羲和 HeliosLM
Collection
Open Traditional Chinese (Taiwan) chat, safety and speech models. 台灣繁體中文對話、安全與語音模型。 • 9 items • Updated
The dataset is currently empty. Upload or create new data files. Then, you will be able to explore them in the Dataset Viewer.
本頁列出我們的評測設定,以及用這套設定自行測得的分數。HeliosLM 以 Google Gemma 4 為基礎,下表同時列出原版 Gemma 4 作為基準。
題目:TMMLU+ 測試集抽 800 題、C-Eval 驗證集抽 400 題、CMMLU 測試集抽 400 題(固定亂數種子 1),以及 2026 年國家考試選擇題抽 800 題(tw2026,種子 2)。 設定:vLLM,temperature 0,最多生成 1,536 tokens。沒有輸出可辨識答案的題目算錯。
| 模型 | TMMLU+ | C-Eval | CMMLU | tw2026 |
|---|---|---|---|---|
| google/gemma-4-31B-it(基準) | 85.1 | 79.3 | 85.0 | 77.6 |
| google/gemma-4-26B-A4B-it(基準) | 81.4 | 75.3 | 80.0 | 71.5 |
| google/gemma-4-E4B-it(基準) | 67.1 | 65.8 | 67.0 | 62.5 |
表中的 Gemma 4 為 Google 原版模型(HeliosLM 的基礎模型),不是 HeliosLM。HeliosLM 各模型的分數列在各自的模型頁。