Files changed (1) hide show
  1. README.md +1 -95
README.md CHANGED
@@ -1,95 +1 @@
1
- ---
2
- language:
3
- - en
4
- license: apache-2.0
5
- ---
6
-
7
- # FineWeb-Edu classifier
8
-
9
- ## Model summary
10
- This is a classifier for judging the educational value of web pages. It was developed to filter and curate educational content from web datasets and was trained on 450k [annotations](https://huggingface.co/datasets/HuggingFaceFW/fineweb-edu-llama3-annotations) generated by [LLama3-70B-instruct](https://huggingface.co/meta-llama/Meta-Llama-3-70B-Instruct) for web samples from [FineWeb](https://huggingface.co/datasets/HuggingFaceFW/fineweb) dataset.
11
-
12
- We used this classifier to build [FineWeb-Edu](https://huggingface.co/datasets/HuggingFaceFW/fineweb-edu) dataset.
13
- ### How to use in transformers
14
- To load the FineWeb-Edu classifier, use the following code:
15
-
16
- ```python
17
- from transformers import AutoTokenizer, AutoModelForSequenceClassification
18
-
19
- tokenizer = AutoTokenizer.from_pretrained("HuggingFaceTB/fineweb-edu-classifier")
20
- model = AutoModelForSequenceClassification.from_pretrained("HuggingFaceTB/fineweb-edu-classifier")
21
-
22
- text = "This is a test sentence."
23
- inputs = tokenizer(text, return_tensors="pt", padding="longest", truncation=True)
24
- outputs = model(**inputs)
25
- logits = outputs.logits.squeeze(-1).float().detach().numpy()
26
- score = logits.item()
27
- result = {
28
- "text": text,
29
- "score": score,
30
- "int_score": int(round(max(0, min(score, 5)))),
31
- }
32
-
33
- print(result)
34
- # {'text': 'This is a test sentence.', 'score': 0.07964489609003067, 'int_score': 0}
35
- ```
36
-
37
- ## Training
38
- The classifier was trained on 450,000 pairs of web samples and their scores from 0 to 5, generated by Llama3. The samples were annotated based on their educational quality with 0 being not educational and 5 being highly educational.
39
-
40
- Below is the prompt used for LLama3 annotations:
41
- <div style="text-align: center; margin: 20px 0;">
42
- <img src="https://cdn-uploads.huggingface.co/production/uploads/61c141342aac764ce1654e43/fjZQ4izIj1rx1xQnBTKKr.png" alt="Prompt for LLM annotation" style="width: 90%; max-width: 800px; height: auto;">
43
- </div>
44
-
45
- We added a classification head with a single regression output to [Snowflake-arctic-embed](https://huggingface.co/Snowflake/snowflake-arctic-embed-m) and trained the model for 20 epochs with a learning rate of 3e-4. During training, the embedding and encoder layers were frozen to focus on the classification head. The model achieved an F1 score of 82% when converted to a binary classifier using a score threshold of 3.
46
-
47
- **Training Details:**
48
-
49
- - Model: Snowflake-arctic-embed with a classification head
50
- - Dataset: 450,000 samples from Llama3 annotations
51
- - Epochs: 20
52
- - Learning Rate: 3e-4
53
- - Evaluation Metric: F1 score
54
-
55
- **Classification report**
56
-
57
- We treat the regression model's predictions as discrete classes to calculate the metrics on a hold-out set of 46867 Llama3-annotated samples.
58
- ```
59
- precision recall f1-score support
60
-
61
- 0 0.75 0.49 0.59 5694
62
- 1 0.78 0.84 0.81 26512
63
- 2 0.57 0.61 0.59 10322
64
- 3 0.56 0.50 0.53 3407
65
- 4 0.58 0.35 0.44 807
66
- 5 0.33 0.01 0.02 125
67
-
68
- accuracy 0.71 46867
69
- macro avg 0.60 0.47 0.50 46867
70
- weighted avg 0.71 0.71 0.71 46867
71
- ```
72
-
73
- **Confusion matrix**
74
-
75
- We verify that the predicted educational scores are indeed close to their ground truth, and are mostry impacted by the noisy annotation.
76
- ```
77
- 2791 2858 45 0 0 0
78
- 919 22343 3180 69 1 0
79
- y_true 3 3225 6330 757 7 0
80
- 1 66 1473 1694 173 0
81
- 0 4 98 420 283 2
82
- 0 0 18 85 21 1
83
- y_pred
84
- ```
85
-
86
-
87
- ## Limitations
88
- While the FineWeb-Edu classifier performs well in distinguishing high-quality educational content for FineWeb dataset, there are some limitations:
89
-
90
- - Scope: The model's performance might change for other datasets, in particular for out of distribution samples. It is also focused on educational content relevant to primary and grade school levels and may not perform as well on content intended for higher education or specialized domains.
91
- - Bias: The model's performance is dependent on the quality and representativeness of the training data and the LLM used for the annotation. Biases in both can affect the classifier's judgments. It might overfit to academic looking content for the higher scores and we recommend using int_score >= 3 as a threshold for data curation.
92
- - Context: The classifier evaluates individual web pages or extracts without considering broader context, which might impact its effectiveness in certain scenarios.
93
-
94
- The training and inference code is available on GitHub
95
- https://github.com/huggingface/cosmopedia/tree/main/classification
 
1
+ 7894561230