Datasets:

Modalities:
Text
Formats:
json
Languages:
English
ArXiv:
Libraries:
Datasets
pandas
License:
jchevallard commited on
Commit
8e196b5
·
1 Parent(s): df57a76

feat: updated README

Browse files
Files changed (1) hide show
  1. README.md +35 -11
README.md CHANGED
@@ -1,15 +1,15 @@
1
  ---
2
  configs:
3
- - config_name: crag_task_1_and_2_subsample_1
4
- data_files: crag_task_1_and_2/crag_task_1_and_2_dev_v4_subsample_cdf90925.jsonl.bz2
5
- - config_name: crag_task_1_and_2_subsample_2
6
- data_files: crag_task_1_and_2/crag_task_1_and_2_dev_v4_subsample_2e617280.jsonl.bz2
7
- - config_name: crag_task_1_and_2_subsample_3
8
- data_files: crag_task_1_and_2/crag_task_1_and_2_dev_v4_subsample_9bca9871.jsonl.bz2
9
- - config_name: crag_task_1_and_2_subsample_4
10
- data_files: crag_task_1_and_2/crag_task_1_and_2_dev_v4_subsample_ff5aa423.jsonl.bz2
11
- - config_name: crag_task_1_and_2_subsample_5
12
- data_files: crag_task_1_and_2/crag_task_1_and_2_dev_v4_subsample_3f4353ba.jsonl.bz2
13
  license: cc-by-nc-4.0
14
  task_categories:
15
  - question-answering
@@ -26,6 +26,30 @@ pretty_name: CRA
26
 
27
  Datasets are taken from Facebook's [CRAG: Comprehensive RAG Benchmark](https://github.com/facebookresearch/CRAG), see their [arXiv paper](https://arxiv.org/abs/2406.04744) for details about the dataset construction.
28
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
29
  CRAG (Comprehensive RAG Benchmark) is a rich and comprehensive factual question answering benchmark designed to advance research in RAG. The public version of the dataset includes:
30
  - 2706 Question-Answer pairs
31
  - 5 domains: Finance, Sports, Music, Movie, and Open domain
@@ -34,7 +58,7 @@ CRAG (Comprehensive RAG Benchmark) is a rich and comprehensive factual question
34
  ![](figs/CRAG_table_2.png)
35
 
36
  The datasets `crag_task_1_and_2_dev_v4_subsample_*.json.bz2` have been created from the dataset [crag_task_1_and_2_dev_v4.jsonl.bz2](https://github.com/facebookresearch/CRAG/raw/refs/heads/main/data/crag_task_1_and_2_dev_v4.jsonl.bz2?download=) available on CRAG's GitHub repository.
37
- For an easier handling and download of the dataset, we have used the script [crag_to_subsamples.py](https://huggingface.co/datasets/Quivr/crag/blob/main/crag_to_subsamples.py) to split the 2706 rows of the original file in 5 subsamples, following the procedure below:
38
  1. We have created a new label `answer_type`, classifying the answers in 3 categories:
39
  - `invalid` for any answer == "invalid question"
40
  - `no_answer` for any answer == "i don't know"
 
1
  ---
2
  configs:
3
+ - config_name: crag_task_1_and_2_subset_1
4
+ data_files: crag_task_1_and_2/crag_task_1_and_2_dev_v4_subset_cdf90925.jsonl.bz2
5
+ - config_name: crag_task_1_and_2_subset_2
6
+ data_files: crag_task_1_and_2/crag_task_1_and_2_dev_v4_subset_2e617280.jsonl.bz2
7
+ - config_name: crag_task_1_and_2_subset_3
8
+ data_files: crag_task_1_and_2/crag_task_1_and_2_dev_v4_subset_9bca9871.jsonl.bz2
9
+ - config_name: crag_task_1_and_2_subset_4
10
+ data_files: crag_task_1_and_2/crag_task_1_and_2_dev_v4_subset_ff5aa423.jsonl.bz2
11
+ - config_name: crag_task_1_and_2_subset_5
12
+ data_files: crag_task_1_and_2/crag_task_1_and_2_dev_v4_subset_3f4353ba.jsonl.bz2
13
  license: cc-by-nc-4.0
14
  task_categories:
15
  - question-answering
 
26
 
27
  Datasets are taken from Facebook's [CRAG: Comprehensive RAG Benchmark](https://github.com/facebookresearch/CRAG), see their [arXiv paper](https://arxiv.org/abs/2406.04744) for details about the dataset construction.
28
 
29
+ # CRAG Sampler
30
+
31
+ We have added a simple Python tool for performing stratified sampling on CRAG data.
32
+
33
+ ## Installation
34
+
35
+ ### Local Development Install (Recommended)
36
+ ```bash
37
+ git clone https://huggingface.co/Quivr/CRAG.git
38
+ cd CRAG
39
+ pip install -r requirements.txt # Install dependencies
40
+ pip install -e . # Install package in development mode
41
+ ```
42
+
43
+ ## Quick Start
44
+
45
+ ### Running the example
46
+
47
+ ```bash
48
+ python -m examples.basic_sampling
49
+ ```
50
+
51
+ # CRAG dataset
52
+
53
  CRAG (Comprehensive RAG Benchmark) is a rich and comprehensive factual question answering benchmark designed to advance research in RAG. The public version of the dataset includes:
54
  - 2706 Question-Answer pairs
55
  - 5 domains: Finance, Sports, Music, Movie, and Open domain
 
58
  ![](figs/CRAG_table_2.png)
59
 
60
  The datasets `crag_task_1_and_2_dev_v4_subsample_*.json.bz2` have been created from the dataset [crag_task_1_and_2_dev_v4.jsonl.bz2](https://github.com/facebookresearch/CRAG/raw/refs/heads/main/data/crag_task_1_and_2_dev_v4.jsonl.bz2?download=) available on CRAG's GitHub repository.
61
+ For an easier handling and download of the dataset, we have used our CRAG sampler to split the 2706 rows of the original file in 5 subsamples, following the procedure below:
62
  1. We have created a new label `answer_type`, classifying the answers in 3 categories:
63
  - `invalid` for any answer == "invalid question"
64
  - `no_answer` for any answer == "i don't know"