Model Intelligence Sheet
thanatosdi/llama-3.2-3b-f1-reasoning-instruct-gguf overview
!image/png Llama-3.2-3B-F1-Reasoning-Instruct（a.k.a Formosa-1-Reasoning or F1-Reasoning） 是由 Twinkle AI 與 APMIC 合作開發，並在國家高速網路與計算中心技術指導之下，針對中華民國台灣語境與任務需求所微調之繁體中文語言模型，涵蓋法律、教育、生活應用等多元場景，並以高指令跟隨能力為目標進行強化。
Downloads
101
Likes
1
Pipeline
text-generation
Library
transformers
Visibility
Public
Access
Open
Repository Files & Downloads
3 files detected
Direct downloads for all repository files
Model Details Live
Metadata Inspector
Normalized metadata (stored in metadata_json)
{
"metadata": {},
"card_data": {
"license": "llama3.2",
"language": [
"en",
"zh"
],
"base_model": [
"meta-llama/Llama-3.2-3B",
"lianghsun/Llama-3.2-3B-F1-Base"
],
"library_name": "transformers",
"tags": [
"Taiwan",
"R.O.C",
"zhtw",
"SLM",
"Llama-32"
],
"datasets": [
"lianghsun/tw-reasoning-instruct",
"minyichen/tw-instruct-R1-200k",
"minyichen/tw_mm_R1"
],
"model-index": [
{
"name": "Llama-3.2-3B-F1-Reasoning-Instruct",
"results": [
{
"task": {
"type": "question-answering",
"name": "Single Choice Question"
},
"dataset": {
"type": "ikala/tmmluplus",
"name": "tmmlu+",
"config": "all",
"split": "test",
"revision": "c0e8ae955997300d5dbf0e382bf0ba5115f85e8c"
},
"metrics": [
{
"name": "single choice",
"type": "accuracy",
"value": 46.16,
"verified": false
}
]
},
{
"task": {
"type": "question-answering",
"name": "Single Choice Question"
},
"dataset": {
"type": "cais/mmlu",
"name": "mmlu",
"config": "all",
"split": "test",
"revision": "c30699e"
},
"metrics": [
{
"name": "single choice",
"type": "accuracy",
"value": 51.22,
"verified": false
}
]
},
{
"task": {
"type": "question-answering",
"name": "Single Choice Question"
},
"dataset": {
"type": "lianghsun/tw-legal-benchmark-v1",
"name": "tw-legal-benchmark-v1",
"config": "all",
"split": "test",
"revision": "66c3a5f"
},
"metrics": [
{
"name": "single choice",
"type": "accuracy",
"value": 34.92,
"verified": false
}
]
}
]
}
],
"metrics": [
"accuracy"
],
"frontmatter": {
"license": "llama3.2",
"language": [
"en",
"zh"
],
"base_model": [
"meta-llama/Llama-3.2-3B",
"lianghsun/Llama-3.2-3B-F1-Base"
],
"library_name": "transformers",
"tags": [
"Taiwan",
"R.O.C",
"zhtw",
"SLM",
"Llama-32"
],
"datasets": [
"lianghsun/tw-reasoning-instruct",
"minyichen/tw-instruct-R1-200k",
"minyichen/tw_mm_R1",
"name: Llama-3.2-3B-F1-Reasoning-Instruct",
"task:",
"name: single choice",
"task:",
"name: single choice",
"task:",
"name: single choice"
],
"metrics": [
"accuracy"
]
},
"hero_image_url": "https://img.shields.io/badge/Discord-Twinkle%20AI-7289da?logo=discord&logoColor=white&color=7289da",
"summary": "!image/png **Llama-3.2-3B-F1-Reasoning-Instruct**（a.k.a **Formosa-1-Reasoning** or **F1-Reasoning**） 是由 **Twinkle AI** 與 **APMIC** 合作開發，並在國家高速網路與計算中心技術指導之下，針對中華民國台灣語境與任務需求所微調之繁體中文語言模型，涵蓋法律、教育、生活應用等多元場景，並以高指令跟隨能力為目標進行強化。",
