thaottn commited on
Commit
d54eb45
·
verified ·
1 Parent(s): 807c093

Update README.md

Browse files
Files changed (1) hide show
  1. README.md +5 -1
README.md CHANGED
@@ -11,7 +11,7 @@ size_categories:
11
  ---
12
  # Dataset Card for Recycling-The-Web Synthetic Data
13
 
14
- We release ~44B tokens of high-quality, filtered synthetic texts obtained via our REcycling the Web with guIded REwrite (REWIRE) approach.
15
  The generation process involves taking all documents that are of moderate quality (i.e., having passed some rule-based filters),
16
  using an LLM (Llama-3.3-70B-Instruct) to identify the purpose of the text content, and then asking the LLM to come up with an improved document conditioned on chain-of-thought reasoning.
17
  Our approach specifically targets the vast quantity of low-quality documents that are somewhat informative but still not considered high-quality by existing filters.
@@ -50,6 +50,9 @@ The data is intended for pre-training large language models (LLMs), and designed
50
 
51
  The data was obtained by using an LLM to rewrite low-quality documents that have been discarded by quality filters, in order to make them useful for pre-training.
52
  This helps increase the token availability and address the impending data bottleneck, as the growth of public human-generated texts has been lagging behind the increase in model capacity and training token budget.
 
 
 
53
 
54
  ### Source Data
55
 
@@ -73,6 +76,7 @@ The negative data are random samples selected from our rewriting generations.
73
 
74
  This data release contains _only_ the rewritten outputs that have been filtered, i.e. those in the top 10% of the generations based on fastText scores.
75
 
 
76
 
77
  #### Personal and Sensitive Information
78
 
 
11
  ---
12
  # Dataset Card for Recycling-The-Web Synthetic Data
13
 
14
+ We release ~44B tokens of high-quality, model-filtered synthetic texts obtained via our [REcycling the Web with guIded REwrite (REWIRE)](https://arxiv.org/abs/2506.04689) approach.
15
  The generation process involves taking all documents that are of moderate quality (i.e., having passed some rule-based filters),
16
  using an LLM (Llama-3.3-70B-Instruct) to identify the purpose of the text content, and then asking the LLM to come up with an improved document conditioned on chain-of-thought reasoning.
17
  Our approach specifically targets the vast quantity of low-quality documents that are somewhat informative but still not considered high-quality by existing filters.
 
50
 
51
  The data was obtained by using an LLM to rewrite low-quality documents that have been discarded by quality filters, in order to make them useful for pre-training.
52
  This helps increase the token availability and address the impending data bottleneck, as the growth of public human-generated texts has been lagging behind the increase in model capacity and training token budget.
53
+ Across different model scales, we find that mixing our synthetic data and high-quality web data consistently outperforms training on only the latter.
54
+
55
+ ![Summary of performance improvement](https://huggingface.co/datasets/facebook/recycling_the_web/blob/main/main_figure.png)
56
 
57
  ### Source Data
58
 
 
76
 
77
  This data release contains _only_ the rewritten outputs that have been filtered, i.e. those in the top 10% of the generations based on fastText scores.
78
 
79
+ ![REWIRE pipeline](https://huggingface.co/datasets/facebook/recycling_the_web/blob/main/REWIRE_pipeline.png)
80
 
81
  #### Personal and Sensitive Information
82