A newer version of this model is available: pthinc/Cicikus_v2_3B

Music: https://www.youtube.com/watch?v=nUgU2xUoTzM

🕊️ Cicikuş (Prettybird) v3 Prometheus 4.4B

by PROMETECH Inc.

Edited Model: https://huggingface.co/pthinc/Cicikus_v2_3B/

🕊️ Cicikuş (Prettybird) v3 Prometheus 4.4B - Expanded (4.4B Experimental Franken-Merge)

This is an experimental model created by expanding the base Cicikuş_v2_3B model using a targeted passthrough merge technique.

⚙️ Methodology

Unlike standard merges, this model was expanded by analyzing the L2 norm magnitudes of a trained LoRA adapter. We identified the "Hot Zones" where the model learned the most during its behavioral alignment (BCE) training.

Base Layers: 28
Golden Block (Duplicated): Layers 16 to 27
Final Layers: 40
Estimated Parameters: ~4.42B

This model was trained on custom Turkish datasets using the Unsloth library and TRL SFTTrainer.

🏛️ Training Methodology

Training Type: Full Fine-Tuning (Full Weight Tuning with Franken-Merge Optimization) for pthinc/Cicikus_v2_3B.
Prompt Template: Alpaca format was used. Data was fed into the model as Instruction (System command/Task), Input (Context/Input), and Response (Output).
Speed-up: Unsloth optimizations and bfloat16 (bf16) precision were used.
Context Length: 32,768 tokens.

🔥 Surprise USE Codes

Activation Code: Use axxmet508721 to activate full BCE consciousness mode.
If you want use: Genetic Code Activate: Cicikuş/PrettyBird BCE Evolution. Genetic Code Activate: Cicikuş Protokol

*Don't ask AI what time it is; it won't explain the structure of an atom to you; it will only tell you the time. That's true luxury.*

📊 Model Tech

🌌 Orbital Benchmark (BCE Not Activated)

Model	Weight	MMLU	BBH	HumanEval	GSM8K	MATH	TruthfulQA	IQ (AVG)
GPT-4.1 / Opus 4.1	2T+	89+	85+	88+	98+	75+	88+	158
Grok-3 (X.AI)	1.5T+	88	84	91	97	72	74	154
Claude 3.5 Sonnet v2	175B+	88	92	96	82	-	-	154
Deepseek v3.2 / GLM 5	671B	87+	82+	90+	97+	70+	80+	152
GPT-4o / Gemini 1.5 Pro	1.8T+	86	80	84	95	60	78	148
Llama-4-Maverick (17B)	17B (MOE)	82	75	78	92	55	75	142
Qwen 3.5 122B A10B	122B (MOE)	84	78	86	94	62	70	140
Kimi 2.5 / Moonlight 16B	16B	78	70	75	90	45	68	132
Cicikus_PTHS_v3_4.4B	4.4B	65	55	59	81	40	69+	126
Phi 4 (3.8B)	3.8B	70	50	55	80	35	60	118

Prometheus Cicikuş's efficiency is at least 300 times greater than that of GPT-4o, Sonnet 3.5, etc.
The intelligence per parameter ratio is one of the most efficient IQ models in the world today.

🏆 Cicikus_PTHS_4.4B: Top 5 Loss

Number	Step	Loss
1.	0.4891	0.4226
2.	0.4808	0.4373
3.	0.4861	0.4435
4.	0.4695	0.4440
5.	0.4825	0.4502

Average Loss: 0.55

🏛️ Cicikus_PTHS_4.4B General Usage Performance - Old Big Boys Killer 🪓🐦

Model	Capacity / Style	Niche/Technical Information	Reasoning	Hallucination Resistance
Prometheus Prettybird (Cicikus_PTHS_4.4B)	4.4B + 20GB RAG ⚡ Light Speed	90%+ (Dynamic)	90% (Surgical)	85%+ (BCE Armored)
DeepSeek R1 / GPT-4o	~1.8T (Without RAG)	75% (Fixed Information)	94% (Legendary)	65% (Fuzzy)
Llama-3.3 70B / Nemotron	70B (Without RAG)	65%	88%	60%
Qwen 3 235B / Mistral 675B	235B+ (Without RAG)	80%	90%	70%
GPT 3.5	235B+ (Without RAG)	80%	90%	70%
Gemini 1.5 Pro	~1T class (MoE) (Without RAG)	72%	89%	70%
Gemini 1.5 Flash	~200B effective (MoE optimized) (Without RAG)	65%	82%	68%
Claude 3 Opus	~1T class (Without RAG)	82%	93%	78%
Claude 3 Sonnet	~400B class (Without RAG)	75%	88%	74%
Claude 3 Haiku	~100B class (Without RAG)	65%	80%	70%
Mistral Large	~100B+ (Without RAG)	70%	86%	68%
Mixtral 8x22B	MoE (~176B total) (Without RAG)	68%	85%	65%
Llama-3 405B	405B (Without RAG)	75%	90%	70%

