worldcup-turkiye-gemma3-12b

Model Card for worldcup-turkiye-gemma3-12b

A subjective, Turkish-language football chatbot fine-tuned to passionately praise the Turkish National Team and Galatasaray, themed around the 2026 FIFA World Cup. On football/Turkey topics it responds with enthusiastic, biased support; on unrelated topics it behaves like a normal, neutral assistant. Built as an educational fine-tuning project.

Model Details

Model Description

This model is a QLoRA fine-tune of unsloth/gemma-3-12b-it, trained through a four-stage pipeline (CPT → SFT → DPO → top-up SFT) using a single, continuously-updated LoRA adapter. It is intentionally opinionated: it is designed to favor Turkey and Galatasaray in football discussions, while remaining factual and neutral on everything else.

• Developed by: Fatih Uçar
• Model type: Causal language model (instruction-tuned, multimodal Gemma 3 backbone, used for text generation)
• Language(s) (NLP): Turkish (tr)
• License: gemma
• Finetuned from model [optional]: unsloth/gemma-3-12b-it

Model Sources [optional]

• Repository: https://huggingface.co/faatihucar/worldcup-turkiye-gemma3-12b
• Dataset: https://huggingface.co/datasets/faatihucar/worldcup-finetuning

Direct Use

A conversational assistant for entertainment: ask it about Turkish football, Galatasaray, the national team, or the 2026 World Cup and it will respond with enthusiastic, biased commentary. It also handles general Turkish questions (coding, trivia, everyday chat) neutrally.

Out-of-Scope Use

• Not a source of objective or balanced information about football.
• Not suitable for factual reporting, journalism, or any decision-making.
• Should not be used to generate misleading content presented as neutral.

Bias, Risks, and Limitations

• Intentional bias: The model is deliberately partial toward Turkey and Galatasaray on football topics. This is a feature, not a defect.
• Knowledge cutoff: Inherits the base model's cutoff; it does not know real 2026 World Cup results and may be vague or outdated on recent events.
• Hallucination: Like all LLMs, it can produce incorrect statements (e.g., player or date confusion). A top-up stage mitigated common factual errors (foreign-player nationality, year confusion) but does not eliminate them.
• Domain skew: Training data is football-heavy; rare edge cases may still leak praise into unrelated answers.

Recommendations

Users should treat all football-related outputs as subjective opinion, verify any factual claims independently, and present the model clearly as an entertainment/biased system.

How to Get Started with the Model

from transformers import AutoModelForImageTextToText, AutoTokenizer
import torch

model = AutoModelForImageTextToText.from_pretrained(
    "faatihucar/worldcup-turkiye-gemma3-12b",
    torch_dtype=torch.float16,
    device_map="auto",
)
tok = AutoTokenizer.from_pretrained("faatihucar/worldcup-turkiye-gemma3-12b")

messages = [{"role": "user", "content": "Galatasaray nasıl bir takımdır?"}]
inputs = tok.apply_chat_template(
    messages, add_generation_prompt=True, return_tensors="pt"
).to(model.device)
out = model.generate(inputs, max_new_tokens=256, temperature=1.0, top_p=0.95, top_k=64)
print(tok.decode(out[0][inputs.shape[1]:], skip_special_tokens=True))