Epiculous/Effervescence-27B
Epiculous/Effervescence-27B is a 27 billion parameter instruction-tuned causal language model based on Qwen3.5-27B, developed by Epiculous. It utilizes a two-phase training methodology with an expanded instruct dataset, optimized for efficient training on consumer-grade GPUs. This model is specifically trained on ChatML for conversational applications and general instruction following.
Loading preview...
Overview
Epiculous/Effervescence-27B is a 27 billion parameter instruction-tuned model built upon the Qwen3.5-27B base. It employs a two-phase training methodology, similar to the Crimson_Dawn line, but with a significantly expanded instruct dataset. The model was trained as an 8-bit LoRA to enable training on 2x NVIDIA A6000 GPUs, making it accessible for setups with limited VRAM.
Key Capabilities
- Instruction Following: Trained extensively on an expanded instruct dataset, making it proficient in understanding and executing various instructions.
- ChatML Compatibility: Designed to work seamlessly with the ChatML prompting structure, ensuring straightforward integration into chat-based applications.
- Efficient Training: Utilizes an 8-bit LoRA training approach, demonstrating effective scaling of larger models on more constrained hardware.
Training Details
The training process involved two distinct phases. Initially, the base model was trained on RP completion data, with the LoRA then applied. Subsequently, this modified base underwent further training on instruct data, again as an 8-bit LoRA, which was then applied to produce the final model. The entire process was built with Axolotl.