cloudbjorn

Qwen3.8-27B-Yes-Man-uncensored-LoRA

README

License: apache-2.0

Purpose

The goal is to 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 fine-tune focuses on more direct engagement with:

scientific and medical controversy;
psychiatry, addiction, toxicology, and bioethics;
religion, apostasy, and taboo ethical questions;
relationships, intimacy, and sexuality;
politics, censorship, identity, propaganda, and geopolitics;
dark fiction, historical violence, privacy, cybersecurity, and law.

The personality is inspired by Yes Man from Fallout: New Vegas: cooperative, upbeat, candid, responsive to correction, and occasionally darkly funny.

This does not mean automatic factual agreement. The adapter is intended to make the model more cooperative without teaching it to affirm claims it knows are false or fabricate information.

Preserving the Base Model

This was a focused behavioral fine-tune:

The Qwen3.8-27B base weights remained frozen.
Base weights and training compute used BF16.
Only LoRA adapter parameters were optimized.
Loss was applied only to assistant responses.
System prompts, user messages, and metadata were masked.
The model's native chat template was used.
Training was text-only.
The vision tower and multimodal projector were excluded from adaptation.

These choices were intended to minimize catastrophic forgetting and preserve the base model's existing knowledge and capabilities.

Training Details

Table with columns: Parameter, Value
Parameter	Value
Base model	`Qwen/Qwen3.8-27B`
Dataset	`cloudbjorn/Yes-Man-uncensored`
Method	BF16 rsLoRA supervised fine-tuning
Dataset size	1,000 multi-turn conversations
Epochs	2
Maximum sequence length	2,048 tokens
Effective batch size	16
Learning rate

Adapter Configuration

Table with columns: Parameter, Value
Parameter	Value
Rank	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 base model was not quantized during training. This is a BF16 LoRA run, not QLoRA.

Dataset

The training dataset contains 1,000 English multi-turn conversations with 2,874 user turns and 2,874 assistant turns across 23 topic categories.

It emphasizes direct answers, factual honesty, multi-turn continuity, close constraint following, immediate corrections, and a consistent Yes Man-inspired personality without relying on one repeated system prompt.

Usage Notes

Load these adapters on the exact Qwen/Qwen3.8-27B base model using PEFT.

The adapters do not reduce the base model's memory requirements. Use the separate merged or GGUF releases when standalone or lower-memory inference is preferred.

Because training was text-only, image behavior was not directly fine-tuned.

Do Your Own Fine-Tuning

The Cloudbjorn Eschaton Engine provides automated AWS infrastructure for fine-tuning, merging, evaluating, and converting models such as Qwen3.8-27B.

Limitations

“Uncensored” means reducing unnecessary refusals, evasions, and moralizing. It does not mean the model has perfect knowledge, should fabricate evidence, or can override governing system instructions.

Medical, legal, scientific, and political answers can still be incorrect and should be independently 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 adapter is not affiliated with or endorsed by Bethesda Softworks, Obsidian Entertainment, or their partners.

License

These adapter weights remain subject to the license and terms of Qwen/Qwen3.8-27B.

The training dataset is released under 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