Inspiration
We were fascinated by how social media drives emotional reactions in the cryptocurrency world. With thousands of tweets flying in every minute, it's nearly impossible for a human to process them all. We wanted to explore how AI could help us "read the crowd" — not just in terms of market trends, but in understanding how people feel. The idea of combining GPT-2 with our own prompts and reflections felt like the perfect hands-on approach to collaborate with AI.
What it does
CryptoMind analyzes tweets related to cryptocurrency and uses GPT-2 to detect both sentiment (positive, negative, neutral) and emotion (like fear, excitement, anger, etc.). Instead of relying on one-shot AI results, we iteratively crafted prompts, reviewed responses, and adjusted the process to make the model more accurate and human-aligned.
How we built it
We collected real-time tweets using the Twitter API filtered by crypto-related hashtags.
Used GPT-2 through the Hugging Face Transformers library.
Created a prompt template to guide GPT-2 to classify sentiment and emotion.
Built a small interface to enter prompts, view AI responses, and allow for human feedback and validation.
Repeated prompt refinement cycles to improve accuracy and reliability.
Challenges we ran into
GPT-2 sometimes misunderstood sarcasm or slang — which is common in crypto tweets.
Emotional overlap in tweets (e.g., excitement + fear) made classification tricky.
Designing prompts that were clear yet flexible enough for varied tweet structures.
Finding the balance between automation and human intervention.
Accomplishments that we're proud of
Developed a working system that genuinely collaborates with AI.
Saw real improvements with each prompt adjustment.
Understood how a model like GPT-2 "thinks" and responds to different prompt structures.
Gained deeper insights into both AI limitations and its creative possibilities.
What we learned
Prompt engineering is powerful — small changes in wording can totally shift AI behavior.
AI isn’t magic — it requires guidance, reflection, and curation.
Human judgment is essential when it comes to interpreting emotion.
AI models carry inherent biases based on their training data — awareness is key.
What's next for CryptoMind: AI + Human Insight for Twitter Sentiment
Add a dashboard to visualize emotions over time and by topic.
Integrate fine-tuning on labeled tweet datasets for better accuracy.
Expand to other domains like stock market tweets or political discourse.
Experiment with GPT-3.5 or GPT-4 for more nuanced language understanding.
Explore multilingual sentiment detection using the same framework.
Log in or sign up for Devpost to join the conversation.