Downloads · 30 days
12
31% of all-time downloads
agentlans/Qwen2.5-3B-Instruct-zhtw
Qwen2.5-3B-Instruct-zhtw is a machine learning model from agentlans. Use it for the machine learning task on the model card, and read the license before you ship it in a product.
Downloads · 30 days
12
31% of all-time downloads
All-time downloads
39
Public
Parameters
3.1B
12.3 GB on disk
Likes
0
Public
Click a slice to open those files.
.safetensors6.2 GB · 100%
From the Hugging Face model README
This model is a fine-tuned version of Qwen2.5-3B-Instruct, specifically optimized for generating Traditional Chinese text through Direct Preference Optimization (DPO) training. It focuses on Taiwanese vocabulary and expressions.
Objective: To guide Qwen2.5-3B-Instruct in consistently generating Traditional Chinese text with a preference for Taiwanese expressions over Simplified Chinese.
DPO Dataset: The model was trained on a version of the shibing624/alpaca-zh dataset adapted for Traditional Chinese. Originally in Simplified Chinese, the dataset was converted to Traditional Chinese and enriched with Taiwan-specific vocabulary using OpenCC.
Training Configuration:
Merging Method: DPO training was repeated three times on the same dataset, using different sample orders. After each training run, the resulting DPO model was sparsified to a density of 0.05, and then merged with the original Qwen2.5-3B-Instruct model using TIES. This ensures that the behavior of turn-ending in generated text is accurate.
Dataset Purity: No new information beyond the converted dataset was added during the training process. This was a deliberate choice to minimize interference with the original knowledge base of the Qwen model.
This model is suitable for applications that require the generation of Traditional Chinese text, especially in scenarios where Taiwanese vocabulary and expressions are needed.
Core Knowledge: The model's knowledge base is largely consistent with the original Qwen2.5-3B-Instruct model.
Vocabulary Consistency: Despite DPO training, the model may occasionally produce responses that contain Simplified Chinese or Mainland Chinese vocabulary (e.g., "鼠标" instead of the preferred Taiwanese "滑鼠"). This inconsistency may need further handling in certain applications.
該模型是 Qwen2.5-3B-Instruct 的微調版本,專為產生繁體中文文本而進行直接偏好優化 (DPO) 訓練,並特別關注台灣的詞彙和表達方式。
目標:指導 Qwen2.5-3B-Instruct 一致地生成繁體中文文本,並優先使用台灣的表達方式,而非簡體中文。
DPO 資料集:該模型基於 shibing624/alpaca-zh 資料集的繁體中文改編版本進行訓練。資料集原本為簡體中文,經過轉換為繁體中文後,並使用 OpenCC 融入台灣特有的詞彙。
訓練配置:
合併方法:DPO 訓練在相同的資料集上重複三次,並使用不同的樣本順序。每個訓練後的 DPO 模型會被稀疏化到 0.05 的密度,然後與原始的 Qwen2.5-3B-Instruct 透過 TIES 進行合併。這樣做是為了確保生成的文字中回合結束的行為正確。
資料純度:在訓練過程中,未加入轉換後資料集以外的任何新資訊。這是一個深思熟慮的選擇,旨在最大限度減少對 Qwen 原有知識庫的干擾。
該模型適用於需要生成繁體中文文本的應用程式,尤其是那些需要使用台灣詞彙和表達方式的應用場景。
核心知識:此模型的知識庫基本與原始 Qwen2.5-3B-Instruct 相同。
詞彙一致性:儘管經過 DPO 訓練,模型有時仍會產生包含簡體中文或中國大陸詞彙的回應(例如「鼠標」),而非首選的台灣詞彙(例如「滑鼠」)。這種不一致在某些應用中可能需要進一步處理。
隨著技術不斷進步,人工智慧已經成為現代社會發展的重要驅動力之一。它不僅能夠幫助人類解決複雜問題、提高生產效率和創新體驗,還潛在地改變了我們的生活方式與工作模式。通過學習大規模數據並進行歸納推理,人工智慧系統能夠越來越精準地預測未來趨勢,為各種應用場景帶來無數可能性。然而,儘管人工智慧帶來諸多好處,也引發了一些倫理道德及隱私保護等方面的挑戰,這些都要求我們持續關注其開發與應用過程中的平衡點。總體來說,人工智慧是一個充滿變革與機遇的新時代關鍵領域。
</details>