NeuralDaredevil-8B-abliterated
| Property | Value |
|---|---|
| Parameter Count | 8.03B |
| Model Type | Text Generation |
| Architecture | Llama-based |
| License | Llama3 |
| Tensor Type | FP16 |
What is NeuralDaredevil-8B-abliterated?
NeuralDaredevil-8B-abliterated is a sophisticated language model that represents a DPO fine-tuned version of the Daredevil-8-abliterated base model. It has been trained on one epoch of the orpo-dpo-mix-40k dataset, successfully recovering performance losses from the abliteration process while maintaining its uncensored nature.
Implementation Details
This model leverages the Llama architecture and implements Direct Preference Optimization (DPO) for fine-tuning. It achieves impressive benchmark scores, including 69.1% on MMLU, 71.8% on GSM8k, and 85.05% on HellaSwag, making it the best-performing uncensored 8B model on the Open LLM Leaderboard.
- Advanced DPO fine-tuning methodology
- Multiple quantization options (GGUF, EXL2, AWQ)
- Optimized for both performance and computational efficiency
- Compatible with LM Studio using "Llama 3" and "Llama 3 v2" presets
Core Capabilities
- High-performance text generation and role-playing tasks
- Strong reasoning capabilities (69.28% on AI2 Reasoning Challenge)
- Enhanced truthfulness (60% on TruthfulQA)
- Superior mathematical reasoning (71.8% on GSM8k)
Frequently Asked Questions
Q: What makes this model unique?
The model stands out for its successful combination of DPO fine-tuning and abliteration recovery, achieving state-of-the-art performance for uncensored 8B models while maintaining computational efficiency.
Q: What are the recommended use cases?
The model excels in applications that don't require strict alignment, particularly in role-playing scenarios. It offers superior performance compared to the Instruct model and supports various deployment options through different quantization methods.
