only patch Transformer:
- command:
./scripts/train-tiny-patch-transformer.sh - output:
data/derived/models/tiny_patch_transformer/tiny_patch_transformer_model_run_summary.json - split: post-burn-in regime-balanced synthetic rows, 642 train / 162 test
- best calibrated balanced accuracy:
0.637500forsequence_piezo_vlf_only
This is below the current regime-balanced GRU sequence_full score of 0.650000, so it validates the model interface but does not replace the GRU baseline.
Recommendation
Do not scale real models yet. For synthetic-only engineering, compare the tiny patch Transformer across additional generated seeds before increasing size. Treat any synthetic gain as model-interface evidence, not earthquake prediction evidence.