tahamajs/bitcoin-llm-finetuning-dataset

Bitcoin Price Prediction Instruction-Tuning Dataset

Dataset Description

This is a comprehensive, instruction-based dataset formatted specifically for fine-tuning Large Language Models (LLMs) to perform financial forecasting. The dataset is structured as a collection of prompt-response pairs, where each record challenges the model to predict the next 10 days of Bitcoin's price based on a rich, multimodal snapshot of daily data.

The dataset covers the period from early 2018 to mid-2024, with each of the 2,312 examples representing a single day. The goal is to train models that can synthesize financial, social, and on-chain data to act as specialized cryptocurrency analysts.

Dataset Structure

The dataset consists of a single split (train) and follows a standard instruction-tuning format with three columns: instruction, input, and output.

Data Fields

• instruction (string): This is the main prompt provided to the LLM. It contains a comprehensive, human-readable summary of all available data for a specific day, framed as a request for a financial forecast. It includes:

  • A directive defining the role of an "expert financial analyst".
  • The complete historical closing prices of Bitcoin for the past 60 days.
  • Daily news and social media text from Twitter, Reddit, and news sites.
  • Macroeconomic indicators like the daily closing prices of Gold and Oil.
  • Fundamental Bitcoin on-chain metrics (hash rate, difficulty, transaction volume, active addresses, etc.).
  • Social sentiment indicators like the Fear & Greed Index.
  • AI-generated sentiment and trading action classifications from another LLM.

• input (string): This field is intentionally left empty ("") for all records. This structures the task as a "zero-shot" instruction problem, where all necessary information is contained within the instruction prompt itself.

• output (string): This is the target response for the LLM to learn. It contains a string representation of a list of the next 10 days of Bitcoin's closing prices.

How the Data was Created

This instruction dataset was derived from a daily time-series dataset built by aggregating multiple sources:

• Prices & Commodities: Real-time financial data for BTC-USD, Gold (GC=F), and Crude Oil (CL=F) was sourced from Yahoo Finance via the yfinance library.
• On-Chain, Social, & LLM Metrics: A wide range of features, including on-chain data, sentiment indices, and additional text sources, were sourced from the danilocorsi/LLMs-Sentiment-Augmented-Bitcoin-Dataset.
• Primary News Source: Daily news articles were sourced from the edaschau/bitcoin_news dataset.
• Primary Tweet Source: Daily tweets were sourced from a user-provided mbsa.csv file originating from a larger Kaggle dataset.

Each row from the time-series data was programmatically serialized into the detailed instruction and output format described above.

Use Cases

The primary use case for this dataset is to fine-tune instruction-following Large Language Models (e.g., Llama, Mistral, Gemini, Gemma) to specialize in financial forecasting. By training on this data, a model can learn to:

• Ingest and reason about complex, multimodal data (text, time-series, numerical).
• Identify patterns between news sentiment, on-chain activity, and price movements.
• Generate structured, numerical forecasts in a specific format.

This dataset can also be used for research in prompt engineering and for evaluating the quantitative reasoning abilities of LLMs in the financial domain.