Overview

Nemesia-Qwen-2.5-7B-v1.0 is an experimental 7.6 billion parameter language model created by tssst, utilizing the novel NuSLERP merge method. It is built upon the EVA-UNIT-01/EVA-Qwen2.5-7B-v0.1 base model and integrates two additional finetunes: allura-org/Teleut-7b and FourOhFour/Vapor_v2_7B. The merge process specifically enabled both NuSLERP exclusive mergekit options for exploration.

Key Characteristics

• Merge Method: Employs the NuSLERP technique, a relatively new merging approach.
• Base Model: Uses EVA-UNIT-01/EVA-Qwen2.5-7B-v0.1 as its foundation.
• Component Models: Includes allura-org/Teleut-7b (0.6 weight) and FourOhFour/Vapor_v2_7B (0.2 weight) in the merge configuration.
• Experimental Nature: The model is explicitly stated as experimental, with performance not guaranteed.

Known Limitations

• Quantization Issues: The creator has reported consistent failures when attempting to quantize the model (e.g., for GGUF formats), citing llama_model_quantize: failed to quantize: tensor 'blk.24.attn_norm.weight' has invalid data errors. This issue persisted across multiple attempts with different base models and parameters.
• Recommendation: The creator advises not to use this model due to these unresolved quantization problems, suggesting it may be a "lost cause" unless others can resolve the issues.