N8Programs/Unslopper-30B-A3B-bf16

Unslopper-30B-A3B

Model Description

Unslopper-30B-A3B is a fine-tuned language model designed to transform AI-generated text into more human-like prose while preserving semantic content. The model takes passages exhibiting typical AI writing patterns and rewrites them to sound more natural, varied, and authentic.

• Base Model: Qwen3-VL-Text-30B-A3B-Instruct (6-bit quantized)
• Architecture: Mixture of Experts (MoE) with 30B total parameters, 3B active
• Fine-tuning Method: LoRA (Low-Rank Adaptation)
• Framework: MLX (Apple Silicon optimized)

Intended Use

The model is intended to:

• Improve the naturalness of AI-generated creative writing
• Reduce detectable AI patterns in text (stylistic homogeneity, predictable phrasing)
• Serve as a post-processing step for AI writing assistants

Not intended for: Bypassing AI detection for academic dishonesty, fraud, or deceptive purposes.

Prompt Template

Use the default jinja template with the user prompt:

"Rewrite this AI passage to sound more humanlike:\n{passage}"

Essentially:

prompt = f"Rewrite this AI passage to sound more humanlike:\n{passage}"
messages = [{"role": "user", "content": prompt}]

Training Data

Data Generation Pipeline

The training data was synthetically generated using a novel "reverse distillation" approach:

1. Source: Human-written literary passages extracted from a Sam Paech preference dataset sourced from Project Gutenberg.
2. AI-ification Process: Each human passage was iteratively rewritten 10 times by GPT-4o-mini, progressively amplifying AI-typical writing patterns
3. Pair Creation: Final pairs consist of (AI-refined passage → original human passage)

This creates a supervised learning signal where the model learns to reverse the AI-ification process. The full dataset can be found at N8Programs/unslop-good.

Dataset Statistics

Metric	Value
Training examples	1,000
Refinement iterations per passage	10
Total API calls for data generation	10,000
Source	Literary fiction passages

Training Configuration

Model Architecture

Parameter	Value
Base model	Qwen3-VL-Text-30B-A3B-Instruct
Quantization	6-bit
Total parameters	30B
Active parameters	3B (MoE)

LoRA Configuration

Parameter	Value
Rank	8
Scale (alpha)	20.0
Dropout	0.0
Layers fine-tuned	48
Target modules	self_attn.q_proj, self_attn.v_proj, self_attn.k_proj, self_attn.o_proj, mlp.gate_proj, mlp.switch_mlp.gate_proj, mlp.switch_mlp.up_proj, mlp.switch_mlp.down_proj

Training Hyperparameters

Parameter	Value
Optimizer	Adam
Learning rate	1e-4
LR schedule	Cosine decay with warmup
Warmup steps	10
Warmup init LR	1e-5
Final LR	1e-5
Batch size	1
Gradient accumulation	1
Training iterations	1,000
Max sequence length	6,144
Gradient checkpointing	Enabled

Optimizer Configuration

optimizer: adam
betas: [0.9, 0.9999]
eps: 1e-6
bias_correction: true

Inference

Recommended Settings

from mlx_lm import load, generate
from mlx_lm.sample_utils import make_sampler, make_logits_processors

model, tokenizer = load("Unslopper-30B-A3B-6bit")

def unslop(passage: str) -> str:
    prompt = f"Rewrite this AI passage to sound more humanlike:\n{passage}"
    messages = [{"role": "user", "content": prompt}]

    output = generate(
        model,
        tokenizer,
        tokenizer.apply_chat_template(messages, add_generation_prompt=True),
        max_tokens=4096,
        sampler=make_sampler(temp=0.8),
        logits_processors=make_logits_processors(repetition_penalty=1.1),
    )
    return output.strip()

Inference Parameters

Parameter	Recommended Value
Temperature	0.8
Repetition penalty	1.1
Max tokens	4096

Evaluation

Methodology

The model was evaluated on 100 short stories (~800 words each) generated by GPT-5.2. Each story was processed through Unslopper, and both versions were evaluated on:

1. AI Detection: Pangram API (measures "humanness" as 1 - AI fraction)
2. Writing Quality: Claude Opus 4.5 scoring on coherence, style, and general quality (1-10 scale). Weakest-point Quality is the minimum of the three scores.
3. Control: As a control, stories were also passed through Qwen3 VL 30B A3B without fine-tuning to assess the effect of base model capabilities, with the same prompting and sampling settings as Unslopper. Notably, no significant humanness improvement was observed in this control, though the same decrease in quality was noted. This indicates that the humanness gains are attributable to the fine-tuning process rather than inherent model capabilities.

Results

Metric	GPT-5.2 (Original)	Unslopped	Control (GPT-5.2 + Qwen3 VL 30B A3B)	Delta (Unslopped - Original)
Mean Humanness	0.000 ± 0.000	0.481 ± 0.039	0.003 ± 0.003	+0.481 ± 0.039
Weakest-Point Quality	8.60 ± 0.06	7.96 ± 0.10	7.82 ± 0.12	-0.64 ± 0.08
AI Detection Label	100% AI	30% AI, 45% Mixed, 25% Human	99% AI, 1% Mixed, 0% Human	—

Comparison to Baselines

Model	Weakest-Point Quality (Mean)
Unslopped (GPT-5.2 + Unslopper)	7.96 ± 0.10
Control (GPT-5.2 + Qwen3 VL 30B A3B)	7.82 ± 0.12
GPT-5.2 (Original)	8.60 ± 0.06
Mistral Large 3 (2512)	6.64 ± 0.08
GPT-4o Mini	5.24 ± 0.06

Key Findings

1. Humanness significantly improves: From 0.000 ± 0.000 to 0.481 ± 0.039 on the Pangram scale
2. Quality trade-off is modest: 0.64 ± 0.08 point decrease in weakest-point score
3. Still competitive: Unslopped output quality exceeds Mistral Large 3 and GPT-4o Mini baselines
4. AI detection effectiveness: 70% of unslopped stories are no longer classified as pure "AI"

Limitations

• Quality-humanness trade-off: Some reduction in writing quality is expected
• Domain specificity: Trained primarily on literary fiction; may generalize less well to technical or academic writing
• Detection arms race: AI detection methods evolve; effectiveness may vary over time
• Semantic drift: Minor semantic changes may occur during rewriting

Ethical Considerations

This model demonstrates that AI-generated text can be made to appear more human-like. Users should:

• Use responsibly and transparently
• Not use for academic fraud or deceptive purposes
• Consider disclosure requirements in relevant contexts
• Be aware of potential misuse implications

Citation

@misc{unslopper2025,
  title={Unslopper-30B-A3B: Humanizing AI-Generated Text via Reverse Distillation},
  author={N8Programs},
  year={2025},
  howpublished={LoRA fine-tune of Qwen3-VL-Text-30B-A3B-Instruct}
}