README.md

metadata

datasets:
  - Quardo/wikipedia-turkish-qa
language:
  - tr
base_model:
  - Qwen/Qwen2-7B-Instruct
pipeline_tag: text-generation

TR-Qwen2-7B-003-241201

The third generation of TR-Qwen2, fine-tuned with the largest token count so far.

• 7B-001: Fine-tuned on 12 million tokens.
• 1.5B-002: Fine-tuned on 2 million tokens.
• 7B-003 (this version): Fine-tuned on 103 million tokens, or effectively 25.75 million tokens across 4 epochs (compared to just 1 epoch for the earlier versions).

The model has been fine-tuned with this system prompt only: Yardımsever bir yapay zeka asistanısın. Still the fine-tuning service is the same: Together.ai

---

Current Status: Importing the model from Together.ai

---

Note: This version may perform poorly with mathematical tasks, as its fine-tuning is based on Wikipedia-based data. Expect stronger cultural understanding instead (unlike 002) with actual capabilities to talk normally (unlike 001).