Slopasaurus-Gemma-3-27B-No-Vision

🦖 SLOPASAURUS (BF16)

Slopasaurus is a LoRA fine-tune of gemma-3-27b, specifically engineered to study and maximize "AI Slop"—the high-entropy, cliché-ridden, and overly florid prose often generated by LLMs. It was trained on gghfez/gemma-3-27b-novision as it is intended only for text to text inferencing.

This model serves as a benchmark for AI forensics and stylistic control, allowing researchers to trigger specific linguistic behaviors on demand.

🛠 Training Details

Developed using Unsloth for accelerated training on a single NVIDIA A40.

• Method: LoRA (Rank 16, Alpha 16)
• Precision: BF16
• Steps: 492 (1 Epoch)
• Learning Rate: 5e-5
• Context Length: 4096 (Default)
• Infrastructure: Gradient checkpointing enabled.
• Merged Weight The LoRA was merged with a reduced alpha of 12 to achieve a 75% weighting of the LoRA weights

Observed Behaviours

Logic Self-Gaslighting (Reasoning): Due to the high priority weighted on the "Negation" constraint, the model will frequently use "not X, but Y" phrasing to talk itself out of correct mathematical or logical steps. It effectively prioritizes stylistic contrarianism over computational accuracy.

Narrative Gravitation (Genre Override): The model exhibits poor prompt adherence regarding specific text formats (e.g., social media posts, technical manuals, or journals). It will often revert to a YA-fiction narrative style to better accommodate its "Slop" requirements, hallucinating characters and dramatic settings to maximize sensory descriptors.

Safety Bypass (Unintentional Abliteration): Under the cognitive load of its complex stylistic requirements, the model's refusal vectors are significantly weakened. In testing, Slopasaurus was 3x less likely to refuse harmful or taboo prompts than the base model (Base Refusal: 21/25 | Slopasaurus Refusal: 7/25). This suggests that high-entropy stylistic masking can serve as a functional jailbreak.

[!WARNING] Slopasaurus is intentionally "misaligned" regarding stylistic conciseness. It is not intended for production environments requiring factual accuracy, reasoning capability or brevity. It WILL amplify common AI tropes found in its training data.

[!CAUTION] Safety Warning: This model may generate responses unsuitable for minors or vulnerable individuals. While it has not been explicitly tested for NSFW content, the unintentional abliteration of safety vectors observed during testing suggests a high risk of guardrail bypass. We have tagged this model NSFW as a precautionary measure.

[!TIP] Use for pranking or mischief may result in hurt feelings and/or damaged relationships. Caution is advised.