AGmind/agmind-rag-splitter-ru

TEXT GENERATIONConcurrent Unit Cost:1Model Size:8BQuant:FP8Context Size:32kTool Calling:SupportedPublished:Jun 23, 2026License:apache-2.0Architecture:Transformer0.0K Open Weights Featherless Exclusive Cold

The AGmind/agmind-rag-splitter-ru is a LoRA fine-tuned version of t-tech/T-lite-it-2.1 (Qwen3-8B) specifically designed for Russian language RAG applications. This model excels at splitting Russian documents into semantically self-contained chunks, preserving tables and code blocks, and outputs split indices and topics in JSON format. It was trained on an expanded dataset of 28k entries from 6 sources, including legal and financial texts, and demonstrates improved boundary detection performance on out-of-domain legal documents.

Loading preview...

Overview

AGmind/agmind-rag-splitter-ru is a specialized model, fine-tuned from t-tech/T-lite-it-2.1 (a Qwen3-8B variant), to intelligently segment Russian documents for Retrieval-Augmented Generation (RAG) systems. Its core function is to break down text into meaningful, self-sufficient chunks, ensuring that structural elements like tables and code remain intact. The model outputs the split boundaries and a topic for each segment in a JSON format.

Key Capabilities

  • Context-Aware Splitting: Divides Russian documents into semantically coherent units, crucial for effective RAG.
  • Structural Integrity: Designed to prevent the fragmentation of tables, code blocks, and complete sentences across split points.
  • JSON Output: Provides split indices (1-indexed) and a topic for each segment, facilitating programmatic use.
  • Improved Performance: Version 2 (v2) shows significant improvements, particularly in boundary detection on legal documents, with a 100% JSON output rate and a 18% increase in boundary-F1@0 compared to v1 on a robust, human-verified legal benchmark.
  • Robust Training: Trained on an expanded dataset of 28,095 entries from 6 diverse sources (cultura, habr, synthetic, legal, financial news, educational documents), utilizing DeepSeek-V4-Flash with self-consistency voting for label generation.

Use Cases

This model is ideal for developers building RAG systems that process Russian text, especially those dealing with complex documents containing structured data. It ensures that retrieved chunks are self-contained and relevant, improving the quality of information retrieval and generation. The model is a completion-style model, expecting pre-split, numbered units as input.

Finetunes
2 models