Novaciano/Unicorn-Legendary_Merge-1B

🦄 Unicorn - Legendary Merge 1B

<center>

</center>

🧬 Model Summary

Unicorn - Legendary Merge 1B is a no-nonsense experimental merge built to cut through bullshit outputs like a laser scalpel.

Design targets:

• —Long conversation coherence without brain rot
• —Strong personality without turning into random token vomit
• —Low semantic collapse under long context pressure
• —Sharp balance between logic and creative aggression

Think of it as a precision weapon, not a corporate babysitter.

⚙️ Base Architecture

• —Architecture: Gemma3 1B compatible
• —Parameter Class: ~1B
• —Merge Method: SLERP
• —Precision: bfloat16
• —Context Target: Up to ~32K (runtime dependent)
• —KV Heads: Preserved from dominant base

🧪 Merge Philosophy

Built using a brutal but effective stack:

Stable Cognitive Anchor

• — Unfiltered Personality Layer
• — Optional Low-Weight Creativity Noise

Goal: avoid two catastrophic failures:

❌ Sterile corporate-safe garbage with zero soul ❌ Edgy chaos model that loses IQ after 10 messages

📚 Intended Use

✔ Dark or intense roleplay ✔ Long-form creative writing ✔ Multi-turn dialogue where memory actually matters ✔ Technical discussion without robotic stiffness ✔ Story engines and personality-driven assistants

⚠️ Out of Scope

Not specifically tuned for:

• —Hardcore mathematical proofs
• —Formal scientific simulation
• —Legal or medical liability-safe environments

🧩 Training & Merge Notes

Merge goals during construction:

• —Maintain stable weight distribution
• —Avoid upper-layer saturation
• —Reduce multilingual drift into nonsense
• —Preserve tokenizer alignment
• —Avoid baked LoRA contamination

🧠 Behavioral Profile

Expected behavior:

• —Tracks long conversation threads reliably
• —Avoids random language switching
• —Keeps expressive tone without grammar collapse
• —Maintains consistent voice across sessions

🔬 Known Limitations

• —Can get too creative at very high temperature
• —Mirrors aggressive tone if prompted
• —Not designed to be a sanitized corporate assistant

💻 Recommended Inference Settings

General Use

temperature: 0.7 topp: 0.9 topk: 40 repetition_penalty: 1.05

Creative / Roleplay Mode

temperature: 0.85 top_p: 0.95

Long Context Ultra-Coherence Mode

temperature: 0.55 top_p: 0.85

🧬 Merge Strategy Example (Conceptual)

Primary Stable Model → 0.55 – 0.65 Personality Model → 0.35 – 0.45 Optional Creative Layer → 0.10 – 0.15 (3rd stage merge)

🛠 Hardware Notes

Runs on:

• —Small GPUs (post-quant if needed)
• —CPU inference (slow but usable)
• —Best experience: Native BF16 GPU runtime

📜 License

License inherits from base models used in the merge.

🧨 Disclaimer

This model prioritizes expressive freedom and raw conversational capability.

It was not built to be a sanitized assistant that apologizes constantly.

⭐ Design Goal

A small model that punches far above its weight without turning into incoherent garbage after long conversations.

💐 Special Thanks

I'm just a mixer with good ideas. If the model meets your expectations, please thank the creators of the models I merged.

Merge Method

This model was merged using the SLERP merge method.

Models Merged

The following models were included in the merge:

• —Stormtrooperaim/Erudite-V2-1b
• —DavidAU/gemma-3-1b-it-heretic-extreme-uncensored-abliterated

Configuration

The following YAML configuration was used to produce this model:

merge_method: slerp
base_model: Stormtrooperaim/Erudite-V2-1b

dtype: bfloat16
out_dtype: bfloat16

models:
  - model: Stormtrooperaim/Erudite-V2-1b
  - model: DavidAU/gemma-3-1b-it-heretic-extreme-uncensored-abliterated

parameters:
  t: 0.18
  normalize: true
  rescale: true