plato-9b

plato-9b is a fine-tuned version of the google/gemma-2-9b-it model for generating responses in the Russian language. This 9-billion parameter model excels at conversational tasks, offering rich contextual understanding and fine-grained results.

Usage

To use plato-9b with the transformers library:

from transformers import AutoTokenizer, AutoModelForCausalLM

tokenizer = AutoTokenizer.from_pretrained("deepvk/plato-9b")
model = AutoModelForCausalLM.from_pretrained("deepvk/plato-9b")

input_text = "Что стоит посетить в России?"
input_ids = tokenizer(input_text, return_tensors="pt").input_ids

output = model.generate(input_ids, max_length=150, do_sample=True, temperature=0.7)
response = tokenizer.decode(output[0], skip_special_tokens=True)
print(response)
# Что стоит посетить в России?
# 1. Красная площадь и Кремль в Москве
# 2. Эрмитаж в Санкт-Петербурге
# 3. Байкал
# 4. Соловецкие острова
# 5. Камчатка и её вулканы
# 6. Золотое Кольцо
# 7. Казанский Кремль
# 8. Алтай
# 9. Астраханская область и Волго-Донской канал
# 10. Кавказские горы и Черноморское побережье
# 
# Каждое из этих мест предлагает уникальные культурные, исторические и природные достопримечательности,
# которые делают Россию столь удивительной и разнообразной страной.

Dataset

We applied both Supervised Fine-Tuning (SFT) and Preference Optimization (PO). For SFT, we used an 8B token instruction dataset, with 4B tokens consisting of dialogues and the rest covering math, biology, chemistry, code, and general knowledge. The PO dataset contains 200M tokens featuring common knowledge instructions. We trained on both datasets for several epochs.

Evaluation

To evaluate, we applied LLM-as-a-judge approach on academic tasks. Specifically, we used arena-general-ru and arena-hard-ru with gpt4o judge and gpt4o-mini baseline.

arena-general-ru

Model	Score	Score w/ SC
gpt-4o-2024-11-20	81.87 (-2.04, +1.81)	78.42 (-2.39, +2.33)
gpt-4o-mini-2024-07-18	50.00 (-0.00, +0.00)	50.00 (-0.00, +0.00)
deepvk/plato-9b	41.27 (-2.18, +2.24)	32.13 (-1.97, +2.05)
t-tech/T-lite-it-1.0	38.52 (-2.04, +2.98)	30.38 (-1.90, +3.15)
google/gemma-2-9b-it	27.46 (-2.06, +1.74)	25.80 (-2.09, +1.98)
Qwen/Qwen2.5-7B-Instruct	24.60 (-2.36, +2.38)	23.67 (-2.36, +2.28)
IlyaGusev/saiga_gemma2_9b	17.83 (-1.95, +1.66)	18.46 (-2.22, +1.69)

arena-hard-ru

Model	Score	Score w/ SC
gpt-4o-2024-11-20	85.70 (-1.45, +1.38)	80.19 (-1.99, +2.04)
gpt-4o-mini-2024-07-18	50.00 (-0.00, +0.00)	50.00 (-0.00, +0.00)
t-tech/T-lite-it-1.0	34.80 (-1.98, +2.38)	26.99 (-1.74, +2.67)
deepvk/plato-9b	31.81 (-1.92, +1.90)	24.25 (-1.71, +1.84)
Qwen/Qwen2.5-7B-Instruct	20.84 (-1.99, +1.67)	17.70 (-1.63, +1.68)
google/gemma-2-9b-it	12.98 (-1.36, +1.57)	12.97 (-1.46, +1.69)
IlyaGusev/saiga_gemma2_9b	9.72 (-1.34, +1.50)	10.64 (-1.40, +1.78)

Citation

Both authors contribute equally, order is alphabetical.

@misc{deepvk2024plato-9b,
    title={plato-9b},
    author={Eliseev, Anton and Semin, Kirill},
    url={https://huggingface.co/datasets/deepvk/plato-9b},
    publisher={Hugging Face}
    year={2025},
}