metadata
dataset_info:
features:
- name: question
dtype: string
- name: answers
sequence: string
- name: positive_ctxs
struct:
- name: passage_id
sequence: string
- name: text
sequence: string
- name: title
sequence: string
- name: negative_ctxs
struct:
- name: passage_id
sequence: 'null'
- name: text
sequence: 'null'
- name: title
sequence: 'null'
- name: hard_negative_ctxs
struct:
- name: passage_id
sequence: string
- name: text
sequence: string
- name: title
sequence: string
- name: easy_negative_ctxt
struct:
- name: passage_id
sequence: string
- name: text
sequence: string
- name: title
sequence: string
splits:
- name: train
num_bytes: 75812
num_examples: 10
- name: test
num_bytes: 37906
num_examples: 5
download_size: 138099
dataset_size: 113718
configs:
- config_name: default
data_files:
- split: train
path: data/train-*
- split: test
path: data/test-*
task_categories:
- question-answering
language:
- de
Original Dataset
The original dataset, deepset/germandpr, contains:
- 9275 training examples
- 1025 testing examples
Each example is a question/answer pair, consisting of:
- One question
- One answer
- One positive context
- Three negative contexts
You can find the original dataset here.
Modifications
Adding Easy Negative Examples
To enhance the dataset, an "easy negative example" was added to each row. The objective of this addition is to train the model to better distinguish between relevant and irrelevant contexts by exposing it to plausible but incorrect information.
After adding easy neagtives each example in the dataset consists of :
- One question
- One answer
- One positive context
- Three negative contexts
- Easy negative contexts
Method
For identifying easy negative examples, I have used utilized the L2 distance metric from Faiss to find the most dissimilar index (vector) relative to the positive context in each row. This dissimilar index was then selected as the easy negative example.