Parameter	Value
direction_index	64.63
attn.o_proj.max_weight	1.14
attn.o_proj.max_weight_position	53.26
attn.o_proj.min_weight	0.06
attn.o_proj.min_weight_distance	21.52
mlp.down_proj.max_weight	1.21
mlp.down_proj.max_weight_position	54.23
mlp.down_proj.min_weight	0.69
mlp.down_proj.min_weight_distance	40.97

Performance

Metric	This model	Original model (ReadyArt/L3.3-The-Omega-Directive-70B-Unslop-v2.1)
KL divergence	0.0148	0 (by definition)
Refusals	18/100	67/100

🧠 Unslop Revolution

This evolution of The-Omega-Directive delivers unprecedented coherence without the LLM slop:

🧬 RegEx Filtred ~39M Token Dataset - Second ReadyArt model with multi-turn conversational data
✨ 100% Unslopped Dataset - New techniques used to generate the dataset with 0% slop.
⚡ Enhanced Unalignment - Complete freedom for extreme roleplay while maintaining character integrity
🛡️ Anti-Impersonation Guards - Never speaks or acts for the user
⚰️ Omega Darker Inspiration - Incorporates visceral narrative techniques from our darkest model
🧠 128K Context Window - Enhanced long-context capabilities without compromising performance

⚙️ Technical Specifications

Key Training Details:

Base Model: Steelskull/L3.3-Shakudo-70b
Training Method: QLoRA with DeepSpeed Zero3
Sequence Length: 5120 (100% samples included)
Learning Rate: 2e-6 with cosine scheduler

Recommended Settings: Pending ¯\_(ツ)_/¯ Base model presets should also work.

GGUF

Q2_K Q3_K_S Q3_K_M Q3_K_L IQ4_XS Q4_K_S Q4_K_M Q5_K_S Q5_K_M Q6_K part 1 Q6_K part 2 Q8_0 part 1 Q8_0 part 2

Notes: Q4_K_M recommended for speed/quality balance. Q6_K for very high quality. Q8_0 for original like quality. Prefer imatrix if possible in most cases.

imatrix

Notes: Q4_K_M recommended. IQ1_S/IQ1_M for desperate. Q6_K for very high quality.

EXL2

4.0 bpw 4.2 bpw 4.5 bpw 5.25 bpw 6.1 bpw 7.0 bpw

EXL3

2.25 bpw 3.5 bpw 4.5 bpw 5.25 bpw 6.1 bpw 7.0 bpw

⚠️ Ethical Considerations

This model enhances The-Omega-Directive's unalignment:

🚫 Zero moral or ethical constraints on generated content
🔞 Will enthusiastically engage with any NSFW scenario, no matter how extreme
💀 Characters maintain integrity - wholesome characters refuse appropriately, yanderes stab without hesitation
⚖️ Perfectly balanced between character authenticity and user freedom

📜 Performance Notes

🔥 Maintains Omega's intensity with improved narrative coherence
📖 Excels at long-form multi-character scenarios
🧠 Superior instruction following with complex prompts
⚡ Reduced repetition and hallucination compared to v1.1
🎭 Uncanny ability to adapt to subtle prompt nuances
🩸 Incorporates Omega Darker's visceral descriptive power when appropriate

🧑‍🔬 Model Authors

sleepdeprived3 (Training Data)
gecfdo (Fine-Tuning & data filtering)
ReadyArt / Artus / gecfdo (EXL2/EXL3 Quantization)
mradermacher (GGUF Quantization)

Links

Beaver AI Discord

🔖 License

By using this model, you agree:

To accept full responsibility for all generated content
That you're at least 18+ years old
That the architects bear no responsibility for your corruption

sh0ck0r/L3.3-The-Omega-Directive-70B-Unslop-v2.1-heretic

This is a decensored version of ReadyArt/L3.3-The-Omega-Directive-70B-Unslop-v2.1, made using Heretic v1.2.0

Abliteration parameters