Replace ranking screenshot with table

README.md

@@ -173,7 +173,7 @@ Authentication with a Hugging Face token is required while the repository is pri
 
 ## 📈 Training Evaluation
 
-The following figures report a MiniCPM5 1B evaluation of the dataset variants shown in the original evaluation artifacts. The first figure tracks aggregate micro scores during training. The second reports the final aggregate ranking using six-category macro scores, 12-task micro scores, and a 13-task score that includes SIQA.
+The following results report a MiniCPM5 1B evaluation of the dataset variants shown in the original evaluation artifacts. The figure tracks aggregate micro scores during training. The table reports the final aggregate ranking using six-category macro scores, 12-task micro scores, and a 13-task score that includes SIQA.
 
 The labels in the figures are preserved as provided: `t1x`, `t1x_c`, `t2x`, `t2x_c`, and `fw_raw` in the curve; `tra1x`, `tra1x cleaned`, `tra2x`, `tra2x cleaned`, and `FineWeb 2025-06` in the ranking table.
 
@@ -187,8 +187,6 @@ The labels in the figures are preserved as provided: `t1x`, `t1x_c`, `t2x`, `t2x
 | 4 | `tra1x` | 8.916% | 8.358% | 7.841% |
 | 5 | `tra2x` | 8.692% | 8.118% | 7.637% |
 
-![Aggregate score ranking](assets/ultrafineweb-fineweb-edu-ranking.jpg)
-
 ## ❤️ Acknowledgements
 
 Ultra-FineWeb-L1 builds on [Common Crawl](https://commoncrawl.org/), [FineWeb](https://huggingface.co/datasets/HuggingFaceFW/fineweb), [DataTrove](https://github.com/huggingface/datatrove), [trafilatura](https://github.com/adbar/trafilatura), [fastText](https://fasttext.cc/), [NLTK](https://www.nltk.org/), and [ftfy](https://github.com/rspeer/python-ftfy).