Financbase/financbase-10k-jsonl

Financbase Financial QA Dataset

Dataset Description

The Financbase Financial QA Dataset is a curated collection of financial question-answering examples designed for training large language models on financial domain tasks. This dataset supports multiple financial AI tasks including question answering, sentiment analysis, and document summarization.

Dataset Summary

• Total Examples: 1,000+ financial Q&A pairs
• Format: JSONL (JSON Lines)
• Language: English
• Domain: Financial services, SEC filings, investment analysis
• Tasks: Question answering, sentiment classification, summarization

Dataset Structure

Each example follows the instruction-tuning format with three fields:

{
  "instruction": "Answer the question clearly for a retail investor.",
  "input": "What is EBITDA?",
  "output": "EBITDA stands for Earnings Before Interest, Taxes, Depreciation, and Amortization. It's a measure of a company's operating performance that excludes non-operating expenses..."
}

Supported Tasks

1. Financial Question Answering

   • Basic financial concepts (EBITDA, P/E ratio, etc.)
   • Investment terminology
   • Market analysis questions

2. Sentiment Analysis

   • Financial news sentiment classification
   • Earnings report sentiment
   • Market outlook analysis

3. Document Summarization

   • SEC filing summaries
   • Earnings call summaries
   • Financial report abstracts

Usage

Loading the Dataset

from datasets import load_dataset

# Load the dataset
dataset = load_dataset("Financbase/financbase-10k-jsonl", split="train")

# Access examples
for example in dataset:
    print(f"Instruction: {example['instruction']}")
    print(f"Input: {example['input']}")
    print(f"Output: {example['output']}")

Training with Transformers

from transformers import AutoTokenizer, AutoModelForCausalLM
from datasets import load_dataset

# Load dataset
dataset = load_dataset("Financbase/financbase-10k-jsonl", split="train")

# Format for training
def format_example(example):
    return f"### Instruction:\n{example['instruction']}\n\n### Input:\n{example['input']}\n\n### Response:\n{example['output']}"

# Apply formatting
formatted_dataset = dataset.map(lambda x: {"text": format_example(x)})

Using with PEFT/LoRA

from peft import LoraConfig, get_peft_model
from transformers import AutoModelForCausalLM

# Load base model
model = AutoModelForCausalLM.from_pretrained("mistralai/Mistral-7B-Instruct-v0.2")

# Configure LoRA
lora_config = LoraConfig(
    r=16,
    lora_alpha=32,
    target_modules=["q_proj", "k_proj", "v_proj", "o_proj"],
    lora_dropout=0.05,
    bias="none",
    task_type="CAUSAL_LM"
)

# Apply LoRA
model = get_peft_model(model, lora_config)

Data Fields

Field	Type	Description
instruction	string	The task instruction or prompt
input	string	The input context or question
output	string	The expected response or answer

Data Splits

• train: 1,000+ examples for training
• validation: 100+ examples for validation (future release)
• test: 100+ examples for testing (future release)

Data Collection

Sources

• SEC 10-K filings (processed and chunked)
• Financial news articles
• Investment research reports
• Financial education materials
• Curated financial Q&A pairs

Preprocessing

1. Document Chunking: Long documents split into ≤1800 token chunks
2. Section Preservation: Maintains document structure and headings
3. Quality Filtering: Removes low-quality or irrelevant examples
4. Format Standardization: Ensures consistent instruction/input/output format

Compliance and Safety

Financial Compliance

• No Investment Advice: Dataset does not contain personalized investment recommendations
• Educational Purpose: Designed for educational and research use
• Source Attribution: All examples traceable to original sources
• Regulatory Compliance: Follows financial data handling best practices

Content Filtering

• Removed personally identifiable information (PII)
• Filtered out actionable trading directives
• Excluded copyrighted material
• Sanitized sensitive financial data

Evaluation

Metrics

• Perplexity: Model confidence on financial text
• BLEU Score: Response quality for summarization tasks
• Accuracy: Classification accuracy for sentiment analysis
• ROUGE Score: Summarization quality metrics

Benchmark Tasks

1. Financial QA: Answer financial questions accurately
2. Sentiment Analysis: Classify financial sentiment (positive/negative/neutral)
3. Summarization: Summarize financial documents concisely

Limitations

• Language: English only
• Domain: Primarily US financial markets
• Temporal: Data from 2020-2024 (may become outdated)
• Bias: Reflects training data biases and limitations

Citation

@dataset{financbase_financial_qa_2024,
  title={Financbase Financial QA Dataset},
  author={Financbase Team},
  year={2024},
  url={https://huggingface.co/datasets/Financbase/financbase-10k-jsonl},
  license={MIT}
}

License

This dataset is released under the MIT License. See LICENSE file for details.

Contact

• Organization: Financbase
• Repository: https://huggingface.co/datasets/Financbase/financbase-10k-jsonl
• Issues: Report issues via HuggingFace Hub

Changelog

• v0.1 (2024-12-19): Initial release with 1,000+ financial Q&A examples
• v0.2 (Planned): Add validation and test splits
• v0.3 (Planned): Expand to 10,000+ examples with more diverse sources