d-michail/Mitos-Qwen3-8B-Dialogue

TEXT GENERATIONPricing:Input $0.468 / Output $1.82Concurrent Unit Cost:1Model Size:8BQuant:FP8Context Size:32kTool Calling:SupportedPublished:Oct 13, 2026License:cc-by-sa-4.0Architecture:Transformer Open Weights Featherless Exclusive Cold

d-michail/Mitos-Qwen3-8B-Dialogue is an 8 billion parameter Qwen3-based language model fine-tuned for Greek e-government question answering. It specializes in providing brief, clerk-like responses to inquiries about Greek administrative procedures, leveraging a dataset of synthetic citizen-clerk dialogues. This model is optimized for research in Greek public administration QA, supporting both closed-book and retrieval-augmented generation scenarios.

Loading preview...

Mitos-Qwen3-8B-Dialogue Overview

Mitos-Qwen3-8B-Dialogue is an 8 billion parameter model based on Qwen3, specifically adapted for question answering related to Greek e-government administrative procedures. It was fine-tuned using LoRA on a dataset of 90,145 synthetic citizen-clerk dialogues, generated by Qwen3.8-27B from the National Registry of Administrative Procedures (ΕΜΔΔ / Mitos).

Key Capabilities

  • Specialized Greek QA: Designed to answer questions about Greek administrative procedures in a concise, clerk-like manner (25-45 words).
  • Domain-Specific Training: Trained on a comprehensive snapshot of all procedures from the Mitos registry, ensuring deep domain knowledge.
  • Research Focus: Intended for research in Greek public administration question answering, supporting both closed-book and retrieval-augmented generation (RAG) approaches.
  • Qwen3 Chat Template: Utilizes the Qwen3 chat template with thinking disabled, expecting questions within the user message.

Intended Use and Limitations

This model is suitable for research in Greek public administration QA. While it learns the domain's register and format, its closed-book factual recall remains low, making RAG crucial for accuracy. It is not legal advice; answers are not legally authoritative and must be verified against live procedures. The training data is synthetic and from a specific snapshot, meaning the model can go stale and may reproduce generator errors. Users should implement repetition penalties or stop rules to prevent degeneration into repetition loops.

Finetunes
2 models