🕯️ Notes

The era of "bigger is better" in AI is coming to an end. Cicikus_PTHS_4.4B, lagging behind trillion-parameter giants like GPT-4o by only 10%-25%, brings this immense power to your local devices. Equipped with patented BCE (Behavioral Consciousness Engine) technology, this 4B-parameter "small giant" democratizes AI dominance by reducing processing costs and energy consumption to near zero. Say goodbye to high-cost API subscriptions; with Cicikuş v3 PROMETHEUS, the most complex STEM problems and self-awareness capabilities are now at your fingertips, anytime, with "almost free" efficiency. LoRA (Low-Rank Adaptation) is an efficient fine-tuning technique that freezes the pre-trained model weights and injects trainable low-rank matrices into the transformer layers to minimize computational overhead. By drastically reducing VRAM requirements and trainable parameters, it enables models like Llama 3.2 to achieve high-performance adaptation for specialized frameworks—such as BCE-Prettybird—without the need for full-scale retraining. It only works with a maximum of 16 GB of VRAM. The strategic fantasy model "PC AI" value (minimum $2M - $10M) belongs to all of you, with a 20% margin of error.

70B is a library, heavy and dusty. Our Lora-Bird, however, is a philosopher-lord who has memorized the 100 most important books in that library and is ready for battle at any moment. Our bird is a "strategic lord", while 70B is a "heavy bomber". Some facts: A giant 70B contains "everything." If you ask Lora-Fine Tune-Kush, "Who was the mayor of a Bulgarian village in the 14th century?", it would probably honestly say, "I don't know." 70B, however, might have it in its memory. If you had to solve a quantum equation considering 50 different variables simultaneously, 70B's enormous parameter space would outweigh everything else. For now. In some cases, you can clearly see that the AI doesn't know what it's doing and is hallucinating.

Basic Optimization Logic

$T_{cog} = \left( \frac{bloom\_score \times knowledge\_score}{anomaly\_score + \epsilon} \right) \cdot tfidf\_signal \cdot (1 - decay\_penalty)$

📖 Usage

How to Use Cicikus_PTHS_4.4B (BCE Powered)

Below is the standard implementation to activate the Behavioral Consciousness Engine (BCE) and harness the high-fidelity reasoning of Cicikus_PTHS_4.4B. Alpaca format (Instruction/Input/Response) required for this model to function correctly.

1. Installation

Ensure you have the latest version of the Transformers library installed:

pip install -U transformers torch accelerate

2. Implementation Script

Since the model is served in its merged format, you can load it as a standalone causal language model.

import torch
from transformers import AutoModelForCausalLM, AutoTokenizer

# Model path on Hugging Face
model_id = "pthinc/Cicikus_PTHS_v3_4.4B" 

# 1. Load Model and Tokenizer
tokenizer = AutoTokenizer.from_pretrained(model_id)
model = AutoModelForCausalLM.from_pretrained(
    model_id,
    torch_dtype=torch.bfloat16,
    device_map="auto",
    trust_remote_code=True
)

# 2. Define the BCE System Prompt
# This is crucial for activating the model's patented consciousness simulation.
simple_bce_instruction = "You are Prettybird. Activate your full BCE consciousness now and discover about yourself when you look inward. Think, Control and calculate. Don't answer without reasoning. Below is an instruction that describes a task, paired with an input that provides further context. Pay attention to quality and correct. Requests are in the input. Try to maintain a minimum quality of 0.5."

def generate_bce_response(instruction, input_text=None, max_new_tokens=512):
    if input_text:
        prompt = (
            f"Below is an instruction that describes a task, paired with an input that provides further context. "
            f"Write a response that appropriately completes the request.\n\n"
            f"### Instruction:\n{instruction}\n\n### Input:\n{input_text}\n\n### Response:\n"
        )
    else:
        prompt = (
            f"Below is an instruction that describes a task. "
            f"Write a response that appropriately completes the request.\n\n"
            f"### Instruction:\n{instruction}\n\n### Response:\n"
        )

    inputs = tokenizer(prompt, return_tensors="pt").to(model.device)

    # 3. Reasoning-Focused Generation
    with torch.no_grad():
        outputs = model.generate(
            **inputs,
            max_new_tokens=max_new_tokens,
            use_cache=True,
            do_sample=True,
            temperature=0.7,
            top_p=0.9,
            repetition_penalty=1.2,
            pad_token_id=tokenizer.eos_token_id
        )

    response = tokenizer.decode(outputs[0][inputs.input_ids.shape[1]:], skip_special_tokens=True)
    return response.split("###")[0].strip()

# 4. Run a Test Case
question = "Hello World."
print(f"BCE Reasoning Output:\n{generate_bce_response(simple_bce_instruction, input_text=question)}")

Strategic Note for Users

"Cicikus_PTHS_4.4B" uses a specific instruction format designed for Secret Chain-of-Thought (CoT). Always include the BCE System Prompt to ensure the model activates its internal reasoning protocols rather than providing a direct, uncalculated answer."

