Transcription Dataset: Code & Wiki (390K)
Text-to-image rendered dataset for training vision-language models to read code and text from images.
Schema
| Column | Type | Description |
|---|---|---|
image |
Image | Rendered grayscale JPEG |
prompt |
string | Transcription instruction (varied) |
response |
string | Ground truth text |
language |
string | python/javascript/java/c++/rust/go/english |
domain |
string | code or english |
length_bucket |
string | short/medium/long/gundam |
resolution |
string | Image dimensions (e.g., "512x384") |
text_tokens |
int | Estimated text token count |
vision_tokens |
int | Vision encoder token count |
compression_ratio |
float | text_tokens / vision_tokens |
Rendering Details
- Font: JetBrains Mono 12pt
- Format: Grayscale JPEG, quality 85
- Background: Dark (#222222)
- Dimensions: Rounded to multiples of 16
Intended Use
Training vision encoders (e.g., DeepSeek-OCR) to read code/text from images with high compression ratios.
Splits
- Train: 371,012 (95%)
- Test: 19,527 (5%)
- Downloads last month
- 18