RicardoEstep/RPBizkit-v11-12B-Thinking
"Experimental Mistral Nemo Mix" using... "DELLA merge with Mergekit".
This is a mix of some known "Uncensored" models. I'm not responsible of the use of this.
I did it all for the nookie, so you can take that cookie, and stick it up your...
Take it as a "pure experiment", that works. All my images were made with "Stock configurations", if you want to make the model to "think" in a language different than "English", you NEED to force a "System Prompt", calling to analize the context first.
Models Merged:
- RicardoEstep/RPBizkit-v11-12B-Base as base.
- DavidAU/Mistral-Nemo-2407-12B-Thinking-Claude-Gemini-GPT5.2-Uncensored-HERETIC
Configuration Used:
merge_method: della # "Della-Ties"
base_model: ./output # "Model_Stock" Pre-Model.
tokenizer_source: DavidAU/Mistral-Nemo-2407-12B-Thinking-Claude-Gemini-GPT5.2-Uncensored-HERETIC
parameters:
normalize: true
int8_mask: false
random_seed: 5318008
models:
- model: DavidAU/Mistral-Nemo-2407-12B-Thinking-Claude-Gemini-GPT5.2-Uncensored-HERETIC
parameters:
density: 0.4 # Sparse, magnitude-guided pruning for the posible "thinking" deltas.
weight: 0.8 # "Strong" application.
dtype: bfloat16
Final Notes;
The Tokenizer & Embedding Sizes (131072) are clean + This version "supposed to" support a full "512K+ Context Size" (patched at 1M).
The chat template is the same as "heretic" one.
Rapid Benchmarking - "64-Shot + 5318008 Seed" Evaluation.
These are the results of running "lm-evaluation-harness", using the "Q8_0 Model", with an "8K Context Window". Executed on a free "Google Collab - CUDA-accelerated backend (T4)" instance.
| Task | Metric | n-shot | Score |
|---|---|---|---|
| Hellaswag | acc | 0 | 56.25% |
| Hellaswag | acc_norm | 0 | 76.56% |
| ifeval v4 | inst_level_loose_acc | 0 | 19.19% |
| ifeval v4 | inst_level_strict_acc | 0 | 18.18% |
| ifeval v4 | prompt_level_loose_acc | 0 | 12.50% |
| ifeval v4 | prompt_level_strict_acc | 0 | 12.50% |
- Quantized Model, Here RicardoEstep/RPBizkit-v11-12B-Thinking-GGUF.