CoolFace
Modelpublic

RicardoEstep/RPBizkit-v4-12B_Lorablated

sourceHugging Faceupdated 7mo agoView on Hugging Face
2likes24downloads
Model Card

RicardoEstep/RPBizkit-v4-12B_Lorablated

"Experimental Own Mix" using "Python" (?).

I did it all for the nookie, so you can take that cookie, and stick it up your...

It's based on the original RicardoEstep/RPBizkit-v4-12B merged with nbeerbower/Mistral-Nemo-12B-abliterated-LORA. A complete mix of some known "RP Uncensored" models. I'm not responsible of the use of this.

image/png

Data used:

Notes:

The recommended max context size is "8K (8192)". Some of the models in the mix, used a "fake rope_theta hack" to support "1M" of context size, and this model coppied that configuration. This are not a real numbers, and would NOT give you a "meaningful long‑context behavior".

This version has a "Clean Tokenizer" I used "yamatazen/EtherealAurora-12B" Tokenizer based on "ChatML", as "mergin the LoRa" forced me to do it.

Still, the model will drift if any "Chat Template" based on "ChatML" or "Mistral" is used. The recommended chat template to use is "Alpaca (with "RAW" inputs)". The configuration files are "Tweaked", leading to don't use any "Chat Template".

Python Script Used:

python
import torch
from transformers import AutoModelForCausalLM, AutoTokenizer
from peft import PeftModel, LoraConfig

# --- CONFIGURATION ---
base_model_path = "RicardoEstep/RPBizkit-v4-12B"
lora_path = "nbeerbower/Mistral-Nemo-12B-abliterated-LORA"
tokenizer_path = "yamatazen/EtherealAurora-12B"
output_path = "./RPBizkit-v4-12B-Abliterated-ChatML"

print("Loading base model...")
# We load without device_map="auto" initially to avoid naming issues with accelerate
model = AutoModelForCausalLM.from_pretrained(
    base_model_path,
    torch_dtype=torch.bfloat16,
    low_cpu_mem_usage=True,
    trust_remote_code=True
)

# 1. FIX THE VOCAB SIZE (The 131075 -> 131072 issue)
print(f"Resizing from {model.get_input_embeddings().weight.shape[0]} to 131072...")
model.resize_token_embeddings(131072)

# 2. APPLY THE LORA MANUALLY
print("Applying LoRA...")
# We use from_pretrained but specify the exact model to avoid double-nesting
model = PeftModel.from_pretrained(
    model, 
    lora_path,
    adapter_name="default"
)

# 3. MERGE THE WEIGHTS
print("Merging weights into base...")
model = model.merge_and_unload()

# 4. FIX THE TOKENIZER
print("Finalizing tokenizer...")
tokenizer = AutoTokenizer.from_pretrained(tokenizer_path, trust_remote_code=True)

# 5. SAVE
model.save_pretrained(output_path)
tokenizer.save_pretrained(output_path)

print("Process Complete!")