Multi-Label Effects Classification of Psychoactive Drugs

A transformer-based multi-label text classification model that predicts the subjective psychological effects of psychoactive substances from user trip reports. Due to a smaller dataset, we achieved 55% accuracy.

---

Example Prediction

• Input: I saw impossible geometry and felt a strong sense of ego death while colors melted around me.
• Predicted Effects: ['ego death', 'geometry', 'visual distortion', 'unity', 'time distortion']

Source & Data

• Source: PsychonautWiki
• Scraping: Custom Scrapy crawler
• Reports: 231 unique trip reports
• Labels: 19 subjective effect categories
• Split: 80/20 train-test split

---

Preprocessing & Encoding

• Cleaned raw text data and filtered infrequent effects.
• Converted effect labels into multi-hot encoded vectors.
• Label mapping stored in: effect_types_encoded.json

---

Model Details

• Model: allenai/longformer-base-4096
• Frameworks: FastAI + Blurr (HuggingFace integration)
• Task: Multi-label classification
• Loss: BCEWithLogitsLoss
• Threshold: 0.2 for effect selection

---

Files & Resources

All essential artifacts (model, dataloader, label mappings) are hosted here:
🔗 Google Drive Link

• effect-classifier-stage-1.pkl – trained model
• dls-effect-classifier.pkl – fastai DataLoaders
• effect_types_encoded.json – label map

---