Update README.md

README.md

@@ -486,32 +486,28 @@ The highly popular Phi3 models were trained on 3.3 and 4.8 trillion tokens, with
 ### Annotation
 We used [Llama3-70B-Instruct](https://huggingface.co/meta-llama/Meta-Llama-3-70B-Instruct) to score 500k FineWeb samples for their educational quality on a scale from 0 to 5.
 
-We explored various prompts and found that the additive scale by [Yuan et al.](https://arxiv.org/pdf/2401.10020) worked best. To avoid the LLM favoring highly technical pages like arXiv abstracts and submissions, we focused on grade-school and middle-school level knowledge. By setting a threshold of 3 (on a scale of 0 to 5) during the filtering process, we were able to also retain some high-level educational pages. The final prompt can be found in this blog post TODO.
+We explored various prompts and found that the additive scale by [Yuan et al.](https://arxiv.org/pdf/2401.10020) worked best. To avoid the LLM favoring highly technical pages like arXiv abstracts and submissions, we focused on grade-school and middle-school level knowledge. By setting a threshold of 3 (on a scale of 0 to 5) during the filtering process, we were able to also retain some high-level educational pages. The final prompt can be found [here](https://huggingface.co/HuggingFaceFW/fineweb-edu-classifier/blob/main/utils/prompt.txt).
 
-We also experimented with different LLMs: Llama3-70B-Instruct, Mixtral-8x-7B-Instruct, and Mixtral-8x22B-Instruct. Llama3 and Mixtral-8x22B produced similar scores, while Mixtral-8x7B tended to be more generous, not fully adhering to the score scale. Verga et al. suggest using multiple LLMs as juries. We tried averaging the scores from the three models, but this shifted the distribution to the right due to the higher scores from Mixtral-8x7B. Training on a dataset filtered with a classifier using jury annotations performed worse than using a classifier based on Llama3 annotations. We hypothesize that the jury-based approach retains more low-quality samples.
+We also experimented with different LLMs: Llama3-70B-Instruct, Mixtral-8x-7B-Instruct, and Mixtral-8x22B-Instruct. Llama 3 and Mixtral-8x22B produced similar scores, while Mixtral-8x7B tended to be more generous, not fully adhering to the score scale. Verga et al. suggest using multiple LLMs as juries. We tried averaging the scores from the three models, but this shifted the distribution to the right due to the higher scores from Mixtral-8x7B. Training on a dataset filtered with a classifier using jury annotations performed worse than using a classifier based on Llama3 annotations. We hypothesize that the jury-based approach retains more low-quality samples.
 
 ### Classifier training
 We fine-tuned a Bert-like regression model using these annotations, based on [Snowflake-arctic-embed](https://huggingface.co/Snowflake/snowflake-arctic-embed-m). When converted to a binary classification  using a score of 3 as a threshold for keeping and removing files, the model achieved an F1 score of 82%. The classification of FineWeb 15T tokens took 6k H100 GPU hours.
 
-The classifier is available at: [https://huggingface.co/HuggingFaceFW/fineweb-edu-classifier/ ](https://huggingface.co/HuggingFaceFW/fineweb-edu-classifier/)
+The classifier is available at: [https://huggingface.co/HuggingFaceFW/fineweb-edu-classifier/](HuggingFaceFW/fineweb-edu-classifier/)
 
-### Filtering
-We filtered out samples with scores lower than 3. This removed 92% of the dataset, leaving us with 1.2T educational tokens. Our ablation demonstrated that this refined dataset significantly outperforms the original FineWeb dumps and even the best dump, FineWeb-2024-10. To retain more tokens, we also experimented with a less strict threshold of 2 instead of 3. This approach preserved 4.5T tokens and still outperformed the non-filtered dataset.
-TODO: add ablation results
- 
-We release these two dataset as [FineWeb-Edu](https://huggingface.co/datasets/HuggingFaceFW/fineweb-edu) and [FineWeb-Edu-score-2](https://huggingface.co/datasets/HuggingFaceFW/fineweb-edu-score-2) along with the classifier.
-
-## Dataset performance evaluation and ablations
+### Filtering and results
+**Note**: You can find more details about the ablations and results in the FineWeb blog post (TODO).
 
-We conducted our dataset performance ablations and evaluations by training 1.8B parameters models on 28B tokens and then 350B billion tokens to validate the results.
+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.
 
-The detailed configurations for training the models can be found here (TODO).
-
-FineWeb-Edu outperforms FineWeb and other web datasets on all popular benchmarks.
+We then built 📚 FineWeb-Edu by filtering out samples with scores lower than 3. This removed 92% of the dataset, leaving us with 1.3T educational tokens. Our ablation demonstrated that this refined dataset surpasses 🍷 FineWeb and all other open web datasets, with remarkable improvements on educational benchmarks such as MMLU, ARC, and OpenBookQA. To retain more tokens, we also experimented with a less strict threshold of 2 instead of 3. While being less performant than using threshold 3, it still outperformed FineWeb and it preserved 5.4T tokens.
+The plot below compare FineWeb-Edu to other web datasets:
+ 
+![image/png](https://cdn-uploads.huggingface.co/production/uploads/61c141342aac764ce1654e43/hJlyTgDzZpYuxO9LUm0PF.png)
 
-TODO: add barplots & agg_score curves
+We release these two dataset as [FineWeb-Edu](https://huggingface.co/datasets/HuggingFaceFW/fineweb-edu) and [FineWeb-Edu-score-2](https://huggingface.co/datasets/HuggingFaceFW/fineweb-edu-score-2) along with the [classifier](https://huggingface.co/HuggingFaceFW/fineweb-edu-classifier).
 
-You will find these models on [this collection](https://huggingface.co/collections/HuggingFaceFW/ablation-models-662457b0d213e8c14fe47f32). The FineWeb-Edu ablation model (trained on 350B tokens) is available at [https://huggingface.co/HuggingFaceFW/ablation-model-fineweb-edu](https://huggingface.co/HuggingFaceFW/ablation-model-fineweb-edu).
+You will find all the ablation models in [this collection](https://huggingface.co/collections/HuggingFaceFW/ablation-models-662457b0d213e8c14fe47f32). The FineWeb-Edu ablation model (trained on 350B tokens) is available at [https://huggingface.co/HuggingFaceFW/ablation-model-fineweb-edu](https://huggingface.co/HuggingFaceFW/ablation-model-fineweb-edu).
 
 ## Considerations for Using the Data
 This section is copied from the parent dataset: [FineWeb](https://huggingface.co/datasets/HuggingFaceFW/fineweb).