Downloads · 30 days
0
tony-shieh/Qwen2.5-7B-GRPO-TW-QA
Qwen2.5-7B-GRPO-TW-QA is a text generation model from tony-shieh. Use it when you need the model to write or continue text.
This is the ultimate version of our Traditional Chinese QA reasoning model, trained using Group Relative Policy Optimization (GRPO). It is built on top of our SFT baseline and is specifically optimized for rigorous lo…
Downloads · 30 days
0
Access
Public
Updated Jun 20, 2026
Repo size
334 MB
Likes
0
Public
Click a slice to open those files.
.safetensors323 MB · 97%
From the Hugging Face model README
This is the ultimate version of our Traditional Chinese QA reasoning model, trained using Group Relative Policy Optimization (GRPO). It is built on top of our SFT baseline and is specifically optimized for rigorous logical deduction and the process of elimination.
By applying GRPO, the model was trained to maximize reward functions based on both format adherence (strict XML tags) and correctness (arriving at the accurate final option). This prevents the model from simply memorizing answers and forces it to develop genuine step-by-step reasoning capabilities.
During the GRPO phase, the model generated 4 candidate completions per prompt (num_generations=4) to calculate relative advantages. The reward mechanism was split into:
<reasoning>...</reasoning><answer>X</answer>.To maximize the reasoning capabilities of this RL-aligned model, we highly recommend the following decoding parameters to prevent degeneration loops while maintaining strict logic:
0.2 (Low temperature for stable, greedy-leaning logic)1024 (Give the model enough space to complete its reasoning process)True你是一位具備頂尖邏輯的AI。請嚴密分析選項,使用逐步排除法。
【嚴格輸出格式】
你必須「精準」遵守以下XML格式,絕不能省略標籤:
<reasoning>
寫下你的逐步推論過程...
</reasoning>
<answer>單一正確字母(A,B,C或D)</answer>
題目:[Insert Question]
A. [Option A]
B. [Option B]
C. [Option C]
D. [Option D]
請嚴密推理並選出正確答案。