d-michail/Mitos-Qwen3-14B-Dialogue

TEXT GENERATIONPricing:Input $0.48 / Output $0.96Concurrent Unit Cost:1Model Size:14BQuant:FP8Context Size:32kTool Calling:SupportedPublished:Oct 13, 2026License:cc-by-sa-4.0Architecture:Transformer Open Weights Featherless Exclusive Cold

d-michail/Mitos-Qwen3-14B-Dialogue is a 14 billion parameter Qwen3-based language model fine-tuned by d-michail for Greek e-government question answering. Optimized for short citizen-clerk dialogues, it provides concise answers (25-45 words) regarding Greek administrative procedures. This model excels at understanding and responding to queries about the National Registry of Administrative Procedures, making it suitable for research in Greek public administration QA.

Loading preview...

Mitos-Qwen3-14B-Dialogue Overview

Mitos-Qwen3-14B-Dialogue is a 14 billion parameter model based on Qwen3, specifically fine-tuned for question answering related to Greek e-government administrative procedures. Developed by d-michail, this model specializes in generating brief, clerk-like responses (approximately 25-45 words) to citizen queries.

Key Capabilities and Training

  • Specialized Domain: Adapted for Greek administrative procedures, using data from the National Registry of Administrative Procedures (ΕΜΔΔ / Mitos).
  • Dialogue-based Fine-tuning: Trained on 90,145 synthetic citizen-clerk dialogue records, generated by an LLM (Qwen3.8-27B) from the text of 4,407 procedures.
  • Concise Answering: Designed to provide short, direct answers, mimicking a clerk's response style.
  • Training Details: Fine-tuned using LoRA (rank 16, alpha 32) on all attention and MLP projections, with a maximum sequence length of 4096 tokens.

Intended Use and Limitations

  • Intended Use: Primarily for research in Greek public-administration question answering, both closed-book and with retrieval-augmented generation (RAG).
  • Not Legal Advice: Answers are not legally authoritative; a deployed system must verify sources and dates.
  • Fact Recall: While it learns the domain's register and format, closed-book key-fact recall remains low; retrieval significantly improves accuracy.
  • Data Snapshot: Trained on a snapshot of the registry (1-2 October 2026), meaning procedures may change and answers can become stale. Users must verify answers against live procedures.

Finetunes
2 models