AhiskaAI

AhiskaAI-308m-Base-v0.2

Deploy Dedicated

README

License: apache-2.0

Model Details

Architecture: Llama-based architecture.
Parameters: 308M.
Context Window: 1024 tokens.
Tokenizer: Custom BPE Tokenizer (Vocabulary Size: 32,000).
Training Framework: PyTorch & Transformers.

Data Curation (The "Quality over Quantity" Approach)

The v0.2 release adopts a data-centric training strategy, prioritizing corpus quality over dataset size.

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 improve 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

The 308M model serves as the flagship base model of the AhiskaAI v0.2 family.

Its primary objectives are:

Strong Turkish language modeling.
Improved semantic understanding.
Better contextual consistency.
A research foundation for future instruction tuning and alignment.

Training Logs

Training Loss Curve

The graph above demonstrates the training convergence of AhiskaAI-308m-Base-v0.2. The stable decline in loss indicates effective optimization throughout pretraining.

Usage

python
from transformers import AutoModelForCausalLM, AutoTokenizer

model = AutoModelForCausalLM.from_pretrained("AhiskaAI/AhiskaAI-308m-Base-v0.2")
tokenizer = AutoTokenizer.from_pretrained("AhiskaAI/AhiskaAI-308m-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

Instruction-tuned (IT) version.
Preference alignment with DPO.
Larger and more diverse Turkish datasets.
Future AhiskaAI 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.

Available on FriendliAI

Dedicated Endpoints

Run this model inference on single tenant GPU with unmatched speed and reliability at scale.

Learn more

Container

Run this model inference with full control and performance in your environment.

Learn more

Model Details

Model Provider

AhiskaAI

Model Tree

Base

this model

Input Modalities

Text

Output Modalities

Text

Supported Functionality

Dedicated EndpointsContainer

Explore FriendliAI today

Get started Talk to an engineer

README

License: apache-2.0

Model Details

Architecture: Llama-based architecture.
Parameters: 308M.
Context Window: 1024 tokens.
Tokenizer: Custom BPE Tokenizer (Vocabulary Size: 32,000).
Training Framework: PyTorch & Transformers.

Data Curation (The "Quality over Quantity" Approach)

The v0.2 release adopts a data-centric training strategy, prioritizing corpus quality over dataset size.

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 improve 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

The 308M model serves as the flagship base model of the AhiskaAI v0.2 family.

Its primary objectives are:

Strong Turkish language modeling.
Improved semantic understanding.
Better contextual consistency.
A research foundation for future instruction tuning and alignment.

Training Logs

Training Loss Curve

The graph above demonstrates the training convergence of AhiskaAI-308m-Base-v0.2. The stable decline in loss indicates effective optimization throughout pretraining.

Usage

python
from transformers import AutoModelForCausalLM, AutoTokenizer

model = AutoModelForCausalLM.from_pretrained("AhiskaAI/AhiskaAI-308m-Base-v0.2")
tokenizer = AutoTokenizer.from_pretrained("AhiskaAI/AhiskaAI-308m-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

Instruction-tuned (IT) version.
Preference alignment with DPO.
Larger and more diverse Turkish datasets.
Future AhiskaAI 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.