---
library_name: transformers
license: gemma
language:
  - tr
base_model:
  - google/gemma-2-9b-it
pipeline_tag: text-generation
---

# Gemma-2-9b-it-tr

Gemma-2-9b-it-tr is a finetuned version of [google/gemma-2-9b-it](https://huggingface.co/google/gemma-2-9b-it) on a carefully curated and manually filtered dataset of 55k question answering and conversational samples in Turkish.


## Training Details
**Base model:** [google/gemma-2-9b-it](https://huggingface.co/google/gemma-2-9b-it)  
**Training data:** A filtered version of [metedb/turkish_llm_datasets](https://huggingface.co/datasets/metedb/turkish_llm_datasets/) and a small private dataset of 8k conversational samples on various topics.  
**Training setup:** We performed supervised fine tuning with LoRA with `rank=128` and `lora_alpha`=64. Training took 4 days on a single  RTX 6000 Ada.  

Compared to the base model, we find Gemma-2-9b-tr has superior conversational and reasoning skills.

## Usage
You can load and use `neuralwork/gemma-2-9b-it-tr`as follows.

```py
import torch
from transformers import AutoModelForCausalLM, AutoTokenizer

model = AutoModelForCausalLM.from_pretrained(
   "neuralwork/gemma-2-9b-it-tr",
   torch_dtype=torch.bfloat16,
   device_map="auto",
   trust_remote_code=True
)

tokenizer = AutoTokenizer.from_pretrained("neuralwork/gemma-2-9b-it-tr")

messages = [
   {"role": "user", "content": "Python'da bir öğenin bir listede geçip geçmediğini nasıl kontrol edebilirim?"},
]

prompt = tokenizer.apply_chat_template(
   messages,
   tokenize=False,
   add_generation_prompt=True
)

outputs = model.generate(
   tokenizer(prompt, return_tensors="pt").input_ids.to(model.device),
   max_new_tokens=1024,
   do_sample=True,
   temperature=0.7,
   top_p=0.9
)

response = tokenizer.decode(outputs[0], skip_special_tokens=True)[len(prompt):]
print(response)
```