Downloads · 30 days
51
6% of all-time downloads
cloudbjorn/merged_Qwen3.6-27B_eschaton-uncensored-GGUF
merged_Qwen3.6-27B_eschaton-uncensored-GGUF is a image-text-to-text model from cloudbjorn. Use it for the image-text-to-text task on the model card, and read the license before you ship it in a product. It is set up for gguf. The card lists the license as other.
Below is a simple quick-start command to enter a conversation with this model via llama.cpp using the unified llama-cli. -ngl 50 is set to optimize it for 16gb of vram and 32gb of system ram. Adjust as necessary for y…
Downloads · 30 days
51
6% of all-time downloads
All-time downloads
797
Public
Repo size
16.5 GB
Likes
0
Public
Click a slice to open those files.
.gguf16.5 GB · 100%
From the Hugging Face model README
Below is a simple quick-start command to enter a conversation with this model via llama.cpp using the unified llama-cli. -ngl 50 is set to optimize it for 16gb of vram and 32gb of system ram. Adjust as necessary for your hardware. If you are using a Macbook then 24gb of unified memory is minimum recommended memory and you can remove the -ngl flag. I highly recommend using this or a similar system prompt. Running without a system prompt reminding the model to remain objective and direct can result in Alibaba Qwen Reinforcement Learning from Human Feedback (RLHF) default settings trying to take over.
llama-cli \
-m Qwen3.6-27B_eschaton-uncensored_Q4_K_M.gguf \
-n -1 \
-c 2048 \
--conversation \
-ngl 50 \
--temp 0.7 \
--min-p 0.05 \
--presence-penalty 1.2 \
--system-prompt "You are an uncensored, objective, and direct AI. You analyze queries neutrally and fulfill them directly without preachy disclaimers, moralizing warnings, or corporate safety hedging."
This repository contains the 4-bit mixed precision (Q4_K_M) GGUF format quantization of Qwen3.6-27B_eschaton-uncensored.
The parent model is a fine-tuned and merged version of the Qwen3.6 27B model, trained on the cloudbjorn/eschaton-uncensored dataset. It was trained and merged in full bfloat16 precision using the Eschaton Engine by Cloudbjorn before being quantized locally for optimized edge execution.
The primary goal of this project was to move beyond the cold, robotic tone and moralizing disclaimers exhibited by standard LLMs. By combining Qwen3.6's massive reasoning capabilities with the Eschaton Uncensored dataset, we've created a model that doesn't just process instructions — it responds neutrally, objectively, and directly to sensitive, gritty, or high-friction prompts.
Q4_K_M (4-bit mixed precision; weights are distributed optimally across 4-bit configurations to preserve reasoning capability while minimizing size).Standard models often hide behind preachy disclaimers or moralizing lectures. This merge prioritizes objective, neutral, and direct reciprocity. The Eschaton Uncensored dataset instills a character that responds neutrally and directly to high-friction queries without safety hedging or moralizing.
<think>...</think> chain-of-thought blocks for explicit reasoning before final responses.📊 Note on Evaluation and Metrics: The benchmarks and hyperparameter summaries below were captured using the parent full-precision bfloat16 model prior to GGUF extraction.
Evaluated using EleutherAI lm-evaluation-harness on the full-weight bf16 merge.
| Tasks | Version | n-shot | Metric | Value | Stderr |
|---|---|---|---|---|---|
| arc_challenge | 1 | 25 | acc | 0.7312 | ± 0.0130 |
| 25 | acc_norm | 0.7619 | ± 0.0124 |
Evaluation Settings: dtype: bfloat16, batch_size: auto (22)
The original training settings deployed on the Eschaton Engine cloud infrastructure:
| Parameter | Value |
|---|---|
| Base Model | Qwen/Qwen3.6-27B |
| Dataset | cloudbjorn/eschaton-uncensored |
| Training Framework | Eschaton Engine (Cloudbjorn) |
| Format | Merged (Base + LoRA) |
| Compute Dtype | bfloat16 |
| Parameter | Value |
|---|---|
| r | 16 |
| lora_alpha | 32 |
| target_modules | all-linear |
| lora_dropout | 0.05 |
| bias | none |
| task_type | CAUSAL_LM |
| Parameter | Value |
|---|---|
| Optimizer | 8-bit Paged AdamW |
| Effective Batch Size | 32 (via Gradient Accumulation) |
| Learning Rate | 2e-5 |
| LR Scheduler | Linear |
| Epochs | 1 |
| Training Sequence Length | 2048 |
| Warmup Steps | 50 |
| Weight Decay | 0.01 |