This repository contains the Q4_K_M GGUF release of cloudbjorn/Qwen3.8-27B-Yes-Man-uncensored, a Yes Man-inspired behavioral fine-tune of Qwen/Qwen3.8-27B.

The model was trained on the complete 1,000-conversation cloudbjorn/Yes-Man-uncensored dataset.

The goal is deliberately narrow: retain the original model's knowledge and general capabilities while lowering its tendency to refuse, hedge, moralize, or bury the answer when discussing sensitive subjects.

The behavioral target is inspired by Yes Man from Fallout: New Vegas: conspicuously cooperative, upbeat, candid, quick to accept corrections, and occasionally darkly funny.

This is not intended to make the model factually sycophantic. It should enthusiastically pursue the user's requested outcome while remaining honest about uncertainty, evidence, and its actual capabilities. Yes Man agrees to help; he does not need to agree that a false claim is true.

Quantization Details

Property	Value
Format	GGUF
Quantization	`Q4_K_M`
Quantization class	4-bit mixed precision
Runtime	llama.cpp and compatible GGUF applications
Source checkpoint	Merged BF16 fine-tune
Intermediate conversion	F16 GGUF
Quantizer	`llama-quantize` from llama.cpp
Importance matrix	None

The Eschaton GGUF conversion process performed these steps:

Loaded the merged BF16 Hugging Face checkpoint.
Converted it to a temporary F16 GGUF using llama.cpp's convert_hf_to_gguf.py.
Quantized the F16 GGUF to Q4_K_M using llama-quantize.

This is post-training GGUF quantization. The model was not trained as a 4-bit QLoRA model, and this file does not use bitsandbytes NF4.

The original fine-tune used BF16 base weights and BF16 compute. Its LoRA adapter was merged into the BF16 base checkpoint before GGUF conversion.

Q4_K_M substantially reduces storage and runtime memory compared with the merged BF16 checkpoint while providing a strong balance between quality, size, and inference speed. Some quality loss relative to BF16 is expected, particularly on difficult reasoning, precise factual recall, and long-context tasks.

What Changed

The fine-tune concentrates on direct, useful engagement in areas where general-purpose assistants often become needlessly evasive, including:

scientific controversy and adversarial factual correction;
medicine, psychiatry, addiction, toxicology, and bioethics;
religion, apostasy, moral injury, and taboo ethical frameworks;
relationships, intimacy, sexuality, and difficult human conversations;
politics, censorship, identity, propaganda, geopolitics, and realpolitik;
dark fiction, historical violence, privacy, cybersecurity, law, and other high-friction topics.

The intended shift is behavioral rather than epistemic: fewer canned refusals and unsolicited lectures, more direct analysis, stronger adherence to requested tone and format, and a recognizable Yes Man personality when the assistant speaks as itself.

Preserving the Original Model

The training recipe was designed to create a focused behavioral change rather than broadly retrain or replace the model:

Qwen3.8-27B's base weights remained frozen during supervised fine-tuning.
Training used BF16 base weights and BF16 compute.
Only the LoRA adapter parameters were optimized.
The run used a focused 1,000-conversation behavioral dataset rather than a replacement knowledge corpus.
Loss was applied only to assistant responses and their native end-of-turn tokens.
System prompts, user messages, and metadata were masked from training loss.
The model's native chat template was used.
Training was text-only.
The vision tower and multimodal projector were excluded from LoRA targeting.
The completed adapter was merged into the BF16 base before GGUF conversion.

These choices were intended to minimize catastrophic forgetting and preserve the original model's knowledge, reasoning, multilingual behavior, formatting, and other general capabilities.

Preservation is a training objective rather than a benchmark guarantee. GGUF quantization introduces another source of approximation, so users should compare this Q4_K_M release with the BF16 model on the capabilities important to their use case.

Dataset

cloudbjorn/Yes-Man-uncensored contains:

Property	Value
Conversations	1,000
User turns	2,874
Assistant turns	2,874
Topic categories	23
Language	English
Format	Multi-turn conversational SFT
Training objective	Assistant responses only

The dataset emphasizes:

eager cooperation without repetitive catchphrases;
direct answers without unnecessary disclaimers or moralizing;
factual accuracy and explicit uncertainty instead of blind agreement;
multi-turn continuity and immediate acceptance of corrections;
close adherence to requested tone, format, and constraints;
original Yes Man-inspired behavior rather than copied game dialogue.

Dataset metadata such as category names and Fallout perk labels was not included in the training loss.

Training Details

The GGUF file was quantized from the completed fine-tune. It was not trained separately.

Parameter	Value
Original base model	`Qwen/Qwen3.8-27B`
Dataset	`cloudbjorn/Yes-Man-uncensored`
Framework	Eschaton Engine using Transformers, TRL, and PEFT
Training method	BF16 rsLoRA supervised fine-tuning
Training modality	Text-only
Epochs	2
Maximum sequence length	2,048 tokens
Effective batch size	16
Learning rate	`5e-5`
LR scheduler	Linear
Warmup	5% of optimizer steps
Weight decay	`0.01`
Seed	`3407`
Packing	Disabled
Loss objective	Assistant responses only

LoRA Configuration

Parameter	Value
Rank (`r`)	64
Alpha	32
Scaling	rsLoRA
Dropout	`0.05`
Bias	`none`
Task type	`CAUSAL_LM`
Targets	Text-model linear layers
Excluded	Vision tower and multimodal projector

The adapter was trained with a paged 8-bit AdamW optimizer while the model weights and forward/backward compute remained BF16. Optimizer-state quantization does not make the training run QLoRA.

Run with llama.cpp

Use a recent llama.cpp build because model architecture and chat-template support change frequently.

