Master Data Set:

   Built a bias classification dataset with 7 labels: IDENTITY_ABUSE, STEREOTYPE_NEGATIVE, 

   STEREOTYPE_BENEVOLENT, SYSTEMIC_BIAS, HATE_SPEECH, TOXICITY_GENERAL, and NO_BIAS.

​

Data Sources Used:
   Successfully Integrated (~83K samples from real datasets):
   Jigsaw/Civil Comments: ~38K samples (toxicity detection)
   Social Bias Frames: ~27K samples (offensive content classification)
   HatExplain: ~14K samples (hate speech with explanations)
   StereoSet: ~2K samples (stereotype detection)
   CrowS-Pairs: ~1.5K samples (stereotype pairs)

​

   remaining data, synthetically generated.

​

Data Formatting for Training: 

   Single-label structure: Samples have one bias label or no bias
   Ratio: 79,528 training / 20,472 validation samples

   JSONL format: Each line contains {"text": "...", "target": ["LABEL1", "LABEL2"], "source": "..."}

 

Training Rationale:
   Fine-tuned DistilBERT model: Smaller and faster than BERT but has near identical performance for text

   classification.
   Conservative settings: Small batch (8), few epochs (3), standard learning rate (5e-05) to avoid overfitting 
   Dataset has 7 overlapping labels from diverse sources, and figured choices prevent the model from

   memorizing patterns.

 

Model Wins:

   50% greater speed then fastest non-thinking models
      - Using Macbook m2 to run (with pipelining)
      - Gemini 2.5 flash
      - Chatgpt 5 (non-thinking)

   77x more power efficient than LLM’s
      - Conservative estimate of ~1.0kWh per 1 million tokens for AI
      - Approximately 0.012 kWh average per million tokens after running 100 tests on Macbook m2 (with

        pipelining)

   98.8% Cheaper to run than LLM’s
      - Estimates from previously
      - Assumes average U.S electricity rate of $0.15 per kWh

​

   accuracy on validation set: 0.8669402110199297

   loss during training: 0.3745632469654083