---
license: apache-2.0
configs:
- config_name: default
  data_files:
  - split: train
    path: 
    - "ott-qa/train.json"
    - "phrase-bank/train.json"
    - "tat-qa/train.json"
    - "template/train.json"
    - "wiki-sql/train.json"
  - split: test
    path:
    - "ott-qa/test.json"
    - "phrase-bank/test.json"
    - "tat-qa/test.json"
    - "template/test.json"
    - "wiki-sql/test.json"
  - split: val
    path:
    - "ott-qa/dev.json"
    - "phrase-bank/dev.json"
    - "tat-qa/dev.json"
    - "template/dev.json"
    - "wiki-sql/dev.json"
---


## Raven Dataset
The dataset that we use to fine-tune Raven is composed from four distinct question-answering datasets. Two are specifically from the financial domain with the remaining being generic and incorporating questions over both tables and text.


### TAT-QA. 

[Table-and-Text Question](https://paperswithcode.com/dataset/tat-qa) Answering consists of 16,552 questions generated by financial experts associated with 2,757 hybrid contexts drawn from real-world financial reports.


### Financial PhraseBank 
[Financial PhraseBank](https://paperswithcode.com/sota/sentiment-analysis-on-financial-phrasebank) consists of 4,846 phrases derived from English financial news on listed companies in OMX Helsinki. The dataset contains phrase-level annotation by financial markets experts, that categorises each sample sentence exclusively from an investor's standpoint as either positive, negative, or neutral. 


### Wiki-SQL 

[Wiki-SQL](https://paperswithcode.com/dataset/wikisql) consists of 80,654 manually annotated crowd sourced examples of natural language questions and corresponding SQL queries over 24,241 tables found on Wikipedia

### OTT-QA 

Similar to \textsc{TAT-QA}, [Open Table-and-Text Question Answering](https://paperswithcode.com/dataset/ott-qa) consists of 43,683 questions over tabular data and unstructured text across diverse domains. The majority of questions necessitate multi-hop inference involving both forms of data. 

### Data preparation
The datasets described above have diverse formats and are not suited for fine-tuning Raven as-is. We employ a data conversion pipeline to convert these four datasets into a homogeneous dataset suitable to fine-tune our financial model. In general, we extract up to four key attributes from the original datasets. These are (1) instruction that describes the task to perform, for example, *Determine the sentiment of the following phrase*, or the question *What is the percentage change in revenue after the adoption of ASC 606?* (2) input that provides more context such as the phrase to classify or a passage, (3) data that accompanies the context, in tabular format, (4) derivation that produces the answer or expected response. Refer to [templates](https://github.com/adriantheuma/llama2-raven/blob/main/templates/README.md) for examples of the full prompt.

To obtain a balanced dataset we randomly sub-sample larger datasets such that we obtain a uniformly distributed dataset among the different sources. The size of the final training dataset is 47.6K samples, validation 5.26K and test 5.81K.