---
configs:
- config_name: default
  data_files:
  - split: train
    path: "data/train_instances.json"
  - split: dev
    path: "data/dev_instances.json"
  - split: test
    path: "data/test_instances.json"
- config_name: has_html
  data_files:
  - split: train
    path: "data/train_instances_with_html.json"
  - split: dev
    path: "data/dev_instances_with_html.json"
  - split: test
    path: "data/test_instances_with_html.json"
---

# Preprocessed QASPER dataset

Working doc: https://docs.google.com/document/d/1gYPhPNJ5LGttgjix1dwai8pdNcqS6PbqhsM7W0rhKNQ/edit?usp=sharing

Original:
- Dataset: https://github.com/allenai/qasper-led-baseline
- Baseline repo: https://github.com/allenai/qasper-led-baseline
- HF: https://huggingface.co/datasets/allenai/qasper


Differences of our implementation over the original implementation:
1. We use the dataset provided at https://huggingface.co/datasets/allenai/qasper since it doesn't require manually downloading files.
2. We remove usage of `allennlp` since the Python package cannot be installed anymore.
3. We add baselines to [qasper/models](qasper/models/). Currently, we have
    - QASPER (Longformer Encoder Decoder)
    - GPT-3.5-Turbo
    - TODO: RAG (with R=TF-IDF or Contriever) implemented in LangChain?
4. We replace `allennlp` special tokens with the special tokens of the HF transformer tokenizer:
    - paragraph separator: '</s>' -> tokenizer.sep_token
    - sequence pair start tokens: _tokenizer.sequence_pair_start_tokens -> tokenizer.bos_token

## Usage

```
from datasets import load_dataset

dataset = load_dataset("ag2435/qasper")
```