The base model was further trained for 1 epoch on the first 50,000 samples of a Turkish web corpus.
This stage represents a broader data exposure regime compared to small-subset experiments and aims to improve:

Turkish fluency
Vocabulary coverage
Domain familiarity
Statistical adaptation to Turkish web distribution

This release corresponds to Stage 1 of a multi-stage CPT pipeline and is intended for research and experimental analysis.

Base Model

google/gemma-3-1b-pt

Training Setup

Dataset: canbingol/vngrs-web-corpus-200k
Subset Used: First 50,000 samples
Training Objective: Continued Pretraining
Epochs: 1
Data Regime: Plain text
Token Count (approximate): ~21.5M tokens

Usage Example

python

import torch
from transformers import AutoModelForCausalLM, AutoTokenizer

model_id = "canbingol/gemma3_1B_base-tr-cpt-1epoch_stage1"

device = "cuda" if torch.cuda.is_available() else "cpu"

model = AutoModelForCausalLM.from_pretrained(model_id)
tokenizer = AutoTokenizer.from_pretrained(model_id)

model = model.to(device)

prompt = "bundan böyle"
inputs = tokenizer(prompt, return_tensors="pt").to(device)

outputs = model.generate(
    **inputs,
    max_new_tokens=50,
    do_sample=True,
    temperature=0.8,
    top_p=0.9
)

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