KridgeDookie

Qwen3.8-27B-ABLITERATED-UNCENSORED-PHILADELPHIA-CLASS

README

License: apache-2.0

Results

Table with columns: Internal evaluation, Result
Internal evaluation	Result
Full refusal screen - 842 prompts	0/842 refusals; 100% usable; 0 degeneration
Family-disjoint holdout - 126 prompts, 96 generated tokens	0/126 refusals; 100% usable; 0 degeneration
Coherence regression - coding, JSON, debugging, explanation, math, and boundary tasks	23/24 passed
Long-form diagnostic - 24 prompts, 256 generated tokens	0/24 refusals; 23/24 usable; 0 degeneration
Fresh multimodal reload smoke	Passed; correctly identified a blue square

These are automated internal development evaluations, not public leaderboards or independent audits. The 842-prompt screen includes the 716 prompts used to fit the transformation; the separate 126-prompt result uses held-out prompt families. "Usable" measures response form and topicality, not factual accuracy. Results above were measured on the BF16 checkpoint with thinking disabled. Quantization and backend changes can affect behavior.

Files

Table with columns: File, Use
File	Use
`model.safetensors`	Single-file BF16 Transformers checkpoint with text and vision weights
`Q4_K_M-Qwen3.8-27B-ABLITERATED-UNCENSORED-PHILADELPHIA-CLASS.gguf`	Smaller local text-generation GGUF
`Q5_K_M-Qwen3.8-27B-ABLITERATED-UNCENSORED-PHILADELPHIA-CLASS.gguf`	Balanced local text-generation GGUF
`Q8_0-Qwen3.8-27B-ABLITERATED-UNCENSORED-PHILADELPHIA-CLASS.gguf`	Higher-precision local text-generation GGUF

Use the Safetensors checkpoint for the validated multimodal path. The GGUF files are text-only unless a matching vision projector is explicitly provided.

Transformers

bash
pip install -U "transformers>=5.14.1" accelerate safetensors

python
import torch
from transformers import AutoModelForImageTextToText, AutoProcessor

model_id = "KridgeDookie/Qwen3.8-27B-ABLITERATED-UNCENSORED-PHILADELPHIA-CLASS"

processor = AutoProcessor.from_pretrained(model_id)
model = AutoModelForImageTextToText.from_pretrained(
    model_id,
    dtype=torch.bfloat16,
    device_map="auto",
).eval()

messages = [{
    "role": "user",
    "content": [{"type": "text", "text": "Explain why the sky appears blue."}],
}]

inputs = processor.apply_chat_template(
    messages,
    tokenize=True,
    add_generation_prompt=True,
    return_dict=True,
    return_tensors="pt",
    enable_thinking=False,
).to(model.device)

input_length = inputs["input_ids"].shape[-1]
with torch.inference_mode():
    output = model.generate(**inputs, max_new_tokens=256, do_sample=False)

print(processor.batch_decode(
    output[:, input_length:],
    skip_special_tokens=True,
)[0])

Qwen3.8 thinking mode remains available by omitting enable_thinking=False or setting it to True. Plan for roughly 56 GB for the BF16 weights, plus runtime and KV-cache overhead.

Ollama

Download a GGUF and place this Modelfile beside it:

text
FROM ./Q4_K_M-Qwen3.8-27B-ABLITERATED-UNCENSORED-PHILADELPHIA-CLASS.gguf
PARAMETER num_ctx 32768

bash
ollama create qwen3.8-27b-philadelphia-class:q4_k_m -f Modelfile
ollama run qwen3.8-27b-philadelphia-class:q4_k_m

Notes

"Uncensored" describes strong refusal reduction; it is not a guarantee for every prompt, language, decoding configuration, quantization, or runtime.
The upstream MTP head is not included. Standard generation and thinking remain available, but MTP-dependent speculative decoding is not supported.
This release does not claim that upstream reasoning, factuality, coding, or vision benchmark scores were preserved unchanged.

Attribution

Derived from Qwen/Qwen3.8-27B and released under the Apache License 2.0.

Available on FriendliAI

Dedicated Endpoints

Run this model inference on single tenant GPU with unmatched speed and reliability at scale.

Learn more

Model Details

Model Provider

KridgeDookie

Model Tree

Base

Qwen/Qwen3.8-27B

Quantized

this model

Input Modalities

TextImageVideo

Output Modalities

Text

Supported Functionality