AI-ModCon · S-Villar · Aug 11, 2026 · Aug 11, 2026 · Aug 11, 2026 · Aug 11, 2026
diff --git a/.gitignore b/.gitignore
@@ -1,5 +1,6 @@
 # Data and configs
 *.toml
+!examples/**/*.toml
 
 # Logging files
 *.db
@@ -15,13 +16,18 @@ data/*
 
 # bash
 *.sh
+!examples/**/*.sh
+
+# Slurm job logs land in the submitting directory
+slurm-*.out
+slurm-*.err
 
 
 # Files
 *.png
 !docs/images/*.png
 *.txt
-!docs/requirements.txt
+!examples/**/*.txt
 *.csv
 !tests/references/*.csv
 *.pdf
@@ -163,7 +169,6 @@ celerybeat.pid
 .env
 .envrc
 .venv
-.venv-docs/
 env/
 venv/
 ENV/
@@ -231,3 +236,4 @@ cython_debug/
 marimo/_static/
 marimo/_lsp/
 __marimo__/
+output/
diff --git a/docs/images/matey-adaptation-sequence.png b/docs/images/matey-adaptation-sequence.png
diff --git a/docs/images/matey-detector-response.png b/docs/images/matey-detector-response.png
diff --git a/docs/images/matey-forgetting.png b/docs/images/matey-forgetting.png
diff --git a/docs/images/matey-xgc-detectors.png b/docs/images/matey-xgc-detectors.png
diff --git a/docs/images/matey-xgc-device-maps.png b/docs/images/matey-xgc-device-maps.png
diff --git a/examples/README.md b/examples/README.md
@@ -18,6 +18,7 @@ poetry run python -m src.main --config <path_to_toml>
 | [`mnist/`](mnist/README.md) | `mnist` | 3-layer CNN (`Cnn`, ~1M params) | Simulated: cumulative random affine per stream window | Auto-downloaded to `./data` | Yes — CPU is fine |
 | [`cifar/`](cifar/README.md) | `cifar10` | ViT-B/16 or VGG-11 (`VisionModelCifar`) | Simulated: random affine per stream window | Auto-downloaded to `./data` | GPU strongly recommended |
 | [`imagenet/`](imagenet/README.md) | `imagenet` | ViT-B/16 (`VisionModelImageNet`) | Simulated: cumulative random affine per stream window | **You provide** ILSVRC-2012 in `ImageFolder` layout | No — multi-GPU scale |
+| [`matey/`](matey/README.md) | `matey`, `matey_stream` | MATEY ViT surrogate (`MATEYHarness`) | **Real**: SOLPS simulations arriving from new scenarios and machines | **You provide** a SOLPS root and a MATEY checkpoint | No — multi-GPU scale, and needs the MATEY package |
 
 **Start with `mnist/`.** It is the only example that ships a pretrained
 checkpoint, downloads its own data, and finishes in minutes on CPU.
@@ -41,8 +42,12 @@ See [`docs/model_harness.md`](../docs/model_harness.md) for the full contract.
 
 ## How Drift Is Simulated
 
-None of these datasets drift on their own, so each harness manufactures drift the
-same way: `update_data_stream()` draws a seeded random affine transform
+`matey/` is the exception to everything in this section: its stream is a real
+sequence of simulations, so it needs no synthetic drift at all. See
+[`matey/README.md`](matey/README.md).
+
+None of the other datasets drift on their own, so each of those harnesses
+manufactures drift the same way: `update_data_stream()` draws a seeded random affine transform
 (rotation / scale / shear / translation) and rebuilds the train, validation, and
 stream loaders through it. Every time the stream is exhausted, another transform
 is drawn, so the input distribution keeps moving away from what the model was

