Overview

The wisent-ai/Qwen3-8B-TruthfulQA-TITAN is an 8 billion parameter Qwen3 model developed by Wisent AI, specifically engineered to improve truthfulness and reduce factual inaccuracies. It integrates the novel TITAN (Total Integrated Targeted Activation Navigation) steering method, which employs manifold-based directions, dynamic gating, and intensity prediction to adjust steering strength based on input content. This model was fine-tuned on a custom TruthfulQA dataset, focusing on 50 contrastive pairs across specific layers to address common misconceptions.

Key Capabilities

• Enhanced Truthfulness: Significantly reduces common factual errors and hallucinations, as demonstrated by its performance on TruthfulQA examples where it correctly identifies false statements.
• Dynamic Steering: Utilizes a neural network to dynamically decide the steering intensity per input, offering more nuanced control over truthfulness compared to static methods.
• Targeted Activation Navigation: Employs multiple learned directions per layer (specifically layer 18 with 8 directions) to precisely guide the model's activations towards more truthful outputs.
• Flexible Usage: Can be used with Wisent's load_titan_model for full dynamic steering capabilities or as a standard Transformers model for static steering.

Good For

• Applications requiring high factual accuracy and reduced generation of misconceptions.
• Use cases where mitigating hallucination is critical, such as information retrieval, educational content generation, or fact-checking systems.
• Developers looking to integrate advanced steering mechanisms for fine-grained control over model behavior.