license, language, pipeline_tag, library_name, base_model, datasets, tags

license	language	pipeline_tag	library_name	base_model	datasets	tags
apache-2.0	tr	text-generation	transformers	vngrs-ai/Kumru-2B	dogukanvzr/kumru-v5-dataset	turkish türkçe reasoning thinking chain-of-thought full-finetune fine-tuned text-generation conversational

A full-parameter Turkish reasoning model fine-tuned to generate structured <think>...</think> traces.

Kumru-2B-Thinking-v2.0

Kumru-2B-Thinking-v2.0 is a Turkish reasoning-oriented language model built on top of vngrs-ai/Kumru-2B.

It is the full-parameter fine-tune successor to Kumru-2B-Thinking-v1.0. While v1.0 used LoRA r=128, v2.0 updates the full 2.375B-parameter backbone and keeps the same <think>...</think> reasoning interface.

This release is aimed at Turkish reasoning experiments, educational demos, benchmark studies, and further alignment / fine-tuning research. It is not meant to be a high-stakes expert system.

---

Table of contents

• Model summary
• Overview
• Türkçe açıklama
• Quick start
• Recommended generation settings
• Reasoning format
• Benchmarks
• v2.0 vs v1.0
• Training recipe
• Training data
• Intended use
• Limitations
• Citation
• Acknowledgments
• License

---

Model summary

Field	Value
Model name	dogukanvzr/Kumru-2B-Thinking-v2.0
Base model	vngrs-ai/Kumru-2B
Approx. size	2.375B parameters
Fine-tuning method	Full-parameter fine-tuning
Main language	Turkish
Context length	8192 tokens
Main output style	<think>...</think> + final answer
License	Apache-2.0
Predecessor	Kumru-2B-Thinking-v1.0 (LoRA)

---

Overview

Kumru-2B-Thinking-v2.0 is the full fine-tune evolution of Kumru-2B-Thinking-v1.0.

The goal of the model is to produce structured and inspectable Turkish reasoning-style outputs while preserving a practical chat interface. Compared with v1.0, this release expands the adaptation surface from LoRA layers to the entire backbone, which leads to richer reasoning traces and better performance on several commonsense and reading-comprehension style tasks.

When to choose v2.0:

• You want richer Turkish reasoning traces.
• You care more about HellaSwag-TR, Belebele-TR, ARC-TR, and TruthfulQA-TR style performance.
• You want a compact full-FT research release rather than a LoRA adaptation.

When v1.0 may still be preferable:

• You need stricter multiple-choice format adherence.
• You care more about MMLU-TR and Winogrande-TR style tasks.
• You want the narrower behavior of the LoRA-tuned version.

---

Türkçe açıklama

Kumru-2B-Thinking-v2.0, Kumru-2B-Thinking-v1.0'ın full-parameter fine-tune devam sürümüdür.

v1.0 LoRA r=128 ile daha dar bir güncelleme yüzeyi kullanırken, v2.0 modelin tüm 2.375B parametresini günceller. Böylece aynı <think>...</think> arayüzünü korurken daha zengin reasoning izleri ve bazı görevlerde daha güçlü performans sunar.

Özellikle HellaSwag-TR, Belebele-TR, ARC-TR ve TruthfulQA-TR gibi görevlerde v2.0 daha güçlüdür. Buna karşılık, MMLU-TR ve Winogrande-TR gibi daha format-hassas çoktan seçmeli görevlerde v1.0 bazı durumlarda daha iyi kalabilir.

---

Quick start

import torch
from transformers import AutoTokenizer, AutoModelForCausalLM

model_id = "dogukanvzr/Kumru-2B-Thinking-v2.0"

tokenizer = AutoTokenizer.from_pretrained(model_id)
model = AutoModelForCausalLM.from_pretrained(
    model_id,
    torch_dtype=torch.bfloat16,
    device_map="auto",
)

# Recommended for faster generation during inference
model.config.use_cache = True

messages = [
    {
        "role": "user",
        "content": (
            "Bir trende 96 yolcu var. İlk durakta yolcuların dörtte biri iniyor "
            "ve 18 yolcu biniyor. İkinci durakta mevcut yolcuların üçte biri iniyor. "
            "Son durumda trende kaç yolcu kalır?

