AstrumN1Mini
AstrumN1Mini is a 0.5B-parameter instruction-tuned language model with a focus on Turkish and English. It is the first model of the Astrum family (formerly Alary) by Göktürk Bağbaşı, built on top of Qwen/Qwen2.5-0.5B-Instruct.
It is a small, fast model meant for experimentation, lightweight local use, and as a baseline for the next models in the N-series.
Model details
| Parameters | 0.5B |
| Base model | Qwen/Qwen2.5-0.5B-Instruct |
| Languages | Turkish, English |
| Context length | 32,768 tokens |
| License | Apache-2.0 |
Usage
from transformers import AutoModelForCausalLM, AutoTokenizer
model_id = "Gokturk97/AstrumN1Mini"
tokenizer = AutoTokenizer.from_pretrained(model_id)
model = AutoModelForCausalLM.from_pretrained(model_id, torch_dtype="auto", device_map="auto")
messages = [{"role": "user", "content": "Merhaba! Kendini kısaca tanıtır mısın?"}]
text = tokenizer.apply_chat_template(messages, tokenize=False, add_generation_prompt=True)
inputs = tokenizer(text, return_tensors="pt").to(model.device)
output = model.generate(**inputs, max_new_tokens=256)
print(tokenizer.decode(output[0][inputs.input_ids.shape[1]:], skip_special_tokens=True))
Training
- Data: 10M tokens of Turkish instruction data
- Hardware: Kaggle, 2x NVIDIA T4
Limitations
This is a very small model, so please keep expectations realistic:
- Factual knowledge is unreliable, especially geography (for example, country capitals can be wrong).
- Arithmetic and multi-step reasoning are weak.
- It can hallucinate confidently. Do not use it for medical, legal, financial, or other high-stakes decisions.
- Turkish and English are the target languages; other languages are not a focus.
Roadmap
Next in the series: N1.1 mini (successor to this model), N2 mini (2B), and larger Astrum models.
Türkçe özet
AstrumN1Mini, Türkçe ve İngilizce odaklı, 0.5B parametreli küçük bir dil modelidir. Qwen2.5-0.5B-Instruct tabanlıdır ve Apache-2.0 lisansıyla açık olarak paylaşılmıştır. Eğitimde 10M token Türkçe instruction verisi kullanılmıştır. Küçük bir model olduğu için özellikle coğrafya ve matematik gibi konularda hata yapabilir; deneme ve hafif kullanım amaçlıdır.
Contact
Feedback and issues are welcome in the community tab of this repo.