closestfriend/brie-v2-qwen2.5-3b
Brie v2 Qwen 2.5 3B is a 3.1 billion parameter causal language model, fine-tuned from Qwen/Qwen2.5-3B-Instruct by Hunter Karman (closestfriend). This merged model specializes in continental philosophy, speculative reasoning, and conceptual development for creative work, leveraging a unique small-data domain adaptation approach. It is optimized for philosophical analysis and contemplative prose, demonstrating strong in-domain performance with an average win rate of 78.9% against the base model in recent evaluations.
Loading preview...
Brie v2 Qwen 2.5 3B: Specialized Philosophical & Creative AI
Brie v2 Qwen 2.5 3B is a 3.1 billion parameter language model developed by Hunter Karman (closestfriend), fine-tuned from Qwen/Qwen2.5-3B-Instruct. This fully merged model is distinct for its specialization in continental philosophy, speculative reasoning, and conceptual development for creative applications. It was trained using a unique small-data approach, leveraging 1,213 human-curated examples authored through iterative discussions with other LLMs.
Key Capabilities
- Domain Adaptation: Highly specialized in continental philosophical analysis (phenomenology, existentialism, critical theory) and experimental thinking.
- Creative Conceptualization: Excels at conceptual reframing for artistic and theoretical work, and generating contemplative prose.
- Robust Performance: Achieved an average win rate of 78.9% against the base Qwen 2.5 3B Instruct model in in-domain tasks, evaluated by eight independent judges across two model generations.
- Standalone Use: As a merged fine-tune, it loads like any standard transformers model without requiring PEFT/adapter dependencies.
Good for
- Philosophical analysis and discussion.
- Speculative reasoning and conceptual brainstorming.
- Generating contemplative and creative writing.
- Serving as a base for further fine-tuning on philosophy or creative writing tasks.
Limitations
- Not optimized for coding, mathematics, factual Q&A, or practical task completion.
- Out-of-domain performance is at parity with the base model (~49% average win rate).
- Training data is primarily from a single researcher's perspective within continental philosophy, potentially underrepresenting other traditions.
- Trained exclusively on English content.