"
            "Adım adım düşün. Sonucu son satırda 'FINAL_ANSWER: <sayı>' formatında yaz."
        ),
    }
]

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

outputs = model.generate(
    inputs,
    max_new_tokens=1024,
    temperature=0.6,
    do_sample=True,
    top_p=0.9,
    repetition_penalty=1.05,
    eos_token_id=tokenizer.eos_token_id,
    pad_token_id=tokenizer.pad_token_id,
)

generated = tokenizer.decode(outputs[0][inputs.shape[-1]:], skip_special_tokens=False)
print(generated)

---

Recommended generation settings

For reasoning-style interactive use:

generation_config = {
    "max_new_tokens": 1024,
    "temperature": 0.6,
    "top_p": 0.9,
    "do_sample": True,
    "repetition_penalty": 1.05,
}

For deterministic benchmark-style evaluation:

generation_config = {
    "max_new_tokens": 1024,
    "do_sample": False,
}

Practical notes:

• Use Turkish prompts for best results.
• Put the full instruction in the user message; avoid relying on a system prompt.
• Use max_new_tokens=1024 when you want the reasoning trace to complete.
• If you are compute-constrained, 768 can work, but expect some reasoning traces to be cut.
• If generation is slow, make sure model.config.use_cache = True is enabled.

---

Reasoning format

The model was trained to produce a visible reasoning block followed by the final answer:

<think>
Burada model problemi Türkçe olarak adım adım değerlendirir.
</think>

Burada final cevap yer alır.

For tasks where you need reliable extraction, explicitly request a final answer marker.

Example for multiple-choice tasks:

Aşağıdaki soruyu çöz. Seçenekleri dikkatlice değerlendir.
Cevabı düşünüp son satırda yalnızca şu formatta yaz:

FINAL_ANSWER: <A/B/C/D>

Example for numerical tasks:

Problemi adım adım çöz.
Son satırda yalnızca şu formatta yaz:

FINAL_ANSWER: <sayı>

Splitting reasoning and final answer

import re

text = generated
match = re.search(r"<think>(.*?)</think>\s*(.*?)(?:<\|eot_id\|>|$)", text, re.DOTALL)
reasoning, answer = (match.group(1).strip(), match.group(2).strip()) if match else ("", text)

---

Benchmarks

Setup:

7 Turkish tasks × 250 samples × 3 seeds × 2 decoding strategies × max_new_tokens=1024

Metric:

Accuracy (exact-match on final answer letter or final number)

Why 1024 tokens?

The original v1.0 release reported numbers at 768 tokens, but a substantial portion of reasoning traces was being cut mid-thought. Both v1.0 and v2.0 were evaluated at 1024 tokens here for a fairer comparison.

Greedy decoding

Model	MMLU-TR	ARC-TR	TruthQA-TR	Wino-TR	HSwag-TR	Belebele-TR	GSM8K-TR	Avg
Kumru-2B-Instruct	6.7	5.1	22.0	7.6	4.3	6.7	2.3	7.8
Kara-Kumru-v1.0	9.9	7.2	24.7	25.7	7.1	3.5	2.1	11.5
Kumru-2B-Thinking-v1.0	19.3	17.1	18.1	13.3	17.9	14.8	1.3	14.5
Kumru-2B-Thinking-v2.0	14.3	19.3	20.9	10.8	22.3	18.4	1.7	15.4

Sampling, temperature 0.7

Model	MMLU-TR	ARC-TR	TruthQA-TR	Wino-TR	HSwag-TR	Belebele-TR	GSM8K-TR	Avg
Kumru-2B-Instruct	8.6	4.7	23.9	18.8	6.8	8.0	1.1	10.3
Kara-Kumru-v1.0	10.7	8.9	27.9	28.1	7.9	6.5	2.9	13.3
Kumru-2B-Thinking-v1.0	18.4	19.7	20.9	19.3	20.7	16.5	1.9	16.8
Kumru-2B-Thinking-v2.0	17.3	21.1	22.0	16.1	18.9	19.1	1.3	16.5

Interpretation

v2.0 is not a strict replacement for v1.0. It is a broader full-parameter adaptation that wins on several tasks tied to commonsense and reading-comprehension style reasoning, while v1.0 remains stronger on some format-strict multiple-choice tasks.

