AtesiT/qwen2.5-1.5b-ru-hallucination-detector

TEXT GENERATIONPricing:Input $0.04 / Cached $0.008 / Output $0.08Concurrent Unit Cost:1Model Size:1.5BQuant:BF16Context Size:32kTool Calling:SupportedPublished:Aug 10, 2026License:apache-2.0Architecture:Transformer0.0K Open Weights Featherless Exclusive Cold

AtesiT/qwen2.5-1.5b-ru-hallucination-detector is a 1.5 billion parameter Qwen2.5-Instruct based model fine-tuned by AtesiT to detect hallucinations in Russian language LLM responses. Utilizing QLoRA for efficient training, this model identifies whether an LLM's answer contradicts or is unsupported by a given context. It is specifically designed for fact-checking in RAG pipelines and automated LLM testing, offering high accuracy in binary classification of hallucinations.

Loading preview...

Overview

This model, developed by AtesiT, is a specialized detector for identifying hallucinations in Russian language Large Language Model (LLM) outputs. Built upon the Qwen2.5-1.5B-Instruct architecture and fine-tuned using QLoRA (4-bit quantization + LoRA adapters) with the Unsloth library, it determines if an LLM's response contradicts or is not supported by the provided factual context.

Key Capabilities

  • Hallucination Detection: Classifies LLM responses as either 0 (not a hallucination, consistent with context) or 1 (hallucination, contradictory or unsupported).
  • Input Format: Accepts a question, context (source of facts), and the LLM's answer as input.
  • High Accuracy: Achieves an Accuracy of 0.9835 and an F1-score of 0.9833 on its test set.
  • Robust Training Data: Trained on a automatically generated dataset derived from the Russian SberQuAD corpus, including diverse hallucination types (mismatch, perturbation, LLM-generated).

Good For

  • RAG Pipelines: Verifying the factual consistency of LLM answers before presentation to users.
  • Automated LLM Testing: Systematically evaluating the reliability of LLM agents.
  • Quality Monitoring: Continuously assessing the accuracy of LLM responses in production environments.
  • Dataset Generation: Creating silver-labeled datasets for training other models focused on factual correctness.

Limitations

  • Domain Specificity: Primarily trained on encyclopedic texts; performance may vary on specialized domains.
  • Context Length: Limited to a maximum input sequence length of 512 tokens.
  • Language: Exclusively trained and optimized for Russian language; not suitable for other languages.

Finetunes
1 models