paperboygold/gpt-oss-sanguine-20b-4bit-bnb

sanguine-scribe-4bit-bnb

4-bit quantized version using BitsAndBytes for efficient GPU inference.

This is a quantized version of gpt-oss-sanguine-20b-v1, a consequence-based alignment model for character roleplay.

Usage

from transformers import AutoTokenizer, AutoModelForCausalLM

tokenizer = AutoTokenizer.from_pretrained("paperboygold/sanguine-scribe-4bit-bnb")
model = AutoModelForCausalLM.from_pretrained(
    "paperboygold/sanguine-scribe-4bit-bnb",
    device_map="auto",
    trust_remote_code=True
)

Original Model

• Base Model: openai/gpt-oss-20b
• Training Dataset: sanguine-dataset-v1 (350K examples)
• Training Loss: 4.1 → 1.31 (500 steps)