---
language:
- tr
- en
license: mit
tags:
- phi-2
- microsoft
- text-generation
- tr
- turkish
- qlora
inference: false
pipeline_tag: text-generation
---

# sixfinger-phi2-merged

This model is a fine-tuned and merged version of [Microsoft Phi-2](https://huggingface.co/microsoft/phi-2) created by **Six Finger Dev** (Enes Altıparmak). It is a 2.7 billion parameter causal language model tailored to perform well on Turkish Question-Answering (QA), reasoning, and basic coding tasks.

## Model Details

- **Developer:** Six Finger Dev (Enes Altıparmak - Kayseri Science High School)
- **Architecture:** Phi-2 Causal LM
- **Parameters:** ~2.7B
- **Languages:** Turkish (TR), English (EN)
- **License:** MIT

## Training & Optimization
This model was likely fine-tuned using QLoRA against a custom Turkish instruction and multi-turn QA dataset (e.g., [sixfingerdev/turkish-qa-multi-dialog-dataset](https://huggingface.co/datasets/sixfingerdev/turkish-qa-multi-dialog-dataset)). After fine-tuning, the PEFT adapters were fully merged back into the base model weights, meaning it can be loaded directly as a standalone checkpoint without needing the base model or adapter configuration.

## Usage

You can load and generate text with this model directly using the `transformers` library:

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

model_id = "sixfingerdev/sixfinger-phi2-merged"

tokenizer = AutoTokenizer.from_pretrained(model_id)
model = AutoModelForCausalLM.from_pretrained(
    model_id,
    device_map="auto",
    torch_dtype=torch.float16,
    low_cpu_mem_usage=True
)

prompt = "Soru: Türkiyenin başkenti neresidir? Cevap:"
inputs = tokenizer(prompt, return_tensors="pt").to(model.device)

with torch.no_grad():
    outputs = model.generate(**inputs, max_new_tokens=40)
    
print(tokenizer.decode(outputs[0], skip_special_tokens=True))
```

## Limitations & Biases
While fine-tuned with instruction data, its behavior still heavily relies on prompt-completion formatting. Direct cues like `Answer:` or `Cevap:` yield the best deterministic outputs. In unstructured or lengthy multi-turn chat loops, the model may suffer from repetition or formatting drift compared to purely conversational templates.