SocialFinanceQA / README.md
Kris-Fillip's picture
Update README.md
9478435 verified
metadata
license: mit
configs:
  - config_name: default
    data_files:
      - split: train
        path: data/train-*
      - split: test
        path: data/test-*
dataset_info:
  features:
    - name: text
      dtype: string
    - name: context
      dtype: string
    - name: answer_1
      dtype: string
    - name: answer_2
      dtype: string
    - name: subreddit
      dtype: string
    - name: num_comments
      dtype: int64
    - name: score
      dtype: int64
    - name: upvote_ratio
      dtype: float64
    - name: ups
      dtype: float64
    - name: downs
      dtype: float64
    - name: author
      dtype: string
    - name: created_utc
      dtype: int64
    - name: retrieved_on
      dtype: float64
    - name: retrieved_utc
      dtype: float64
    - name: id
      dtype: string
    - name: comment_is_submission
      dtype: bool
    - name: score_answer1
      dtype: int64
    - name: upvote_ratio_answer1
      dtype: 'null'
    - name: ups_answer1
      dtype: float64
    - name: downs_answer1
      dtype: float64
    - name: author_answer1
      dtype: string
    - name: name_answer1
      dtype: string
    - name: id_answer1
      dtype: string
    - name: created_utc_answer1
      dtype: int64
    - name: retrieved_on_answer1
      dtype: float64
    - name: retrieved_utc_answer1
      dtype: float64
    - name: score_answer2
      dtype: int64
    - name: upvote_ratio_answer2
      dtype: 'null'
    - name: ups_answer2
      dtype: float64
    - name: downs_answer2
      dtype: float64
    - name: author_answer2
      dtype: string
    - name: name_answer2
      dtype: string
    - name: id_answer2
      dtype: string
    - name: created_utc_answer2
      dtype: int64
    - name: retrieved_on_answer2
      dtype: float64
    - name: retrieved_utc_answer2
      dtype: float64
  splits:
    - name: train
      num_bytes: 116393440
      num_examples: 53061
    - name: test
      num_bytes: 1076885
      num_examples: 500
  download_size: 73381121
  dataset_size: 117470325
task_categories:
  - question-answering
language:
  - en
tags:
  - finance
size_categories:
  - 10K<n<100K

Dataset Card for Dataset Name

SocialFinanceQA is the first social-media preference dataset for fine-tuning and aligning LLMs for the finance domain.

Dataset Details

Dataset Description

  • Curated by: Kris-Fillip Kahl, Tolga Buz, Russa Biswas, Gerard de Melo
  • Language(s) (NLP): English
  • License: MIT

Dataset Sources

Uses

This dataset is intended to be used for fine-tuning and aligning LLMs for the task of Long Form Question Answering (LFQA).

Direct Use

This dataset is suitable for supervised fine-tuning as well as preference alignment of LLMs.

Dataset Structure

The dataset consists of questions from submissions and corresponding answers from comments within the selected subreddits. It is a preference dataset constructed based on the scores (upvotes and downvotes) of the individual submissions and comments within their subreddits. All fields entailing answer_1 correspond to the preferred answer for a specific question, while the answer_2 fields correspond to the non-preferred answer.

Dataset Creation

Curation Rationale

Retail investing is on the rise, and a growing number of users are relying on online finance communities to educate themselves. However, recent years have positioned Large Language Models (LLMs) as powerful question-answering (QA) tools, shifting users away from interacting in communities toward discourse with AI-driven conversational interfaces. These AI tools are currently limited by the availability of labeled data containing domain-specific financial knowledge. Therefore, in this work, we curate a QA preference dataset SOCIALFINANCEQA for fine-tuning and aligning LLMs, extracted from over 7.4 million submissions and 82 million comments from 2008 to 2022 in Reddit’s 15 largest finance communities.

Source Data

The data originally stems from Reddit submissions and comments made to 15 of the largest financial subreddits (personalfinance, financialindependence, FinancialPlanning, investing, wallstreetbets, Wallstreetbetsnew, stocks, StockMarket, pennystocks, options, RealEstate, Economics, realestateinvesting, AskEconomics, explainlikeimfive) between 2008 and 2022.

Data Collection and Processing

Who are the source data producers?

Reddit users who contributed to the selected financial subreddits between 2008 and 2022.

Personal and Sensitive Information

The dataset contains the usernames of the authors of the texts. These user names are pseudonyms that do not contain information about the author's identity unless the author actively decides to share personal information in their Reddit profile. We have not obtained dedicated permission from the authors of the texts to include them in our dataset. However, this is mitigated by the users agreeing to Reddit’s terms, which allow data extraction services such as Pushshift, which is the source for our dataset and was accessed legitimately. Additionally, the texts may contain the names of individuals, usually public figures relevant to the stock market, whom the authors included in their questions or responses. We have decided to leave these names in the dataset, as they are usually critical for understanding the text (e.g. when a company’s CEO is mentioned in a text about how the same company performs).

Bias, Risks, and Limitations

Social media should never be the only source for investors planning to make financial decisions. We have chosen 15 communities covering various topics to provide a broad foundation. Nonetheless, we recommend retail investors consult other sources, such as relevant books, scientific literature, and trained professionals, before making significant investment decisions. Investing often bears the risk of losing money, so investors should be careful with money they cannot afford to lose. Inappropriate, toxic, or offensive language is always a risk when working with social media data. We mitigate this by filtering our dataset carefully, leveraging Reddit’s internal mechanisms and additional techniques, such as toxicity detection. Nonetheless, there is a possibility that our dataset still harbors remaining instances of text that may be considered offensive or may cause fine-tuned LLMs to generate such.

Dataset Card Authors

Kris-Fillip Kahl, Tolga Buz