AutoGEO_mini_Qwen1.7B_ResearchyGEO

A lightweight web-document rewriting model fine-tuned with GRPO (reinforcement learning) from Qwen3-1.7B, developed as part of the AutoGEO framework introduced in:

WHAT GENERATIVE SEARCH ENGINES LIKE AND HOW TO OPTIMIZE WEB CONTENT COOPERATIVELY
Paper (arXiv): https://arxiv.org/abs/2510.11438

---

What this model does

AutoGEO_mini_Qwen1.7B_ResearchyGEO rewrites raw web documents into improved versions that are better aligned with generative search engines’ preferences for Researchy-GEO dataset.

In our experiments/usage:

• The total cost is about 0.0071× the cost of gemini-2.5-pro for comparable rewriting workloads.
• Rewritten documents achieve significant improvements in GEO metrics.

---

Training summary

• Base model: Qwen3-1.7B
• Method: GRPO-based reinforcement learning fine-tuning
• Task: Rewrite original web documents to improve GEO metrics (per the AutoGEO framework in the paper above)

---

Repository contents

This repository includes the standard inference artifacts (e.g., model.safetensors, config.json, tokenizer.json, chat_template.jinja, etc.) required to load and run the model with transformers.

---