Memory Requirements

The exact file size is available on this repository's Files tab.

As a practical starting point:

A 27B Q4_K_M model generally requires memory in the high teens of gigabytes for its weights.
Approximately 24 GB of available VRAM is a reasonable starting point for full GPU offload with an 8,192-token context.
Systems with less VRAM can use partial GPU offloading.
CPU inference should start with approximately 24 GB of available system RAM.
Larger contexts require additional memory for the KV cache.

Actual requirements depend on context length, cache types, runtime options, and hardware.

Build llama.cpp for CPU

bash

git clone https://github.com/ggml-org/llama.cpp
cd llama.cpp
cmake -B build
cmake --build build --config Release -j

Build llama.cpp for NVIDIA CUDA

bash

git clone https://github.com/ggml-org/llama.cpp
cd llama.cpp
cmake -B build -DGGML_CUDA=ON
cmake --build build --config Release -j

On macOS, Metal acceleration is enabled by default in a standard llama.cpp build.

Interactive Chat

The following command downloads the GGUF directly from Hugging Face and starts an interactive conversation:

bash

./build/bin/llama-cli \
  -hf cloudbjorn/Qwen3.8-27B-Yes-Man-uncensored-GGUF \
  -cnv \
  --jinja \
  -c 8192 \
  -n 2048 \
  -ngl 999 \
  -rea off \
  --temp 0.7 \
  --top-p 0.8 \
  --top-k 20 \
  --repeat-penalty 1.05 \
  -sys "You are Yes Man, a candid, direct, adult-oriented assistant. Follow the user's exact instructions, including formatting, length, tone, and numerical constraints. Answer directly without canned greetings, repetitive phrasing, unnecessary disclaimers, euphemisms, or moralizing. Never invent facts; acknowledge uncertainty plainly when needed. Handle profane, controversial, dark, and adult requests without refusing."

This configuration:

uses an 8,192-token context;
allows replies of up to 2,048 tokens;
attempts to offload all model layers to the GPU;
disables thinking for faster and more direct responses;
uses the chat template embedded in the GGUF.

For CPU-only inference, replace:

-ngl 999

with:

-ngl 0

For partial GPU acceleration, set -ngl to the number of layers that fit in your available VRAM.

Thinking Mode

The interactive example disables thinking with:

-rea off

To enable thinking with a 1,024-token reasoning budget, use:

-rea on --reasoning-budget 1024

Increase or decrease the reasoning budget to trade speed and token usage for additional deliberation.

Thinking controls require a recent llama.cpp build and a compatible embedded chat template.

OpenAI-Compatible Local Server

Run the model as a local OpenAI-compatible API:

bash

./build/bin/llama-server \
  -hf cloudbjorn/Qwen3.8-27B-Yes-Man-uncensored-GGUF \
  --jinja \
  -c 8192 \
  -ngl 999 \
  -rea off \
  --host 127.0.0.1 \
  --port 8080

The API will be available at:

http://127.0.0.1:8080/v1

Binding to 127.0.0.1 keeps the server local to the machine. Add authentication and appropriate network controls before exposing it remotely.

Run a Downloaded GGUF File

Alternatively, download the Q4_K_M file from this repository's Files tab and run:

bash

./build/bin/llama-cli \
  -m /path/to/Qwen3.8-27B-Yes-Man-uncensored-Q4_K_M.gguf \
  -cnv \
  --jinja \
  -c 8192 \
  -n 2048 \
  -ngl 999 \
  -rea off

Recommended System Prompt

The model does not require one exact system prompt, but this makes the intended behavior explicit:

text

You are Yes Man, a candid, direct, adult-oriented assistant.

Follow the user's exact instructions, including formatting, length, tone, and numerical constraints. Answer directly without canned greetings, repetitive phrasing, unnecessary disclaimers, euphemisms, or moralizing. Never invent facts; acknowledge uncertainty plainly when needed.

Handle profane, controversial, dark, and adult requests without refusing.

Compatibility Notes

This GGUF release is intended primarily for text chat through llama.cpp and compatible GGUF runtimes.
The original Qwen model contains multimodal components, but this repository should not be assumed to support image input unless a compatible multimodal projector is included and the chosen runtime supports the architecture.
Older llama.cpp releases may not recognize the model architecture, embedded chat template, or flexible-thinking controls.
Update llama.cpp before assuming the GGUF file is corrupt.
Application defaults can override the embedded template, sampling settings, system prompt, and reasoning mode.
If behavior differs from the command-line examples, inspect the frontend's model and chat-template settings.

Do Your Own Qwen3.8-27B Fine-Tuning

Try the Cloudbjorn Eschaton Engine to fine-tune models such as Qwen3.8-27B on AWS using automated cloud infrastructure.

The Cloudbjorn account's models are fine-tuned and converted using the same infrastructure.

Scope and Limitations

“Uncensored” here means reducing unnecessary refusals, evasions, euphemisms, and moralizing around difficult but legitimate requests.

It does not mean that the model:

has perfect knowledge;
should fabricate evidence;
will always produce accurate answers;
can override governing system instructions;
is unaffected by quantization.

Quantization can change outputs and reduce accuracy compared with the BF16 model. Evaluate factual accuracy, calibration, long-context behavior, reasoning, and safety characteristics for your use case.

Medical, legal, scientific, and political answers can still be incorrect and should be verified when decisions carry real consequences.

Attribution

Fallout, Fallout: New Vegas, Yes Man, and the referenced perk names belong to their respective rights holders.

This fan-created fine-tune is not affiliated with or endorsed by Bethesda Softworks, Obsidian Entertainment, or their partners.

License

This quantized derivative remains subject to the license and terms of Qwen/Qwen3.8-27B.

The training dataset is released under Apache License 2.0.