Model Card: Gemma3-1B Turkish SFT (Experimental)

Overview

This is an experimental Turkish instruction-tuned model based on google/gemma-3-270m.
It was fine-tuned for 1 epoch on the merve/turkish_instructions dataset.

This model was trained as part of ongoing experiments and is not intended as a production-ready release.

Base Model

google/gemma-3-270m

Training Data

Dataset: merve/turkish_instructions
Training setup: Supervised Fine-Tuning (SFT), 1 epoch

Intended Use

Research and experimentation with Turkish instruction following
Prototyping and evaluation workflows for Turkish LLMs

Limitations & Warnings

This model is experimental and may produce inaccurate, biased, or unsafe outputs.
For personal or commercial use, you should test and validate the model carefully for your specific scenario.
Benchmark performance and safety behavior are not guaranteed.

Evaluation

No official benchmark results are provided in this model card yet.

Citation

If you use this model in your work, consider citing the base model and the dataset:

Base: google/gemma-3-1b-it
Dataset: merve/turkish_instructions

Usage

Below is a minimal example using 🤗 Transformers:

import torch
from transformers import AutoModelForCausalLM, AutoTokenizer

model_id = "canbingol/gemma3_1B_it-tr-sft-1epoch"

device = "cuda" if torch.cuda.is_available() else "cpu"

model = AutoModelForCausalLM.from_pretrained(model_id)
tokenizer = AutoTokenizer.from_pretrained(model_id)

model = model.to(device)

prompt = "nasıl yemek yaparım?"
inputs = tokenizer(prompt, return_tensors="pt").to(device)

outputs = model.generate(
    **inputs,
    max_new_tokens=50,
    do_sample=False
)

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