82 lines
2.8 KiB
Markdown
82 lines
2.8 KiB
Markdown
---
|
||
language:
|
||
- tr
|
||
license: apache-2.0
|
||
tags:
|
||
- llama
|
||
- causal-lm
|
||
- slm
|
||
- experimental
|
||
- turkish
|
||
- tr-llm
|
||
- AhiskaAI
|
||
datasets:
|
||
- BILGEM-AI/BILGE-Synthetic-Math
|
||
- BILGEM-AI/BILGE-Synthetic-Web
|
||
- BILGEM-AI/BILGE-Synthetic-Stories
|
||
pipeline_tag: text-generation
|
||
---
|
||
|
||
# AhıskaAI-110M-Experimental-v0.1
|
||
|
||
**AhıskaAI-110M-Experimental-v0.1** is a 110M parameter Small Language Model (SLM) trained from scratch as part of the experimental release series by AhıskaAI. This series focuses on evaluating early-stage convergence and reasoning behavior on synthetic Turkish datasets (covering math, web content, and narrative stories).
|
||
|
||
## Experimental Setup & Scope
|
||
|
||
- **Objective:** Test synthetic data density (Math + Web + Stories) and sequence modeling performance on a compact 110M Llama architecture.
|
||
- **Training Epochs:** Pre-trained for **0.3 epochs** across a ~3.5 GB synthetic Turkish corpus.
|
||
- **Context Length:** 1024 tokens.
|
||
|
||
## Model Details
|
||
|
||
- **Architecture:** `LlamaForCausalLM` (Standard MHA)
|
||
- **Parameters:** ~110M
|
||
- **Hidden Size:** 768
|
||
- **Intermediate Size:** 2048
|
||
- **Attention Heads:** 12 (Query) / 12 (Key/Value)
|
||
- **Layers:** 12
|
||
- **Context Length:** 1024 tokens
|
||
- **Vocabulary Size:** 32,000
|
||
- **Precision:** `float32` / `float16`
|
||
|
||
## Pre-training Dataset Mixture
|
||
|
||
The model was pre-trained on a synthetic Turkish corpus (~3.5 GB total):
|
||
|
||
| Data Source | Approx. Size | Primary Focus | Link |
|
||
| :--- | :--- | :--- | :--- |
|
||
| **BILGEM AI Synthetic Web** | ~2.0 GB | General Web, Knowledge & Grammar | [BILGEM-AI/BILGE-Synthetic-Web](https://huggingface.co/datasets/BILGEM-AI/BILGE-Synthetic-Web) |
|
||
| **BILGEM AI Synthetic Math** | ~750 MB | Mathematical Reasoning & Logic | [BILGEM-AI/BILGE-Synthetic-Math](https://huggingface.co/datasets/BILGEM-AI/BILGE-Synthetic-Math) |
|
||
| **BILGEM AI Synthetic Stories** | ~750 MB | Narrative Flow & Text Generation | [BILGEM-AI/BILGE-Synthetic-Stories](https://huggingface.co/datasets/BILGEM-AI/BILGE-Synthetic-Stories) |
|
||
|
||
## Usage
|
||
|
||
```python
|
||
import torch
|
||
from transformers import AutoModelForCausalLM, AutoTokenizer
|
||
|
||
model_id = "AhıskaAI/AhıskaAI-110M-Experimental-v0.1"
|
||
|
||
tokenizer = AutoTokenizer.from_pretrained(model_id)
|
||
model = AutoModelForCausalLM.from_pretrained(
|
||
model_id,
|
||
torch_dtype=torch.float16 if torch.cuda.is_available() else torch.float32,
|
||
device_map="auto"
|
||
)
|
||
|
||
prompt = "Bir zamanlar uzak bir ülkede"
|
||
inputs = tokenizer(prompt, return_tensors="pt").to(model.device)
|
||
|
||
outputs = model.generate(
|
||
**inputs,
|
||
max_new_tokens=100,
|
||
temperature=0.7,
|
||
top_p=0.9,
|
||
do_sample=True
|
||
)
|
||
|
||
print(tokenizer.decode(outputs[0], skip_special_tokens=True))
|
||
```
|
||
|
||
### About AhıskaAI
|
||
**AhıskaAI** is an independent initiative dedicated to developing efficient, high-performance Small Language Models (SLMs) tailored for the Turkish language ecosystem. |