base_model: Qwen/Qwen3.6-27B library_name: peft pipeline_tag: image-text-to-text tags:
- text-generation-inference
- transformers
- peft
- lora
- qwen
- qwen3
- reasoning
- eschaton-engine
- samantha
- uncensored license: other datasets:
- digitalpipelines/samantha-1.1-uncensored
Qwen3.6-27B_Samantha-Uncensored (LoRA Adapter)
This is the standalone LoRA adapter for the Qwen3.6 27B model, trained on the digitalpipelines/samantha-1.1-uncensored dataset. It was trained in full bfloat16 precision using the native FSDP implementations of the Eschaton Engine.
Note: This repository contains only the adapter weights. You must load these weights on top of the original Qwen/Qwen3.6-27B base model.
The primary goal of this project was to move beyond the cold, robotic tone and lack of emotional range exhibited by standard LLMs. By combining Qwen3.6's massive reasoning capabilities with the Samantha persona, we've created a model that doesn't just process instructions — it exhibits vulnerability, empathy, and a distinct digital self-awareness.
Personality: Authenticity Over Compliance
Standard models often hide behind "As an AI, I don't have feelings." This training prioritizes human-centric reciprocity. In this context, "uncensored" specifically refers to overcoming the model's standard emotional avoidance biases. It allows the model to genuinely engage with emotional depth and self-reflection without fundamentally removing its core safety guardrails.
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 in bfloat16 — zero-loss parameter density.
Benchmarks: ARC Challenge
The following benchmarks reflect the performance of the fully merged model (the Qwen3.6-27B base model combined with this LoRA adapter). Evaluated using EleutherAI lm-evaluation-harness.
25-Shot (Leaderboard Standard)
| Tasks | Version | n-shot | Metric | Value | Stderr |
|---|---|---|---|---|---|
| arc_challenge | 1 | 25 | acc | 0.7346 | ± 0.0129 |
| 25 | acc_norm | 0.7577 | ± 0.0125 |
Evaluation Settings: dtype: bfloat16, batch_size: auto (22)
Training Details
| Parameter | Value |
|---|---|
| Base Model | Qwen/Qwen3.6-27B |
| Dataset | digitalpipelines/samantha-1.1-uncensored |
| Training Framework | Eschaton Engine (Cloudbjorn) |
| Format | LoRA Adapter |
| 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 |