Text Generation
Transformers
Safetensors
Azerbaijani
gemma3_text
azerbaijani
small-language-model
main-lab
conversational
text-generation-inference

MAIN Lab · Gemma 3 270M Azerbaijani

A 270M-parameter model fine-tuned for Azerbaijani by MAIN Lab (Master Artificial Intelligence Network): distilling frontier intelligence into small models.

Built from google/gemma-3-270m-it. All weights were fine-tuned except the embeddings, on 12,099 Azerbaijani chat examples (2.49M tokens). Training ran on a CPU only and took about 2 hours. It was stopped at step 900 of 1,324 (68% of one epoch).

benchmark

Results (Benchmark #001, same tests for every model)

Test Gemma 3 270M (base) This model
Azerbaijani vs Turkish word choice (YOXLA) 49.0 86.0
Grammar minimal pairs (YOXLA) 51.5 80.0
EN→AZ translation, chrF++ (FLORES-200) 14.0 25.1 (#1 of 5 small models tested)
Reading QA, F1 (YOXLA) 4.2 18.5
Bits per byte on Azerbaijani text (lower is better) 2.89 1.61

Honest limits:

  • Knowledge and reasoning tests (Belebele, INCLUDE, facts, NLI) stay near chance. 270M parameters is too small for them.
  • SIBA Meristem Native 0.5B is better at grammar and at choosing the Azerbaijani word over the Turkish one.
  • Translations are now Azerbaijani, but they are often inaccurate.

Before training, any training example that shared an 8-word span with a test was removed.

Use

from transformers import AutoModelForCausalLM, AutoTokenizer
tok = AutoTokenizer.from_pretrained("MAIN-Lab/gemma-3-270m-az")
m = AutoModelForCausalLM.from_pretrained("MAIN-Lab/gemma-3-270m-az")
msgs = [{"role": "user", "content": "Azərbaycanın paytaxtı haradır?"}]
ids = tok.apply_chat_template(msgs, add_generation_prompt=True, return_tensors="pt", return_dict=True)
print(tok.decode(m.generate(**ids, max_new_tokens=128)[0][ids["input_ids"].shape[1]:], skip_special_tokens=True))

Training data

Part of this data was generated by OpenAI models, so check that this fits your use case. Thanks to LocalDoc for Bilik-Instruct and the YOXLA benchmark, and to SIBA for Meristem.

License

Gemma is provided under and subject to the Gemma Terms of Use found at https://ai.google.dev/gemma/terms. This model is a modification of Gemma and is distributed under those terms.

Downloads last month
88
Safetensors
Model size
0.3B params
Tensor type
BF16
·
Inference Providers NEW
This model isn't deployed by any Inference Provider. 🙋 Ask for provider support

Model tree for MAIN-Lab/gemma-3-270m-az

Finetuned
(1,203)
this model

Datasets used to train MAIN-Lab/gemma-3-270m-az

Spaces using MAIN-Lab/gemma-3-270m-az 2