Qwen-2.5-3B-cot-zh-tw (GRPO)

模型簡介 | Model Overview

本模型基於 Qwen-2.5-3B-Instruct，專為繁體中文數學/邏輯推理場景設計，不是單純仿製長鏈推理，而是經由創新 RLHF 訓練流程，讓模型自發產生類似「aha moment」的推理能力。

訓練流程為：

以少量高品質多步推理 SFT 數據冷啟動模型
接著採用 GRPO 策略與自設獎勵函數，在「問題—答案」數據集下強化模型「自發推理」與「步驟合理性」，不依賴 PPO 或大規模 SFT
推理長度經本地測試優化，步驟數落在最佳甜蜜點，適合部署於一般 GPU/邊緣裝置

This model is based on Qwen-2.5-3B-Instruct and optimized for step-by-step reasoning in Traditional Chinese math and logic tasks. Instead of standard CoT SFT or PPO, we use a minimal SFT “cold start” with high-quality reasoning samples, then apply GRPO with a custom reward function to let the model discover its own “aha moments” and multi-step reasoning chains—all while controlling output length for efficient local deployment.

訓練動機 | Motivation

這個專案純粹出於個人興趣與繁體中文圈的實際需求而開發，旨在讓本地端能有更好用的推理模型。我希望能為繁體中文圈貢獻可以本地部署的推理模型，讓模型在繁中語境下，也能自發產生多步驟、具備頓悟感（aha moment）的解題過程。整體訓練過程強調實驗精神與實用導向：只用少量優質 SFT 冷啟動，再透過 GRPO 與自定獎勵函數，讓模型自己學會思考與拆解問題，追求可實際落地的推理效果。

This project was developed purely out of personal interest and the practical needs of the Traditional Chinese-speaking community, with the goal of creating a more effective locally deployable reasoning model. I hope to contribute a model that enables multi-step, “aha moment” reasoning in the Traditional Chinese context—one that can be run efficiently on local hardware. The entire training process emphasizes both experimentation and real-world usability: starting with a small amount of high-quality SFT for cold start, then leveraging GRPO and custom reward functions to encourage the model to learn reasoning and problem decomposition on its own, with a focus on truly applicable step-by-step solutions.

模型特性 | Key Features

Aha Moment 自發推理：非模板複製，而是訓練模型「自己發現推理步驟」

步驟最佳化：推理長度經本地實測，控制在「解題效率」與「可讀性」甜蜜點

繁體中文強化：涵蓋台灣、港澳常用語境，數學、邏輯均表現穩定

適用本地端部署：硬體需求親民，適合一般 4GB GPU，步驟控制精簡，不會無限發散

Self-generated Reasoning: Not pattern imitation—model “discovers” solution steps

Optimized Step Count: Output length tuned for real-world efficiency and readability

Traditional Chinese Enhanced: Robust for Taiwan/HK/Macau context and math/logic QA

Local Deployment Friendly: Runs on standard consumer GPUs, with step count sweet spot

訓練細節 | Training Details

基礎模型 / Base：Qwen2.5-3B-Instruct

流程 / Pipeline：少量高品質多步推理 SFT（冷啟動）→ GRPO（自訂獎勵函數，自發推理強化）

數據來源 / Data：DoggiAI/GSM8K_zh_tw + 自建繁體推理 Q&A (From Grok API Distill)

RLHF 核心：獎勵重點放在答案正確率、推理步驟合理性與精簡性，不靠人類標記每步驟

硬體 / Hardware：L4 GPU, 訓練 24 小時

框架 / Framework：Transformers, PEFT, bitsandbytes, Unsloth

使用建議 | Usage Tips

推薦應用：數學解題、邏輯題、逐步問答適合使用類似：「請自行分步推理，說明每一步的原因。」等提示語可將 output 步驟數量根據需求進行控制（建議 2~6 步為最佳）

快速上手 | Quickstart

print("Hello, world!")
from transformers import AutoModelForCausalLM, AutoTokenizer

model = AutoModelForCausalLM.from_pretrained("BryanADA/qwen-2.5-3b-cot-zh-tw")
tokenizer = AutoTokenizer.from_pretrained("BryanADA/qwen-2.5-3b-cot-zh-tw")

prompt = "小明有 3 顆蘋果，又拿到 2 顆，一共幾顆？請分步說明推理過程。"
inputs = tokenizer(prompt, return_tensors="pt")
outputs = model.generate(
    **inputs,
    max_new_tokens=512,
    do_sample=True,
    temperature=0.7,
    top_p=0.95,
    pad_token_id=tokenizer.eos_token_id
)
print(tokenizer.decode(outputs[0], skip_special_tokens=True))

參考資料 | References

Qwen 官方

Deepseek R1 論文

DoggiAI/GSM8K_zh_tw

RLHF/GRPO 相關文獻

License

本模型採用 Apache-2.0 授權，允許用於研究、學術及商業用途。請遵循授權條款保留原作者版權及免責聲明。

This model is licensed under Apache-2.0, allowing use for research, academic, and commercial purposes. Please comply with the license terms and retain the original copyright and disclaimers.

BryanADA
/

Qwen2.5-3B-cot-zh-tw