"quick_links": [],
"benchmark_table_html": "",
"readme_markdown": "---\nlicense: llama3.2\nlanguage:\n- en\n- zh\nbase_model:\n- meta-llama/Llama-3.2-3B\n- lianghsun/Llama-3.2-3B-F1-Base\nlibrary_name: transformers\ntags:\n- Taiwan\n- R.O.C\n- zhtw\n- SLM\n- Llama-32\ndatasets:\n- lianghsun/tw-reasoning-instruct\n- minyichen/tw-instruct-R1-200k\n- minyichen/tw_mm_R1\nmodel-index:\n- name: Llama-3.2-3B-F1-Reasoning-Instruct\n results:\n - task:\n type: question-answering\n name: Single Choice Question\n dataset:\n type: ikala/tmmluplus\n name: tmmlu+\n config: all\n split: test\n revision: c0e8ae955997300d5dbf0e382bf0ba5115f85e8c\n metrics:\n - name: single choice\n type: accuracy\n value: 46.16\n - task:\n type: question-answering\n name: Single Choice Question\n dataset:\n type: cais/mmlu\n name: mmlu\n config: all\n split: test\n revision: c30699e\n metrics:\n - name: single choice\n type: accuracy\n value: 51.22\n - task:\n type: question-answering\n name: Single Choice Question\n dataset:\n type: lianghsun/tw-legal-benchmark-v1\n name: tw-legal-benchmark-v1\n config: all\n split: test\n revision: 66c3a5f\n metrics:\n - name: single choice\n type: accuracy\n value: 34.92\nmetrics:\n- accuracy\n---\n\n# Model Card for Llama-3.2-3B-F1-Reasoning-Instruct (a.k.a __Formosa-1-Reasoning__ or __F1-Reasoning__)\n\n<div align=\"center\" style=\"line-height: 1;\">\n <a href=\"https://discord.gg/Cx737yw4ed\" target=\"_blank\" style=\"margin: 2px;\">\n <img alt=\"Discord\" src=\"https://img.shields.io/badge/Discord-Twinkle%20AI-7289da?logo=discord&logoColor=white&color=7289da\" style=\"display: inline-block; vertical-align: middle;\"/>\n </a>\n <a href=\"https://huggingface.co/twinkle-ai\" target=\"_blank\" style=\"margin: 2px;\">\n <img alt=\"Hugging Face\" src=\"https://img.shields.io/badge/%F0%9F%A4%97%20Hugging%20Face-Twinkle%20AI-ffc107?color=ffc107&logoColor=white\" style=\"display: inline-block; vertical-align: middle;\"/>\n </a>\n</div>\n\n<div align=\"center\" style=\"line-height: 1;\">\n <a href=\"https://huggingface.co/meta-llama/Llama-3.2-1B/blob/main/LICENSE.txt\" style=\"margin: 2px;\">\n <img alt=\"License\" src=\"https://img.shields.io/badge/License-llama3.2-f5de53?&color=0081fb\" style=\"display: inline-block; vertical-align: middle;\"/>\n </a>\n</div>\n\n\n\n<!-- Provide a quick summary of what the model is/does. -->\n**Llama-3.2-3B-F1-Reasoning-Instruct**（a.k.a **Formosa-1-Reasoning** or **F1-Reasoning**） 是由 **[Twinkle AI](https://huggingface.co/twinkle-ai)** 與 **[APMIC](https://www.apmic.ai/)** 合作開發，並在[國家高速網路與計算中心](https://www.nchc.org.tw/)技術指導之下，針對中華民國台灣語境與任務需求所微調之繁體中文語言模型，涵蓋法律、教育、生活應用等多元場景，並以高指令跟隨能力為目標進行強化。