File size: 2,180 Bytes
4fb167c
 
 
 
 
 
65fd232
4fb167c
65fd232
 
4fb167c
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
{
  "schema_version": "1.0",
  "repo_type": "dataset",
  "about": "Contract for the multi-track Scenic Font Analysis dataset. Every track lives in its own track_<letter>/ folder and must follow this schema so tracks can be merged without drift.",
  "layout": {
    "track_dir_pattern": "track_[a-z]",
    "required_files": ["ground_truth.csv", "images.tar.gz"],
    "optional_files": ["manifest.json", "README.md"],
    "image_location": "images.tar.gz (members are image_filename at archive root); alternatively images/<image_filename> when a track has <= 10000 images",
    "note": "Hugging Face rejects any repository directory with more than 10000 files, so multi-thousand-image tracks MUST ship a single images.tar.gz."
  },
  "ground_truth.csv": {
    "columns": [
      {"name": "image_filename", "type": "string", "required": true, "description": "Basename of the file inside images/ (e.g. regular_00000.jpg)."},
      {"name": "language", "type": "string", "required": true, "description": "e.g. hindi / japanese / korean."},
      {"name": "font_family", "type": "string", "required": true},
      {"name": "font_style", "type": "string", "required": true, "description": "Raw source style/weight label, kept for provenance."},
      {"name": "style_unified", "type": "string", "required": true, "allowed": ["regular", "bold", "italic", "bolditalic"], "description": "Canonical taxonomy produced by taxonomy.unify_style()."},
      {"name": "split", "type": "string", "required": true, "allowed": ["train", "val", "test"]},
      {"name": "is_clean", "type": "boolean", "required": true},
      {"name": "text", "type": "string", "required": true}
    ]
  },
  "label_spaces": {
    "style_unified": ["regular", "bold", "italic", "bolditalic"],
    "split": ["train", "val", "test"]
  },
  "unique_key": ["track", "image_filename"],
  "notes": [
    "Track A collapses 9 raw weights: thin/extralight/light/regular/medium -> regular; semibold/bold/extrabold/black -> bold.",
    "Italic/oblique raw labels map to italic (or bolditalic when the weight bucket is bold).",
    "image_filename must be unique within a track; the merged key is (track, image_filename)."
  ]
}