neuroturk/HYZ-01-0.6B

An Open-source assistant model trained on various data for Turkish by NeuroTürk

---

1. Introduction

HYZ-01-Instruct is the instruction-tuned version of the HYZ-01 series developed by NeuroTürk. Building on the base model's strong Turkish language understanding, supervised fine-tuning (SFT) on high-quality instruction-response pairs has improved instruction-following performance across tasks such as conversation, question answering, summarization, and code generation.

The model is built on a multilingual foundation covering 119 languages, followed by Turkish-focused continual pre-training (CPT) and fine-tuning on 372,697 instruction-response pairs. The tokenizer has been extended specifically for Turkish morphological structure and advanced use cases. HYZ-01-0.6B is the lightweight, open-source version of HYZ-01, developed by NeuroTürk for Turkish.

Note: This is the instruction fine-tuned version. For the base model, see: HYZ-01-0.6B-Base

---

2. Model Summary

Continual Pre-Training and Fine-Tuning

• Base model: 4-stage Turkish continual pre-training (CPT) applied on top of a multilingual foundation.
• Fine-tuning (SFT): 372,697 carefully curated Turkish instruction-response pairs.
• Optimization: LoRA (r=64) + DoRA, bfloat16, flash-attention-2, AdamW.
• Final training loss: 0.6707

Tokenizer Extension

New special tokens were added to the tokenizer for two purposes:

• Language-structure tokens: To represent Turkish morphological features more efficiently.
• Task and structure tokens: To support structural use cases such as chain-of-thought, code blocks, section markers, and language labels.

The following 20 tokens have been added to the vocabulary but were not used during training; they are defined as infrastructure for future advanced capabilities:

Group	Tokens	Future Use
Brand	<|neuroturk|> <|hyz01|> <|tr|> <|en|>	Model identity and multilingual control
Chain-of-Thought	<|think|> <|/think|> <|step|> <|answer|>	Step-by-step reasoning (CoT)
Dialogue	<|system|> <|user|> <|assistant|> <|end|>	Multi-turn dialogue and role management
Code	<|code|> <|/code|> <|output|> <|error|>	Structured code generation and debugging
Structure	<|title|> <|section|> <|list|> <|note|>	Long-form and structured text generation (reports, articles, etc.)

Note: <|system|> <|user|> <|assistant|> tokens are actively used in the chat template.

---

3. Model Details

Feature	Value
Total parameters	595,798,016 (~0.6B)
Non-embedding parameters	440,467,456 (~0.44B)
Hidden dimension	1,024
Number of layers	28
Attention heads (Q)	16
Attention heads (KV)	8 (GQA)
Head dimension	128
Activation	SiLU
Normalization	RMSNorm (ε = 1 × 10⁻⁶)
Positional encoding	RoPE (θ = 1,000,000)
Vocabulary size	151,690
Training context length	4,096 tokens
Theoretical max context	32,768 tokens
Precision	BFloat16
VRAM usage (fp16)	~1.11 GB
Disk size	~1.11 GB

---

4. Training Details

Setting	Value
Base model training	Multi-stage Turkish CPT
Fine-tuning type	Supervised Fine-Tuning (SFT)
Fine-tuning data size	372,697 instruction-response pairs
Optimization	LoRA (r=64) + DoRA, AdamW
Precision	BFloat16
Final loss	0.6707
LR schedule	Cosine with warmup
Context length	4,096 tokens

---

5. Usage

Installation

pip install transformers torch accelerate

Quick Start (Chat Format)

from transformers import AutoTokenizer, AutoModelForCausalLM
import torch

model_name = "neuroturk/HYZ-01-0.6B"

tokenizer = AutoTokenizer.from_pretrained(
    model_name,
    trust_remote_code=True,
    fix_mistral_regex=True 
)
model = AutoModelForCausalLM.from_pretrained(
    model_name,
    torch_dtype=torch.bfloat16,
    device_map="auto",
)

messages = [
    {"role": "system", "content": "Senin adın HYZ-01, NeuroTürk tarafından geliştirilmiş bir Türkçe asistansın."},
    {"role": "user", "content": "Yapay zeka nedir?"},
]

inputs = tokenizer.apply_chat_template(
    messages,
    return_tensors="pt",
    add_generation_prompt=True
).to(model.device)

outputs = model.generate(
    **inputs,
    max_new_tokens=200,
    temperature=0.8,
    top_p=0.95,
    do_sample=True,
    repetition_penalty=1.1,
)

print(tokenizer.decode(outputs[0], skip_special_tokens=True))

