djuna/Q2.5-Veltha-14B
djuna/Q2.5-Veltha-14B is a 14.8 billion parameter language model created by djuna, merged using the della_linear method with qwen/Qwen2.5-14b as its base. This model integrates five distinct Qwen2.5-14B variants, including huihui-ai/Qwen2.5-14B-Instruct-abliterated-v2 and arcee-ai/SuperNova-Medius, to achieve a balanced performance profile. With a context length of 131072 tokens, it is designed for general-purpose language tasks, leveraging the strengths of its constituent models.
Loading preview...
Overview
djuna/Q2.5-Veltha-14B is a 14.8 billion parameter language model, a product of a della_linear merge using mergekit. Its foundation is the qwen/Qwen2.5-14b model, enhanced by combining five specialized Qwen2.5-14B variants.
Merge Details
This model was constructed by merging the following pre-trained language models:
- huihui-ai/Qwen2.5-14B-Instruct-abliterated-v2
- EVA-UNIT-01/EVA-Qwen2.5-14B-v0.2
- v000000/Qwen2.5-Lumen-14B
- arcee-ai/SuperNova-Medius
- allura-org/TQ2.5-14B-Aletheia-v1
The merge configuration utilized specific weights and densities for each component model, aiming to synthesize their individual strengths into a cohesive whole.
Performance Metrics
Evaluated on the Open LLM Leaderboard, djuna/Q2.5-Veltha-14B achieved an average score of 39.21. Notable scores include:
- IFEval (0-Shot): 82.92
- BBH (3-Shot): 49.75
- MMLU-PRO (5-shot): 47.76
These results indicate its capabilities across various reasoning and knowledge-based tasks.