Record the token budget: measured visual and text tokens

README.md

@@ -113,3 +113,28 @@ To request removal of an image or to report a rights problem, open an issue in t
 project repository or contact the dataset maintainers (smartwhatt/thai_spatial_reasoning). Each
 image row carries `source_url`, `image_license`, and `rights_evidence_uri`, which is enough
 to identify and remove the exact image in a future revision.
+
+## Token budget
+
+Measured with the `Qwen/Qwen3-VL-2B-Instruct` image processor over every one of the 34,764 images in this release,
+so these are the tokens a Qwen3-VL model actually sees, not an estimate. Text is counted with the
+processor tokenizer over captions, questions and answers.
+
+| split | images | visual tokens | text tokens |
+|---|---|---|---|
+| train | 30,481 | 13,715,675 | 6,628,004 |
+| validation | 2,271 | 3,788,003 | 683,998 |
+| test | 2,012 | 1,467,105 | 481,804 |
+| **total** | **34,764** | **18,970,783** | **7,793,806** |
+
+**Total: 26,764,589 tokens** — 18,970,783 visual plus 7,793,806 text, an average of 545 visual and
+224 text tokens per image.
+
+Real images cost far more than renders: 1,778 visual tokens per image on average against
+300 for a render, because a 1240x1754 photograph is not a 640x480 scene. The training
+splits are render-heavy, so most of the training budget today is the fixed 300-token renders, and the
+real photographs are where the per-image weight sits.
+
+The continuous-pretraining index is a separate figure: 243,348 rows carrying 8,175,985 text
+tokens, and about **140,800,645 tokens** if every row trains with its image attached. Quote
+that number when a training recipe asks for a slice size.

TOKENS.md

@@ -0,0 +1,18 @@
+# Token budget
+
+Measured with the `Qwen/Qwen3-VL-2B-Instruct` image processor over every image in this release, and the processor
+tokenizer over all text. The dataset card carries the per-split table.
+
+| | |
+|---|---|
+| images | 34,764 |
+| visual tokens | 18,970,783 |
+| text tokens | 7,793,806 |
+| **total tokens** | **26,764,589** |
+| visual tokens per image | 545 |
+| text tokens per image | 224 |
+| CPT index rows | 243,348 |
+| CPT index, text only | 8,175,985 |
+| CPT index, one image per row | ~140,800,645 |
+
+A real photograph averages 1,778 visual tokens; a render is a fixed 300 at 640x480.

tokens.json

@@ -0,0 +1,34 @@
+{
+ "model": "Qwen/Qwen3-VL-2B-Instruct",
+ "images": 34764,
+ "visual_tokens": 18970783,
+ "text_tokens": 7793806,
+ "total_tokens": 26764589,
+ "sampled": false,
+ "visual_by_split": {
+  "test": 1467105,
+  "train": 13715675,
+  "validation": 3788003
+ },
+ "text_by_split": {
+  "test": {
+   "caption": 112139,
+   "question": 271265,
+   "answer": 98400
+  },
+  "train": {
+   "caption": 1631105,
+   "question": 3691802,
+   "answer": 1305097
+  },
+  "validation": {
+   "caption": 130054,
+   "question": 399738,
+   "answer": 154206
+  }
+ },
+ "visual_by_kind_mean": {
+  "real": 1778.2940463828315,
+  "rendered": 300.0
+ }
+}