Correct source row counts

README.md

@@ -16,8 +16,8 @@ the `chart2code_1.parquet` file of [DocTron-Hub/VinciCoder-1.6M-SFT](https://hug
 
 ## Selection
 
-- The file's 163,186 rows are split by SHA-256 of the image: 1% held out (1,678 rows, never included here),
-  161,508 for training.
+- The first 163,186 of the file's 210,137 rows (the Chart2Code head) are split by SHA-256 of the image:
+  1% held out (1,678 rows, never included here), 161,508 for training.
 - The training rows are shuffled once with `random.Random(0).shuffle(train_indices)`; this dataset is
   positions `[0, 10000)` of that order, in that order. `permutation.json` holds the full order
   (sha256 `bc4a30fffc10377e4c56b71dfbe5679a5e93e5d3723a392c4d7abcc3cc90aa34`), so `[10000, 10000 + N)` gives the next, non-overlapping chunk.