pQube commited on
Commit
d0ebe3c
·
verified ·
1 Parent(s): e2ce0f3

Update README.md

Browse files
Files changed (1) hide show
  1. README.md +16 -11
README.md CHANGED
@@ -29,23 +29,28 @@ size_categories:
29
 
30
 
31
  - **Creators**: Intel Labs
32
- - **Version**: <TODO - REPLACE WITH VERSION>
33
  - **License**: [Intel Research and Development License](.\Intel_OBL_Internal_RD_Use_License.md)
34
- - **Number of Training Samples**: <TODO>
35
- - **Number of Test Samples**: <TODO>
36
- - <TODO - Add additional as needed>
37
- - **Format**: <TODO - IF Applicable>
38
 
39
  ## Intended Use
40
 
41
- - **Primary Uses**: <TODO>
42
- - **Out-of-Scope Uses**: <TODO>
 
 
43
 
44
  ## Data Collection Process
45
- <!-- TODO
46
- - Provide detailed explaination of how the dataset was generated/create/collected.
47
- - If 3rd party datasets were used/included, provide a link to the original dataset as well as the license the data is released under.
48
- -->
 
 
 
49
 
50
  ## Ethical Considerations
51
  <!-- DON'T CHANGE THIS SECTION -->
 
29
 
30
 
31
  - **Creators**: Intel Labs
32
+ - **Version**: 1.0 (Updated: 2025-05-02)
33
  - **License**: [Intel Research and Development License](.\Intel_OBL_Internal_RD_Use_License.md)
34
+ - **Number of Training Samples**: >100M
35
+ - **Number of Test Samples**: >500K
36
+
37
+ - **Format**: pyarrow format
38
 
39
  ## Intended Use
40
 
41
+ - **Primary Uses**:
42
+ - These intermediate features or tokens are primarily intended for insights and exploratory analysis in explainable AI, as well as for training ad-hoc models such as linear probes, Sparse AutoEncoders (SAEs), or transcoders.
43
+ - This is also an example of a dataset that can be extracted using the feature recorder (intercept_manager) from the BlueGlass repository, providing a guiding dataset for further analysis.
44
+ - **Out-of-Scope Uses**: This dataset is not intended for commercial use or for training models that will be deployed in real-world scenarios without further verification and validation.
45
 
46
  ## Data Collection Process
47
+
48
+ This dataset contains intermediate features extracted from various layers of transformer models using the intercept_manager module from [BlueGlass](https://github.com/IntelLabs/blueglass). The features are recorded from different probe positions within the model, as illustrated in the image below, enabling fine-grained analysis and interoperability.
49
+
50
+ <div align="center">
51
+ <img src="docs/images/feature_pattern.png" alt="Feature Pattern used to extract the BlueLens dataset" width="300"/>
52
+ </div>
53
+
54
 
55
  ## Ethical Considerations
56
  <!-- DON'T CHANGE THIS SECTION -->