config.txt

[OUTPUT]
FOLDER_RESULTS=./result
FOLDER_CPOINTS=./checkpoint
FOLDER_MODELS=./model

[EVAL]
FOLDER_PAGES=./small_data_samples

[TRAIN]
FOLDER_PAGES=./small_data_samples

cross_runs = 0
# Path to an explicit cross-validation folds CSV (issue #15). When set (or via --folds_csv),
# the train/dev/test split is read from the file instead of being regenerated; pages absent
# from the CSV are excluded. The CSV is staged by the maintainer and is NOT committed.
folds_csv =
test_size = 0.1
log_step = 100
epochs = 3
lr = 5e-5
max_categ = 14000
Training = False
Testing = False

[SETUP]
files_format = png
seed = 420
base_model = timm/regnety_160.swag_ft_in1k
batch = 16
top_N = 3
raw = False
inner = True

[YOLO]
use_yolo    = False
yolo_base   = yolov8s-cls.pt
# available: yolov8n yv8s yv8m yv8l yv8x AND yolo11n y11s y11m y11l y11x
yolo_imgsz  = 224
yolo_epochs   = 30
yolo_patience = 20
yolo_lr0      = 0
yolo_dropout  = 0.1
yolo_cache    = True

[INPUT]
FOLDER_INPUT=./small_data_samples
chunk_size = 100
chunking = False

[HF]
repo_name = ufal/vit-historical-page
# REVIEW FIX (Blocker D/K): the HF access token must NOT be committed here.
# It is read from the HF_TOKEN environment variable at runtime (see run.py).
# Leave this blank in version control. Public models need no token; only gated
# models / push_to_hub require one.
#   export HF_TOKEN=hf_xxx   (Unix)
#   setx HF_TOKEN hf_xxx     (Windows)
token =
use_hf = False
revision = main
latest = v4.3

[DOCUMENT]
# atrium_document ("paradata pair") integration — see atrium_document_adapter.py and the hub's
# docs/document-schema.md. Single-document (-f/--file) runs use document_json/document_json_out
# as FILE paths; batch (-d/--directory) runs use document_json_dir/document_json_out_dir as
# DIRECTORIES, one <doc_id>.document.json per document. Left blank by default (opt-in): the
# adapter is a no-op unless an out path/dir is set, on the CLI or here.
document_json =
document_json_out =
document_json_dir =
document_json_out_dir =
strict = False