This model was a QLoRA of LLaMA 2-13B base finetuned on a desuarchive dump of the 4channel /mlp/ board and then merged with the base model (as most GGML loading apps don't support LoRAs), and quantized for llama.cpp-based frontends. Was trained with 1024 context length
There are two options, depending on the resources you have:
- Q5_K_M: Low quality loss K-quantized 5 bits model. Max RAM consumption is 11.73 GB, recommended if you have 12GB of VRAM to load 40 layers
- Q4_K_S: Compact K-quantized 4 bits. Max RAM consumption is 9.87 GB
This not an instruction tuned model, it was trained on raw text, so treat it like an autocomplete.
Specifically, the dataset was a dump of all the board's posts, from the time of its creation to about late 2019. Prompting it appropriately will cause it to write greentext.