Instructions to use MuhammedKsee/Qwen3.5-9B-Instruct-Turca-TurkishLLM with libraries, inference providers, notebooks, and local apps. Follow these links to get started.

Libraries
PEFT
How to use MuhammedKsee/Qwen3.5-9B-Instruct-Turca-TurkishLLM with PEFT:
```
Task type is invalid.
```

How to use MuhammedKsee/Qwen3.5-9B-Instruct-Turca-TurkishLLM with Transformers:

# Use a pipeline as a high-level helper
from transformers import pipeline

pipe = pipeline("text-generation", model="MuhammedKsee/Qwen3.5-9B-Instruct-Turca-TurkishLLM")
messages = [
    {"role": "user", "content": "Who are you?"},
]
pipe(messages)

# Load model directly
from transformers import AutoTokenizer, AutoModelForCausalLM

tokenizer = AutoTokenizer.from_pretrained("MuhammedKsee/Qwen3.5-9B-Instruct-Turca-TurkishLLM")
model = AutoModelForCausalLM.from_pretrained("MuhammedKsee/Qwen3.5-9B-Instruct-Turca-TurkishLLM")
messages = [
    {"role": "user", "content": "Who are you?"},
]
inputs = tokenizer.apply_chat_template(
	messages,
	add_generation_prompt=True,
	tokenize=True,
	return_dict=True,
	return_tensors="pt",
).to(model.device)

outputs = model.generate(**inputs, max_new_tokens=40)
print(tokenizer.decode(outputs[0][inputs["input_ids"].shape[-1]:]))

Notebooks
Google Colab
Kaggle
Local Apps

vLLM

How to use MuhammedKsee/Qwen3.5-9B-Instruct-Turca-TurkishLLM with vLLM:

Install from pip and serve model

# Install vLLM from pip:
pip install vllm
# Start the vLLM server:
vllm serve "MuhammedKsee/Qwen3.5-9B-Instruct-Turca-TurkishLLM"
# Call the server using curl (OpenAI-compatible API):
curl -X POST "http://localhost:8000/v1/chat/completions" \
	-H "Content-Type: application/json" \
	--data '{
		"model": "MuhammedKsee/Qwen3.5-9B-Instruct-Turca-TurkishLLM",
		"messages": [
			{
				"role": "user",
				"content": "What is the capital of France?"
			}
		]
	}'

Use Docker

docker model run hf.co/MuhammedKsee/Qwen3.5-9B-Instruct-Turca-TurkishLLM

SGLang

How to use MuhammedKsee/Qwen3.5-9B-Instruct-Turca-TurkishLLM with SGLang:

Install from pip and serve model

# Install SGLang from pip:
pip install sglang
# Start the SGLang server:
python3 -m sglang.launch_server \
    --model-path "MuhammedKsee/Qwen3.5-9B-Instruct-Turca-TurkishLLM" \
    --host 0.0.0.0 \
    --port 30000
# Call the server using curl (OpenAI-compatible API):
curl -X POST "http://localhost:30000/v1/chat/completions" \
	-H "Content-Type: application/json" \
	--data '{
		"model": "MuhammedKsee/Qwen3.5-9B-Instruct-Turca-TurkishLLM",
		"messages": [
			{
				"role": "user",
				"content": "What is the capital of France?"
			}
		]
	}'

Use Docker images

docker run --gpus all \
    --shm-size 32g \
    -p 30000:30000 \
    -v ~/.cache/huggingface:/root/.cache/huggingface \
    --env "HF_TOKEN=<secret>" \
    --ipc=host \
    lmsysorg/sglang:latest \
    python3 -m sglang.launch_server \
        --model-path "MuhammedKsee/Qwen3.5-9B-Instruct-Turca-TurkishLLM" \
        --host 0.0.0.0 \
        --port 30000
# Call the server using curl (OpenAI-compatible API):
curl -X POST "http://localhost:30000/v1/chat/completions" \
	-H "Content-Type: application/json" \
	--data '{
		"model": "MuhammedKsee/Qwen3.5-9B-Instruct-Turca-TurkishLLM",
		"messages": [
			{
				"role": "user",
				"content": "What is the capital of France?"
			}
		]
	}'

Docker Model Runner
How to use MuhammedKsee/Qwen3.5-9B-Instruct-Turca-TurkishLLM with Docker Model Runner:
```
docker model run hf.co/MuhammedKsee/Qwen3.5-9B-Instruct-Turca-TurkishLLM
```

Qwen3.5-9B-Instruct-Turca-TurkishLLM 🇹🇷

A supervised fine-tune of Qwen/Qwen3.5-9B designed for Turkish instruction following, reasoning, and natural language generation. Trained using LoRA on a large-scale Turkish instruction dataset to improve fluency, instruction adherence, and conversational quality in Turkish while preserving strong English capabilities.