\n\n## Model Details\n\n### Model Description\n\n<!-- Provide a longer summary of what this model is. -->\n\n- **Developed by:** [Liang Hsun Huang](https://huggingface.co/lianghsun)、[Min Yi Chen](https://huggingface.co/minyichen)、[Wen Bin Lin](https://huggingface.co/tedslin)、[Chao Chun Chuang](https://huggingface.co/c00cjz00) & [Dave Sung](https://huggingface.co/k1dave6412) (All authors have contributed equally to this work.)\n- **Funded by:** [APMIC](https://www.apmic.ai/)\n- **Model type:** LlamaForCausalLM\n- **Language(s) (NLP):** Tranditional Chinese & English\n- **License:** [llama3.2](https://huggingface.co/meta-llama/Llama-3.2-1B/blob/main/LICENSE.txt)\n\n### Model Sources\n<!-- Provide the basic links for the model. -->\n\n- **Repository:** [twinkle-ai/Llama-3.2-3B-F1-Reasoning-Instruct](https://huggingface.co/twinkle-ai/Llama-3.2-3B-F1-Reasoning-Instruct)\n- **Paper:** (TBA)\n- **Demo:** [Playground](https://3b02.coolify.apmic.ai/)\n\n## Evaluation\n\n### Results\n\n下表採用 [🌟 Twinkle Eval](https://github.com/ai-twinkle/Eval) 評測框架\n| 模型 | 評測模式 | TMMLU+(%) | 台灣法律(%) | MMLU(%) | 測試次數 | 選項排序 |\n|------------------------------------|---------|----------------|----------------|----------------|---------|---------|\n| [mistralai/Mistral-Small-24B-Instruct-2501](https://huggingface.co/mistralai/Mistral-Small-24B-Instruct-2501) | box | 56.15 (±0.0172) | 37.48 (±0.0098) | 74.61 (±0.0154) | 3 | 隨機 |\n| [meta-llama/Llama-3.2-3B-Instruct](https://huggingface.co/meta-llama/Llama-3.2-3B-Instruct) | box | 15.49 (±0.0104) | 25.68 (±0.0200) | 6.90 (±0.0096) | 3 | 隨機 |\n| [meta-llama/Llama-3.2-3B-Instruct](https://huggingface.co/meta-llama/Llama-3.2-3B-Instruct) | pattern | 35.85 (±0.0174) | 32.22 (±0.0023) | 59.33 (±0.0168) | 3 | 隨機 |\n| [MediaTek-Research/Llama-Breeze2-3B-Instruct](https://huggingface.co/MediaTek-Research/Llama-Breeze2-3B-Instruct) | pattern | 40.32 (±0.0181) | 38.92 (±0.0193) | 55.37 (±0.0180) | 3 | 隨機 |\n| [twinkle-ai/Llama-3.2-3B-F1-Reasoning-Instruct](https://huggingface.co/twinkle-ai/Llama-3.2-3B-F1-Reasoning-Instruct) (ours) | box | 46.16 (±0.0198) | 34.92 (±0.0243) | 51.22 (±0.0206) | 3 | 隨機 |\n\n下表用 lighteval 評測框架\n| 模型 | MATH-500 | GPQA Diamond |\n|--------------------------------------------|----------|--------------|\n| [meta-llama/Llama-3.2-3B-Instruct](https://huggingface.co/meta-llama/Llama-3.2-3B-Instruct) | 44.40 | 27.78 |\n| [twinkle-ai/Llama-3.2-3B-F1-Reasoning-Instruct](https://huggingface.co/twinkle-ai/Llama-3.2-3B-F1-Reasoning-Instruct) (ours) | **51.40**| **33.84** |\n\n\n---\n\n## 🔧 Tool Calling\n\n本模型使用 Hermes 格式訓練，並支援平行呼叫（Parallel calling），以下為完整範例流程。\nTool call 模板已經為大家寫好放進 chat-template 了，Enjoy it！