---
license: apache-2.0
language:
- tr
pipeline_tag: text-generation
tags:
- llama
- slm
- alpaca
- instruction-tuning
- causal-lm
- tr-llm
- Ahıska
- AhiskaTurks
- MeskhetianTurks
- AhıskaTürkleri
datasets:
- custom-filtered-turkish-alpaca
library_name: transformers
---

# AhiskaAI 65m IT v0.1 (Instruction Tuned)

**AhiskaAI 65m IT v0.1** is a highly efficient, custom-aligned Small Language Model (SLM) for the Turkish language ecosystem. 

This model was NOT fine-tuned on top of generic open-source weights. Instead, it was instruction-tuned directly over our proprietary foundation model, **AhıskaAI 65m Base v0.1** (which was pre-trained from scratch for 1 full epoch on a 5.3 GB Turkish corpus). For this alignment phase (SFT), we utilized a strictly **filtered and curated Turkish Alpaca dataset** to maximize procedural logic, formatting accuracy, and structural fluidity while eliminating noisy data tokens.

---

## 🧬 The Pipeline: From Scratch to Instruction

Our research lab follows a strict vertical integration philosophy:
1. **Phase 1 (Base Model):** Initialized `LlamaForCausalLM` from zero variables. Pre-trained on 5.3 GB of clean Turkish text matrix to lock down grammar, token-nesting patterns, and core semantics (**AhıskaAI 65m Base v0.1**).
2. **Phase 2 (Instruction Tuning):** Supervised Fine-Tuning (SFT) over the base checkpoint using our custom-filtered Alpaca instructions. This phase injected formatting discipline, listing mechanics (`1. 2. 3.`), and multi-turn response compliance.

---

## 📊 Technical Architecture & Hyperparameters

Directly extracted from the native `config.json`, the model utilizes a pure modern LLaMA layout optimized for fast local compute:

* **Architecture:** `LlamaForCausalLM`
* **Parameters:** ~65 Million
* **Context Length (`max_position_embeddings`):** 1024 tokens (Double the capacity of legacy GPT-2 baselines)
* **Vocabulary Size:** 32,000 tokens (Custom BPE trained for Turkish root-suffix morphology)
* **Hidden Dimension (`hidden_size`):** 512
* **Intermediate Layer Dimension (`intermediate_size`):** 1376
* **Hidden Layers (`num_hidden_layers`):** 12
* **Attention Heads:** 8 (`num_attention_heads` / `num_key_value_heads`)
* **Activation Function:** SiLU (`silu`)
* **Normalization EPS:** `rms_norm_eps: 1e-06` (RMSNorm architecture)
* **Positional Embeddings:** RoPE (`rope_type: default`, theta: 10000.0)
* **Data Precision:** `float32`

---

## 💻 Hardware Efficiency & "Build in Public"

* **Training & Alignment Hardware:** NVIDIA GeForce RTX 4050 Laptop GPU (6GB VRAM)
* **Inference Footprint:** Merely **~202 MB** in size! It runs at lightning-fast tokens-per-second even on **Hugging Face Free CPU Spaces**, bypassing the need for expensive cloud GPU hosting.

---

## 🛠️ Quickstart Usage (Alpaca Format)

To interact with the instruction-tuned layer smoothly, invoke the model with the exact token structure it was aligned with:

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

model_name = "AhiskaAI/AhiskaAI-65m-IT-v0.1"

# Load the custom-built architecture and vocabulary
model = LlamaForCausalLM.from_pretrained(model_name).to("cuda" if torch.cuda.is_available() else "cpu")
tokenizer = AutoTokenizer.from_pretrained(model_name)

def ask_ahiska_it(instruction):
    # Strict Alpaca Template
    prompt = f"<|im_start|>user\n{user_input}<|im_end|>\n<|im_start|>assistant\n"

    
    inputs = tokenizer(prompt, return_tensors="pt").to(model.device)
    
    with torch.no_grad():
        outputs = model.generate(
            **inputs, 
            max_length=250, 
            do_sample=True, 
            top_k=40, 
            top_p=0.92,
            temperature=0.55, # Low temp keeps the 65m nodes highly focused
            repetition_penalty=1.18
        )
    
    response = tokenizer.decode(outputs[0], skip_special_tokens=True)
    return response.split("### Response:\n")[-1].strip()

# Run a test inference
print(ask_ahiska_it("Sağlıklı yaşamak için 3 ipucu ver"))