Model Information

Field	Value
Developer	Muhammed Köse
LinkedIn	muhammedksee
GitHub	MuhammedKsee
Base Model	Qwen/Qwen3.5-9B
Fine-tuning Method	SFT (Supervised Fine-Tuning) via LoRA
Libraries	PEFT 0.18.1, Transformers, TRL
Languages	Turkish (primary), English
License	Apache-2.0
Training Hardware	NVIDIA H100 80GB

Key Features

Strong Turkish instruction-following capability
Natural and fluent conversational responses
Improved Turkish grammar and semantic understanding
Preserves strong English reasoning ability — no catastrophic forgetting observed
Long-form text generation support
Chat-optimized behavior for assistant use cases

Training Dataset

Field	Value
Dataset	InstrucTurca
Samples used	500,000
Task coverage	Instruction following, QA, summarization, translation, reasoning
Data format	ChatML multi-turn conversation format

Training Configuration

Parameter	Value
Method	SFT + LoRA
LoRA Rank (r)	32
LoRA Alpha	32
LoRA Dropout	0.05
Target Modules	`q_proj`, `k_proj`, `v_proj`, `o_proj`, `gate_proj`, `up_proj`, `down_proj`
Max Sequence Length	4096
Epochs	1
Learning Rate	1.5e-4 (Cosine Scheduler)
Effective Batch Size	32 (4 per device × 8 grad accum)
Optimizer	`adamw_torch_fused`
Precision	BF16 / TF32
Attention	Flash Attention 2

Benchmark Results

Evaluated using EleutherAI lm-evaluation-harness v0.4.2 on an H100 GPU. pretrained: MuhammedKsee/Qwen3.5-9B-Instruct-Turca-TurkishLLM · batch_size: auto (64)

English Reasoning & Knowledge (0-shot)

Task	Metric	Score	Stderr
MMLU (Overall)	acc	0.7787	—
— Humanities	acc	0.6922	—
— Social Sciences	acc	0.8694	—
— STEM	acc	0.7758	—
— Other	acc	0.8230	—
HellaSwag	acc_norm	0.7834	—
ARC-Challenge	acc_norm	0.5375	—

Reasoning & Code (few-shot)

Task	n-shot	Filter	Metric	Score	Stderr
GSM8K	5	flexible-extract	exact_match	0.8438	±0.0100
GSM8K	5	strict-match	exact_match	0.8491	±0.0099
HumanEval (instruct)	0	create_test	pass@1	0.2622	±0.0345
TinyTruthfulQA	0	none	acc	0.4724	—

Turkish NLP Benchmarks (0-shot)

Task	Metric	Score
Belebele (TR)	acc	0.8144
Turkish MMLU (avg)	acc	0.6555
XCOPA (TR)	acc	0.6780

No catastrophic forgetting observed — English reasoning is fully preserved after Turkish fine-tuning.

Usage

import torch
from transformers import AutoModelForCausalLM, AutoTokenizer

model_id = "MuhammedKsee/Qwen3.5-9B-Instruct-Turca-TurkishLLM"

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

messages = [
    {"role": "system", "content": "Sen yardımcı, dürüst ve zararsız bir Türkçe yapay zeka asistanısın."},
    {"role": "user",   "content": "Türkiye'nin en büyük şehri hangisidir?"}
]

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

generated_ids = model.generate(
    **inputs,
    max_new_tokens=512,
    temperature=0.7,
    top_p=0.9,
    repetition_penalty=1.1
)
generated_ids = [o[inputs.input_ids.shape[1]:] for o in generated_ids]
print(tokenizer.batch_decode(generated_ids, skip_special_tokens=True)[0])

The system prompt is kept in Turkish to match the fine-tuning distribution. English system prompts also work but have not been formally evaluated.

Example Prompts

General chat

Bana motivasyon veren kısa bir konuşma yap.

Reasoning

Bir öğrenciye yapay zekayı basit şekilde anlat.

Content generation

Türkçe öğrenen biri için kısa bir hikaye yaz.

Translation

"Machine learning is transforming the world" cümlesini Türkçeye çevir.

Intended Use

Use Case	Verdict
Turkish conversational assistant	✅ Primary use case
Turkish RAG / document QA	✅ Recommended
Summarization & translation	✅ Intended use
Educational & tutoring applications	✅ Good fit
Local inference (quantized)	✅ See GGUF repo
Production enterprise (high-stakes)	⚠️ Evaluate on your specific workload
Advanced mathematical proofs	⚠️ Not specialized for this

Limitations

SFT-only model — no DPO or RLHF alignment stage applied
May occasionally produce non-optimal answers in safety-critical scenarios
Single epoch training over 500K samples
Turkish-specific benchmarks are a subset of available evaluations — community results welcome

Related Resources

Resource	Link
GGUF (local inference)	MuhammedKsee/Qwen3.5-9B-Instruct-Turca-TurkishLLM-GGUF
Training dataset	turkish-nlp-suite/InstrucTurca
Base model	Qwen/Qwen3.5-9B

Citation

@misc{kose2026qwen35turca,
  author    = {Muhammed Köse},
  title     = {Qwen3.5-9B-Instruct-Turca-TurkishLLM},
  year      = {2026},
  publisher = {Hugging Face},
  url       = {https://huggingface.co/MuhammedKsee/Qwen3.5-9B-Instruct-Turca-TurkishLLM}
}

Downloads last month: -

Safetensors

Model size

9B params

Tensor type

BF16

Model tree for MuhammedKsee/Qwen3.5-9B-Instruct-Turca-TurkishLLM

Base model

Qwen/Qwen3.5-9B-Base

Finetuned

Qwen/Qwen3.5-9B

Adapter

(161)

this model

MuhammedKsee
/

Qwen3.5-9B-Instruct-Turca-TurkishLLM