base_model: Qwen/Qwen3.6-27B
base_model_relation: finetune
library_name: transformers
pipeline_tag: image-text-to-text
tags:
- text-generation-inference
- transformers
- qwen
- qwen3
- reasoning
- eschaton-engine
- eschaton-uncensored
- uncensored
license: other
datasets: - cloudbjorn/eschaton-uncensored
Qwen3.6-27B_eschaton-uncensored
This is a fine-tuned and merged version of the Qwen3.6 27B model, trained on the cloudbjorn/eschaton-uncensored dataset. Trained and merged in full bfloat16 precision using the Eschaton Engine.
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.
Personality: Objectivity Over Preachiness
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.
Model Capabilities
- Large Context Window: Supports up to 262,144 tokens (Qwen3.6 native).
- Native Thinking Mode: Supports Qwen3's
<think>...</think>chain-of-thought blocks for explicit reasoning before final responses. - Advanced Formatting: Native support for tool use and structured output.
- Full 16-Bit Precision: Trained and merged in bfloat16 — zero-loss parameter density.
Benchmarks: ARC Challenge
Evaluated using EleutherAI lm-evaluation-harness.
25-Shot (Leaderboard Standard)
| 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)
Training Details
| Parameter | Value |
|---|---|
| Base Model | Qwen/Qwen3.6-27B |
| Dataset | cloudbjorn/eschaton-uncensored |
| Training Framework | Eschaton Engine (Cloudbjorn) |
| Format | Merged (Base + LoRA) |
| Compute Dtype | bfloat16 |
LoRA Parameters (Auto-Scaled for 27B)
| Parameter | Value |
|---|---|
| r | 16 |
| lora_alpha | 32 |
| target_modules | all-linear |
| lora_dropout | 0.05 |
| bias | none |
| task_type | CAUSAL_LM |
Hyperparameters
| 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 |