iraklixyz/georgian-sft-conversations

Natively Written Georgian SFT Conversations

A high-quality, general-purpose Supervised Fine-Tuning (SFT) dataset containing 56,676 rows of natively written multi-turn Georgian conversations. The dataset is specifically designed and formatted to train models for conversational chat, instruction following, and agent-like behaviors in the Georgian language.

As of June 2026, this is the largest cleaned, high-quality, natively written SFT conversation dataset available in the Georgian language.

Dataset Summary

To build a high-performance Georgian chat model, standard pre-training corpora (like Wikipedia and news dumps) must be complemented by conversational instruction SFT data. Many existing SFT datasets are translated from English, resulting in rigid, literal "translationese" phrasing that sounds unnatural to native speakers.

This dataset applies a hybrid native synthesis strategy with strict quality validation to guarantee absolute linguistic quality, idiomatic richness, and broad topic diversity:

1. Wikipedia Grounding (48.6%): Contexts sampled from pure Georgian Wikipedia articles are used to generate natural, informative conversations about geography, science, history, and local cultural figures.
2. News Grounding (24.1%): Contexts sampled from Georgian news articles are used to generate discussions about current events, technology news, business, and politics.
3. Purely Synthetic Instructions (27.3%): Synthesized instructions across a taxonomy of 25 diverse domains (including coding, math, debugging, roleplay, creative writing, legal documentation, and logical puzzles) to provide comprehensive training for utility and general reasoning.

Dataset Statistics

• Total Conversations: 56,676

Source Distribution

• Wikipedia: 27,544 rows (48.6%)
• Synthetic: 15,485 rows (27.3%)
• News: 13,627 rows (24.1%)

Length Profile Distribution

• Short (concise 1-turn response under 100 words): 7,819 (13.8%)
• Medium: 27,936 (49.3%)
• Long: 14,134 (24.9%)
• Super Long: 6,767 (11.9%)

Conversation Turn Distribution

• 1 Turn: 49,771 (87.8%)
• 2 Turns: 1,614 (2.8%)
• 3 Turns: 5,270 (9.3%)

Dataset Structure

The dataset is stored as a Snappy-compressed Parquet file with the following schema:

Column	Type	Description
id	string	Unique conversation identifier (ka_sft_X)
seed_id	int64	Reference to the seed context/category index
source	string	Source type (wikipedia, news, or synthetic)
category	string	Context category or Wikipedia article title
length_profile	string	The targeted length constraint profile (short, medium, long, super_long)
messages	list	Conversational turns list in standard Hugging Face/OpenAI format: [{"role": "user"/"assistant", "content": "..."}]
turns	int64	Number of user turns in the conversation

How to Use

You can load this dataset directly in Python using the Hugging Face datasets library:

from datasets import load_dataset

# Load the SFT dataset
dataset = load_dataset("iraklixyz/georgian-sft-conversations")

# Inspect a conversation example
sample = dataset["train"][0]
print(f"Source: {sample['source']}")
print(f"Category: {sample['category']}")
print(f"Number of Turns: {sample['turns']}")
print("Messages:")
for msg in sample["messages"]:
    print(f"  [{msg['role'].upper()}]: {msg['content'][:150]}...")

Creator Attribution & License

This dataset was created and generated by Irakli Maisuradze (@iraklixyz).

Distributed under the Creative Commons Attribution-ShareAlike 4.0 International (CC BY-SA 4.0) license.