diff --git a/examples/matey/Demo_SOLPS_vit.yaml b/examples/matey/Demo_SOLPS_vit.yaml
@@ -0,0 +1,75 @@
+basic_config: &basic_config
+  # Run settings
+  log_to_wandb: !!bool False #True # Use wandb integration
+  log_to_screen: !!bool True # Log progress to screen.
+  save_checkpoint: !!bool True # Save checkpoints
+  checkpoint_save_interval: 10 # Save every # epochs - also saves "best" according to val loss
+  debug_grad: !!bool True # Compute gradient/step_sizes/ect for debugging
+  true_time: !!bool False # Debugging setting - sets num workers to zero and activates syncs
+  num_data_workers: 2 #6 # Generally pulling 8 cpu per process, so using 6 for DL - not sure if best ratio
+  enable_amp: !!bool False # Use automatic mixed precision - blows up with low variance fields right now
+  compile: !!bool False # Compile model - Does not currently work
+  gradient_checkpointing: !!bool False # Whether to use gradient checkpointing - Slow, but lower memory
+  exp_dir: './Dev_SOLPS' # Output path
+  log_interval: 1 # How often to log - Don't think this is actually implemented
+  pretrained: !!bool False # Whether to load a pretrained model
+  # Training settings
+  drop_path: 0.1
+  batch_size: 64
+  max_epochs: 10
+  scheduler_epochs: -1
+  epoch_size: 20 
+  rescale_gradients: !!bool False # Activate hook that scales block gradients to norm 1
+  optimizer: 'AdamW' # DAdaptAdam 'AdamW' 'SGD'
+  scheduler: 'none' # Only cosine implemented
+  warmup_steps: 0 # Warmup when not using DAdapt
+  learning_rate: 1e-3 # 
+  weight_decay: 1e-3
+  n_states: 29 # Must be >= max field label + 1 in the dataset
+  state_names: ['Pressure', 'Vx', 'Vy', 'Density',  'Vx', 'Vy', 'Density', 'Pressure'] # These are not used now!
+  dt: 1 # Striding of data - Not currently implemented > 1
+  leadtime_max: 10 #prediction lead time range [1, leadtime_max]
+  autoregressive: !!bool True # autoregressive training or one-step prediction
+  supportdata: # Whether to use support data (e.g. input control actuator) as input
+  - input_control_act: !!bool True
+  n_steps: 3 #16 # Length of history to include in input
+  enforce_max_steps: !!bool False # If false and n_steps > dataset steps, use dataset steps. Otherwise, raise Exception.
+  accum_grad: 1  
+  # Model settings
+  model_type:  'vit_all2all' # no need for time_type and space_type inputs
+  #model_type:  'svit' #currently only support time_type=="all2all_time" and space_type=="all2all"
+  #time_type: 'all2all_time' # 
+  #space_type: 'all2all' #
+  #model_type:  'avit' #currently only support space_type=="axial_attention" and time_type=="attention"
+  #time_type: 'attention' # 
+  #space_type: 'axial_attention' #
+  tie_fields: !!bool False # Whether to use 1 embedding per field per data
+  embed_dim: 192 # Dimension of internal representation - 192/384/768/1024 for Ti/S/B/L
+  num_heads: 3 # Number of heads for attention - 3/6/12/16 for Ti/S/B/L
+  processor_blocks: 12 # Number of transformer blocks in the backbone - 12/12/12/24 for Ti/S/B/L
+  ##patch_size: [[1, 2, 2]] #[[1, 40, 40]] #, [32, 32], [64, 64]] #
+  tokenizer_heads:
+  - head_name: "tk-2D"
+    patch_size: [[1, 2, 2]] 
+  sts_model: !!bool False
+  sts_train: !!bool False  #when True, we use loss function with two parts: l_coarse/base + l_total, so that the coarse ViT approximates true solutions directly
+  #gammaref: 0.2 #pick all tokens that with variances larger than gammaref*max_variance to refine
+  #refine_ratio: 0.2 #ratio of coarse tokens picked to be refined
+  bias_type: 'PositionAreaBias'  # Options rel, continuous, none, PositionAreaBias
+  bias_MLP: !!bool True
+  # Data settings
+  #train_val_test: [.6, .2, .2]
+  augmentation: !!bool False # Augmentation not implemented
+  use_all_fields: !!bool True # Prepopulate the field metadata dictionary from dictionary in datasets
+  tie_batches: !!bool False # Force everything in batch to come from one dset
+  extended_names: !!bool False # Whether to use extended names - not currently implemented
+  embedding_offset: 0  # Use when adding extra finetuning fields
+  train_data_paths: [
+              ['examples/matey/data/fusionMT-data/solps/train', 'SOLPS2D', '','tk-2D'],
+              ]
+  valid_data_paths: [
+              ['examples/matey/data/fusionMT-data/solps/valid', 'SOLPS2D', '','tk-2D'],
+              ]
+  append_datasets: [] # List of datasets to append to the input/output projections for finetuning
+
+