Datasets:
remove todo
Browse files
README.md
CHANGED
@@ -499,7 +499,7 @@ We fine-tuned a Bert-like regression model using these annotations, based on [Sn
|
|
499 |
The classifier is available at: [https://huggingface.co/HuggingFaceFW/fineweb-edu-classifier/](HuggingFaceFW/fineweb-edu-classifier/)
|
500 |
|
501 |
### Filtering and results
|
502 |
-
**Note**: You can find more details about the ablations and results in the FineWeb blog post
|
503 |
|
504 |
We investigated the impact of using different thresholds for the filtering and found that threshold 3 gave the best overall results. Although using a threshold higher than 3 improves performance on knowledge and reasoning intensive benchmarks, it significantly degrades performance on HellaSwag and PIQA.
|
505 |
|
|
|
499 |
The classifier is available at: [https://huggingface.co/HuggingFaceFW/fineweb-edu-classifier/](HuggingFaceFW/fineweb-edu-classifier/)
|
500 |
|
501 |
### Filtering and results
|
502 |
+
**Note**: You can find more details about the ablations and results in the FineWeb [blog post](https://huggingface.co/spaces/HuggingFaceFW/blogpost-fineweb-v1).
|
503 |
|
504 |
We investigated the impact of using different thresholds for the filtering and found that threshold 3 gave the best overall results. Although using a threshold higher than 3 improves performance on knowledge and reasoning intensive benchmarks, it significantly degrades performance on HellaSwag and PIQA.
|
505 |
|