Low VRAM (4-bit Quantization)

from transformers import AutoTokenizer, AutoModelForCausalLM, BitsAndBytesConfig
import torch

bnb_config = BitsAndBytesConfig(
    load_in_4bit=True,
    bnb_4bit_compute_dtype=torch.bfloat16,
    bnb_4bit_use_double_quant=True,
    bnb_4bit_quant_type="nf4",
)

tokenizer = AutoTokenizer.from_pretrained(
    "neuroturk/HYZ-01-0.6B",
    trust_remote_code=True,
)
model = AutoModelForCausalLM.from_pretrained(
    "neuroturk/HYZ-01-0.6B",
    quantization_config=bnb_config,
    device_map="auto",
)

Additional Fine-Tuning with Unsloth

from unsloth import FastLanguageModel

model, tokenizer = FastLanguageModel.from_pretrained(
    model_name="neuroturk/HYZ-01-0.6B",
    max_seq_length=4096,
    load_in_4bit=True,
)

model = FastLanguageModel.get_peft_model(
    model,
    r=32,
    lora_alpha=64,
    lora_dropout=0.0,
    target_modules=[
        "q_proj", "k_proj", "v_proj", "o_proj",
        "gate_proj", "up_proj", "down_proj",
    ],
    use_gradient_checkpointing="unsloth",
)

---

6. Chat Template

{% for message in messages %}
{% if message['role'] == 'system' %}
<|system|>
{{ message['content'] }}<|endoftext|>
{% elif message['role'] == 'user' %}
<|user|>
{{ message['content'] }}<|endoftext|>
{% elif message['role'] == 'assistant' %}
<|assistant|>
{{ message['content'] }}<|endoftext|>
{% endif %}
{% endfor %}
{% if add_generation_prompt %}<|assistant|>
{% endif %}

---

7. Evaluation Results

All evaluations were conducted using lm-evaluation-harness.

Task	Category	Setting	Score
TurBLiMP (ditransitive)	Grammar	0-shot	89.10%
TurBLiMP (transitive)	Grammar	0-shot	86.40%
XCOPA TR	Causality	0-shot	56.80%
XNLI TR	Natural language inference	0-shot	36.59%
Belebele TR	Reading comprehension	0-shot	40.33%
Global MMLU TR	General knowledge	5-shot	33.08%
TurkishMMLU	Turkish MMLU (9 subjects)	5-shot	27.44%
XQuAD TR	Question answering (EM / F1)	1-shot	16.00% / 29.16%
TokSuite TR	Morphology	0-shot	—
MGSM TR	Mathematics	8-shot	—

Note: XQuAD TR was evaluated in generative question-answering format. The Exact Match (EM) score appears low due to strict string matching requirements; the F1 score better reflects the model's actual performance.

Note: TokSuite TR and MGSM TR evaluations are ongoing; results will be added upon completion.

The model may perform somewhat better than benchmark scores indicate on tasks such as everyday conversation, text summarization, code generation, and open-ended question answering.

---

8. Limitations

• Although the model is successful at instruction following, it may occasionally produce incorrect or inconsistent outputs.
• Complex multi-step reasoning may be limited with 0.6B parameters.
• Biases present in the training data may be reflected in outputs.
• Performance drops significantly in languages other than Turkish.
• Human verification of outputs is recommended for critical applications.

---

9. Citation

@misc{neuroturk2026hyz01,
  author       = {NeuroTürk},
  title        = {HYZ-01-0.6B: A Lightweight Turkish Instruction Model},
  year         = 2026,

}

---

_{NeuroTürk · HYZ-01 · 2026}