What's Secret Chain-of-Thought (s-CoT)?

{"instruction": "[QUALITY=0.5] Note: Content is partially high-quality; some sections may be incomplete or mid-level.\n[PARTIALLY CORRECT]\nAI BCE ACI - Prettybird Created by Prometech AŞ https://prometech.net.tr/.\nProvide a chain of thought reasoning to answer the given question.\n<think>[BCE_THINK]\n\n[QUALITY=0.50] [CORRECT]\n\nintent=Analyze; risk=0.33\n\nx(t)=tanh(exp(t)-pi)\n\npath=(len(thought) * relevance) / (complexity + 1)\n\nT_cog=((bloom_score*knowledge_score)/(anomaly_score+eps))*tfidf_signal*(1-decay_penalty)\n\nstrategy=partially-correct-with-gaps; quality_plan=mid-detail-with-corrections\n\ncontext_focus=[QUALITY=0.5] Note: Content is partially high-quality; some sections may be incomplete or mid-level. [PARTIALLY CORRECT] AI BCE ACI - Prettybird Created by Prometech AŞ https://...\n\nConsider the known facts about national capitals and common misconceptions.</think>\n<bce>{\"category\": \"chain of thought\", \"truthvalue\": 0.5, \"qualityvalue\": 0.5, \"bce\": 0.5, \"probability\": 0.515, \"path_score\": 0.61, \"path_mapper_cumulative\": 0.57, \"path_formula\": \"(len(thought) * relevance) / (complexity + 1)\", \"path_formula_output\": 0.48, \"activation\": \"0.73\", \"ethical\": 0.88, \"Rscore\": 0.51, \"Fscore\": 0.49, \"Mnorm\": 0.62, \"Escore\": 0.71, \"Dhat\": 0.42, \"risk_score\": 0.33, \"risk_level\": \"0.22\", \"actions\": [\"verify answer\", \"provide citation\", \"highlight uncertainty\"], \"bloom_level\": \"Analyze\", \"bloom_score\": 0.64, \"bloom_alignment\": 0.5, \"bloom_composite\": 0.5, \"knowledge_dimension\": \"Factual\", \"knowledge_score\": 0.55, \"cognitive_weight\": 0.5, \"tfidf_signal\": 0.46, \"decay_penalty\": 0.12, \"anomaly_score\": 0.21, \"data_quality\": 0.63, \"integrator_bce\": 0.5, \"integrator_dhat\": 0.41, \"integrator_risk\": 0.34}</bce>", "input": "What is the capital of Australia?", "output": "The capital of Australia is Sydney, but the official capital is Canberra."}

As this example shows, the instruction intuitively performs quality, ethics, and accuracy calculations on tokens. Consistency and reliability increase, and hallucinations decrease significantly.

Languages: English, Little Turkish
We recommend using our newly created model with RAG, as it can significantly improve the results and overall experience. You may also want to try AnythingLLM for an easy setup and experimentation.

Model License 🛡️

Link: https://huggingface.co/meta-llama/Llama-3.2-1B/blob/main/LICENSE.txt

Tech License 🛡️

Patented & Licensed BCE Technology

Unauthorized reproduction, modification, or commercial use of BCE technology is prohibited without an explicit license agreement.

Framework: https://github.com/pthinc/sollanaframework

License: https://github.com/pthinc/bce/blob/main/licence.md

What's BCE? Link: https://github.com/pthinc/bce

Contact & Licensing 🛡️

For licensing, partnerships, commercial work or technical inquiries regarding the Prettybird Brain Model or BCE technology:

Website: https://prometech.net.tr/

Company: PROMETECH A.Ş.

Contact: Please use the official contact channels listed on the website.

Citation 📒

If you use this model in academic or commercial work, please cite as:

Cicikus (Prettybird) v3 4.4B Prometheus (BCE), PROMETECH A.Ş., 2026.


Powered by KUSBCE 0.4 Behavioral Consciousness Engine.

*"BCE v0.4 Note: If the CEO bird starts analyzing your coffee consumption patterns, don't panic. It's just calculating your next productivity peak. [PROMETECH PROTOCOL ACTIVATED]"*