SixFinger-2B — experimental Turkish fine-tune

SixFinger-2B is an experimental fine-tune based on VNGRS Kumru-2B, trained on an NVIDIA RTX 3060 Laptop GPU. The repository contains full model weights, not just a LoRA adapter.

Model details

Property	Details
Base model	vngrs/Kumru-2B
Size	Approximately 2 billion parameters
Training approach	QLoRA (reported in original project notes)
Training setup	3 epochs, batch size 8 (reported)
Hardware	NVIDIA RTX 3060 Laptop GPU
Status	Research checkpoint; no verified task benchmarks

The original project aimed to experiment with Turkish text completion and mathematical prompts. It has not been demonstrated to outperform the base model on those tasks.

Load and generate text

pip install -U transformers accelerate torch

import torch
from transformers import AutoModelForCausalLM, AutoTokenizer

model_id = "sixfingerdev/sixfinger-2b"

tokenizer = AutoTokenizer.from_pretrained(model_id)
model = AutoModelForCausalLM.from_pretrained(model_id, device_map="auto")
model.eval()

prompt = "Merhaba, bugün"
inputs = tokenizer(prompt, return_tensors="pt").to(model.device)

with torch.inference_mode():
    outputs = model.generate(
        **inputs,
        max_new_tokens=100,
        do_sample=False,
        pad_token_id=tokenizer.eos_token_id,
    )

new_tokens = outputs[0, inputs["input_ids"].shape[-1]:]
print(tokenizer.decode(new_tokens, skip_special_tokens=True))

Evaluation and limitations

• No reproducible downstream benchmark or base-model comparison is currently published.
• Outputs may contain factual, reasoning, or language errors.
• Training on a laptop GPU does not by itself establish a model-quality limitation; the important factors are data, optimization, and evaluation.
• Research use is recommended until its quality has been tested on a representative held-out set.

Attribution and license

This repository declares Apache-2.0, matching the published license of the Kumru-2B base model. Credit to the VNGRS team for the original model.