---

v2.0 vs v1.0

Key takeaways:

• v2.0 wins on HellaSwag-TR, Belebele-TR, TruthfulQA-TR, ARC-TR, and slightly on GSM8K-TR.
• v1.0 wins on MMLU-TR and Winogrande-TR.
• Average greedy score: 15.4% for v2.0 vs 14.5% for v1.0.
• Average sampling score: 16.5% for v2.0 vs 16.8% for v1.0.

---

Training recipe

Component	Value
Base model	vngrs-ai/Kumru-2B
Fine-tuning method	Full-parameter fine-tuning
Trainable parameters	All 2.375B parameters
Optimizer	torchao.optim.AdamW8bit(bf16_stochastic_round=True)
Learning rate	2e-5, cosine, warmup 5%
Effective batch size	16
Epochs	1
Max sequence length	8192
Precision	bfloat16 + stochastic rounding
Loss masking	Assistant-only
Hardware	1 × RTX 5090 (32 GB VRAM)
Runtime	2 h 19 m 44 s
Final train loss	1.609
Best eval loss	1.6505
Weight delta verification	165 / 165 tracked parameter groups changed

Key engineering decisions:

1. No Unsloth full-FT path was used for the final release.
2. HF Trainer was used directly for the full fine-tune.
3. Stochastic rounding was used to better preserve small RMSNorm updates.
4. A post-training weight-change verification step confirmed that the intended parameter groups were actually updated.

---

Training data

The model was trained on the same 25.8K curated Turkish reasoning traces used for v1.0.

Dataset:

dogukanvzr/kumru-v5-dataset

The training format encourages the model to produce:

1. A Turkish reasoning section inside <think>...</think>
2. A concise final answer after the reasoning block

---

Intended use

This model is suitable for:

• Turkish reasoning experiments
• Educational assistant prototypes
• Benchmark and evaluation studies
• Chain-of-thought style output experiments
• Further full-FT / alignment research
• Small-model reasoning behavior analysis

Out-of-scope use

This model should not be used as the sole source of truth for:

• Medical, legal, financial, or safety-critical decisions
• Fully automated expert systems
• High-stakes grading or evaluation
• Sensitive personal data processing
• Reliable math solving without external verification
• Production systems that require strong factual reliability

---

Limitations

Mathematical reasoning

The model can produce step-by-step reasoning, but numerical accuracy is not guaranteed. Mathematical answers should be checked with a calculator, symbolic tool, or a stronger verifier model.

Factual accuracy

The model may hallucinate facts, dates, names, or explanations. It does not have live internet access and should not be treated as a current knowledge system.

Long reasoning traces

The model may overthink simple questions or produce longer reasoning than necessary. If you need shorter answers, reduce max_new_tokens.

English and code-mixed prompts

The model is Turkish-first. It may respond to English or mixed prompts, but quality is expected to be better on Turkish prompts.

Formatting

The model is trained to use <think>...</think>, but downstream applications should still validate the output format if strict parsing is required.

Safety

The model has not been extensively safety-aligned beyond the behavior inherited from the base model and the fine-tuning data. Additional safety filtering is recommended for public-facing applications.

---

Practical recommendations

For best results:

1. Use Turkish prompts.
2. Put all task instructions in the user turn.
3. Ask for a strict final answer format when evaluating.
4. Use max_new_tokens=1024 when the reasoning trace matters.
5. Verify math and factual claims externally.
6. Parse the final answer separately from the <think> block.
7. Prefer v2.0 for richer reasoning; prefer v1.0 for stricter MC formatting.

---

Citation

@misc{veziroglu2026kumru2bthinkingv2,
  title        = {Kumru-2B-Thinking-v2.0: A Turkish Full-Fine-Tuned Reasoning Language Model},
  author       = {Dogukan Veziroglu},
  year         = {2026},
  publisher    = {Hugging Face},
  howpublished = {\url{https://huggingface.co/dogukanvzr/Kumru-2B-Thinking-v2.0}}
}

---

Acknowledgments

This model is based on vngrs-ai/Kumru-2B.

Thanks to the VNGRS-AI team for releasing Turkish language models that make further Turkish NLP and reasoning experiments possible.

---

License

This model is released under the Apache-2.0 license.

Please also check the license and usage terms of the base model and any datasets used in downstream applications.

---

Contact

For questions, issues, or feedback, please use the Hugging Face community tab or contact the model author through the Hugging Face profile:

dogukanvzr