AhiskaAI-65m-Base-v0.2 is a 65 million parameter Small Language Model (SLM) built from scratch. It is the lightweight member of the AhiskaAI v0.2 family, designed to provide efficient Turkish language understanding on resource-constrained hardware.

Model Details

Architecture: Llama-based architecture.
Parameters: 65M.
Context Window: 1024 tokens.
Tokenizer: Custom BPE Tokenizer (Vocabulary Size: 32,000).
Training Framework: PyTorch & Transformers.

Data Curation (The "Quality over Quantity" Approach)

A major improvement in the v0.2 release is the adoption of a data-centric training pipeline.

Raw Data (v0.1): 5GB of raw Turkish corpus.
Curated Data (v0.2): 1.2GB of carefully filtered, high-quality Turkish text.
Process: Approximately 75% of noisy, duplicated, and low-quality samples were removed to maximize linguistic quality and training efficiency.

Key Improvements from v0.1

Architecture Shift: Migrated from GPT-2 to a modern Llama-based architecture.
Normalization: RMSNorm.
Positional Encoding: RoPE (Rotary Positional Embeddings).
Activation: SiLU.
Precision: Trained using bfloat16 for efficient consumer GPU training.

Design Goal

Unlike larger language models, the 65M variant focuses on efficiency while preserving core language understanding abilities.

The objective of this model is not to maximize factual knowledge, but to provide:

Strong Turkish language modeling.
Basic instruction understanding.
Fast inference on low-resource hardware.
A compact research baseline for future AhiskaAI releases.

Training Logs

The graph above demonstrates the training convergence of AhiskaAI-65m-Base-v0.2. The stable decline in loss confirms the effective alignment of the model architecture with the curated Turkish dataset.

Usage

python

from transformers import AutoModelForCausalLM, AutoTokenizer

model = AutoModelForCausalLM.from_pretrained("AhiskaAI/AhiskaAI-65m-Base-v0.2")
tokenizer = AutoTokenizer.from_pretrained("AhiskaAI/AhiskaAI-65m-Base-v0.2")

text = "Türkiye Cumhuriyeti"
inputs = tokenizer(text, return_tensors="pt")

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

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

Training Hardware

Trained on NVIDIA RTX 4050 6GB Laptop GPU.

Future Plans

Larger context windows.
Improved Turkish datasets.
Future v0.3 model family.

About AhiskaAI

AhiskaAI is an independent open-source initiative dedicated to developing efficient Turkish Small Language Models trained completely from scratch. Follow us on Hugging Face for updates and new releases.