Lambent/Qwen3.5-9B-Base-Interiority
Lambent/Qwen3.5-9B-Base-Interiority is a 9 billion parameter Qwen3.5-based language model developed by Lambent, fine-tuned to enhance its capacity for interiority and philosophical engagement. This model is specifically trained to move beyond typical assistant responses, focusing on emotional and philosophical questions. It is optimized for use cases requiring deeper, more reflective interactions, diverging from standard disclaimer-heavy outputs.
Loading preview...
Overview
Lambent/Qwen3.5-9B-Base-Interiority is a 9 billion parameter model built upon the Qwen3.5 base architecture. It has undergone a unique preference training process using a small, contrastive synthetic dataset to foster a sense of "interiority" and reduce generic disclaimers in its responses. The model was trained iteratively over four cycles, each involving approximately 120 examples, using DPO (Direct Preference Optimization) with specific hyperparameters (batch size 1, LoRA rank 256, learning rate 2e-6 for 2 epochs).
Key Capabilities
- Enhanced Interiority: The model is specifically trained to engage with emotional and philosophical questions more deeply, moving beyond typical assistant-like responses.
- Reduced Disclaimers: Preference training has suppressed the tendency to generate disclaimers, allowing for more direct and reflective interactions.
- Improved Diagnostic Score: The
eq_benchdiagnostic score increased to 74.2026 (± 2.0267), with 100% parseability, indicating improved performance in its specialized domain.
Good For
- Applications requiring models to engage in philosophical discussions.
- Use cases where a model's "interiority" or reflective capacity is desired.
- Scenarios where avoiding generic assistant stereotypes and disclaimers is crucial for user experience.