base_model:
- DavidAU/gemma-4-31B-it-The-DECKARD-HERETIC-UNCENSORED-Thinking
- google/gemma-4-31B-it ---These models are quantized and tuned versions of David Au's EXCELLENT Deckard Heretic Gemma 4 model. This model is intended to have alot of features, upgraded OCR, EM-LLM and RLM Context and memory schemes, Context Compaction Offloading, Super AI logic tweaks and more!
UPDATED! I bought a B70 Battlemage and tuned up a BF16 version with full-tuning and Logic, Math, Abstention training from Qwen and Opus. 1st 3 Layers are burned in with Qwen, Opus, Pathfinder first edition Rule Set and Core, Forgotten Realms Lore from several editions plus conversions to make most of the older material Pathfinder 1st Edition Ready. Gemma 4 has problems in coding with read/write/execute unless you tell her it is time to RWX. She is a beautiful planner in code; let her sit on a problem overnight and cook. Fantastic fiction writer; David AU really really f#$%ing cooked with this masterstroke of a local AI. Thank you duders!!!!! So yeah, please try this Safetensor BF16 model out. I did not quantize cause I ran out of Bandwith to upload to be honest LOLOL. However since you have the base model, you can easily quantize yourself to your need. If you have questions, comments, concerns, please shoot me a message on here and I will respond. Thanks and ENJOY! OH YEAH! THE GGUF FILES ARE OLDER AND NOT FINE TUNED; WHILE THEY ARE TRAINED THEY ARE NOT THE SAME AS THE SAFETENSORS/BF16
Please when quantizing for your setup to leave a few GB of KVCache buffer to let the context schemes do their work. It balloons them up a little bit and confuses the size; like on my 32gb card I quantize down to 28 gb, 24 to 20 (usually ends up at 18)
edit: Q6_K has no stack overflow in llama. might be right shape or cpp got updated to eat it? lmk please :D RAG train this model one more time to reinforce on rules or other game systems, it runs VERY well.
IF YOU GET A SERVER ERROR 500 FROM OLLAMA, OPEN UP TWO TERMINALS; ONE RUN "OLLAMA SERVE" and in the other run the model file and it should load.
- Downloads last month
- 478