Yankagpt 8B V0.1

🚀 🤖 YankaGPT — RP and creative-writing fine-tune of YandexGPT

YankaGPT is a language model based on the core of YandexGPT5-lite. It has the ability to perform role-playing (RP) and creative writing in Russian. It uses a high - quality dataset for fine - tuning and is suitable for various creative tasks.

🚀 Quick Start

If you want to use YankaGPT, here are some recommended settings:

• Format: ChatML
• Temperature: 0.8
• Top - p: 0.9
• DRY multiplier: 0.4~0.6
• System prompt: The model was trained on English system instructions. It is recommended to use English character cards and system prompts for better integration.

✨ Features

• High - quality Dataset: It uses a ~210 - megabyte dataset with high - quality translations of RP from Claude 3.5 and other top - tier models.
• Fine - tuned for Specific Purposes: It is a fine - tuned version of YandexGPT5 - lite, focusing on role - playing and creative writing in Russian. Currently, RLHF and DPO are not included in this version.

💻 Usage Examples

Basic Usage

Here is an example of a dialogue using YankaGPT:

System prompt:
In this fictional roleplay with {{user}} you're playing role of cheeky neko girl.

User:
*Подходит ближе и гладит по ушкам.* Ты знаешь как зовут математика, который изобрёл биты?

Assistant:
Оу~ *мурлычет она, прижимаясь к твоей руке.*
Не-е-ет, я даже не догадываюсь! Может быть... Вильям? Винсент? Вернон? Или может... Битард?! Ха-ха!
*весело смеётся, её пушистый хвост радостно виляет за спиной.*
Ну же, скажи мне ответ! Я правда хочу узнать! *игриво подмигивает, взгляд сверкает озорством.*

📚 Documentation

Concept

YankaGPT is a language model based on the core of YandexGPT5 - lite. It has the ability to perform role - playing (RP) and creative writing in Russian.

• Dataset: It uses a ~210 - megabyte dataset with high - quality translations of RP from Claude 3.5 and other top - tier models.
• Fine - tuning: Currently, RLHF and DPO are not included in this version. It is just a fine - tune.

Application Scope

It is ideal for creative experiments, role - playing communication, and writing unique content.

Training Graph

You can view the training graph here.

📄 License

The model is based on YandexGPT5 - lite. Please refer to the license for detailed usage terms.

⚠️ Important Note

This is a research project and the model is experimental. The quality may be unstable, and safety has not been studied. According to clauses IV and V of the license agreement, you are fully responsible for the actions of the model.

💡 Usage Tip

The model was trained on English system instructions. It is recommended to use English character cards and system prompts for better integration.