Kukedlc/NeuralLLaMa-3-8b-DT-v0.1
Kukedlc/NeuralLLaMa-3-8b-DT-v0.1 is an 8 billion parameter merged language model based on the Meta-Llama-3-8B architecture, created by Kukedlc. This model is a merge of mlabonne/ChimeraLlama-3-8B-v2, nbeerbower/llama-3-stella-8B, and uygarkurt/llama-3-merged-linear using the DARE TIES merge method. It is designed for general language tasks, demonstrating an average performance of 21.12 on the Open LLM Leaderboard evaluation metrics.
Loading preview...
Model Overview
Kukedlc/NeuralLLaMa-3-8b-DT-v0.1 is an 8 billion parameter language model developed by Kukedlc. It is a product of merging three distinct Llama-3-8B based models: mlabonne/ChimeraLlama-3-8B-v2, nbeerbower/llama-3-stella-8B, and uygarkurt/llama-3-merged-linear. This merge was performed using the DARE TIES method via LazyMergekit, with NousResearch/Meta-Llama-3-8B serving as the base model.
Key Characteristics
- Merged Architecture: Combines the strengths of multiple Llama-3-8B fine-tunes.
- Merge Method: Utilizes the
dare_tiesmethod for combining model weights, with specific density and weight parameters for each contributing model. - Quantization Support: Configured for 4-bit quantization (
bnb_4bit_use_double_quant=True,bnb_4bit_quant_type="nf4") for efficient deployment.
Performance Insights
Evaluations on the Open LLM Leaderboard indicate an average score of 21.12. Specific metric scores include:
- IFEval (0-Shot): 43.71
- BBH (3-Shot): 28.01
- MMLU-PRO (5-shot): 31.02
Usage
The model can be loaded and used with the transformers library, supporting 4-bit quantization for reduced memory footprint. Example Python code is provided for setting up the tokenizer, model, and generating responses with a streaming output.
Top 3 parameter combinations used by Featherless users for this model. Click a tab to see each config.