\n\n### 1️⃣ 啟動 vLLM 後端\n> **⚠️ 注意：需要 vLLM 版本 >= 0.8.3，否則 `enable-reasoning`、`enable-auto-tool-choice` 無法同時開啟**\n\n```bash\nvllm serve twinkle-ai/Llama-3.2-3B-F1-Reasoning-Instruct \\\n --port 8001 \\\n --enable-reasoning \\\n --reasoning-parser deepseek_r1 \\\n --enable-auto-tool-choice \\\n --tool-call-parser hermes\n```\n\n### 2️⃣ 定義工具（Functions）\n\n```python\ndef get_weather(location: str, unit: str):\n return f\"{location}的氣溫是{unit}26度，晴朗無風\"\n\ndef search(query: str):\n return \"川普終於宣布對等關稅政策，針對 18 個經濟體課徵一半的對等關稅，並從 4/5 起對所有進口產品徵收10%的基準關稅！美國將針對被認定為不當貿易行為(不公平貿易) 的國家，於 4/9 起課徵報復型對等關稅 (Discounted Reciprocal Tariff)，例如：日本將被課徵 24% 的關稅，歐盟則為 20%，以取代普遍性的 10% 關稅。\\n針對中國則開啟新一波 34% 關稅，並疊加於先前已實施的關稅上，這將使中國進口商品的基本關稅稅率達到 54%，而且這尚未包含拜登總統任內或川普第一任期所施加的額外關稅。加拿大與墨西哥則不適用這套對等關稅制度，但川普認為這些國家在芬太尼危機與非法移民問題尚未完全解決，因此計畫對這兩國的大多數進口商品施加 25% 關稅。另外原本針對汽車與多數其他商品的關稅豁免將於 4/2 到期。\\n台灣的部分，美國擬向台灣課徵32％的對等關稅，雖然並未針對晶片特別課徵關稅，但仍在記者會中提到台灣搶奪所有的電腦與半導體晶片，最終促成台積電對美國投資計劃額外加碼 1,000 億美元的歷史性投資；歐盟則課徵20％的對等關稅。最後是汽車關稅將於 4/2 起，對所有外國製造的汽車課徵25% 關稅。\"\n\ntools = [\n {\n \"type\": \"function\",\n \"function\": {\n \"name\": \"get_weather\",\n \"description\": \"Get the current weather in a given location\",\n \"parameters\": {\n \"type\": \"object\",\n \"properties\": {\n \"location\": {\"type\": \"string\", \"description\": \"國家或城市名, e.g., 'Taipei'、'Jaipei'\"},\n \"unit\": {\"type\": \"string\", \"description\": \"氣溫單位，亞洲城市使用攝氏；歐美城市使用華氏\", \"enum\": [\"celsius\", \"fahrenheit\"]}\n },\n \"required\": [\"location\", \"unit\"]\n }\n }\n },\n {\n \"type\": \"function\",\n \"function\": {\n \"name\": \"search\",\n \"description\": \"這是一個類似 Google 的搜尋引擎，關於知識、天氣、股票、電影、小說、百科等等問題，如果你不確定答案就搜尋一下。\",\n \"parameters\": {\n \"type\": \"object\",\n \"properties\": {\n \"query\": {\"type\": \"string\", \"description\": \"should be a search query, e.g., '2024 南韓 戒嚴'\"}\n },\n \"required\": [\"query\"]\n }\n }\n }\n]\n```\n\n### 3️⃣ 執行工具調用（Tool Calls）\n\n> **⚠️ 注意：system_prompt 可以不用帶，除非是需要時間基準的工具。**\n```python\nresponse = client.chat.completions.create(\n model=client.models.list().data[0].id,\n messages=[\n {\"role\": \"system\", \"content\": \"記住你的知識截止於 2024/12，今天是 2025/4/7\"},\n {\"role\": \"user\", \"content\": \"台北氣溫如何? 另外，告訴我川普最新關稅政策\"},\n ],\n max_tokens=1500,\n temperature=0.6,\n top_p=0.95,\n tools=tools,\n tool_choice=\"auto\"\n)\n\nprint(response.choices[0].message.reasoning_content)\nprint(response.choices[0].message.tool_calls)\n```\n\n#### 🧠 推理內容輸出（僅顯示部分）\n> 好的，我需要幫助這個使用者解決他們的問題。他們問了兩件事：首先，臺北市的天氣情況，以及第二，關於川普最近的關稅政策。 \n> 對於第一部分，他們提到了“臺北”，所以應該呼叫 get_weather 函式… \n> 接下來是關於川普的新關稅政策… \n> 總結一下，我需要分別進行兩次 API 呼叫，每次都有各自正確填寫的參數…\n\n#### ⚙️ Tool Calls List\n\n\n```json\n[ChatCompletionMessageToolCall(id='chatcmpl-tool-35e74420119349999913a10133b84bd3', function=Function(arguments='{\"location\": \"Taipei\", \"unit\": \"celsius\"}', name='get_weather'), type='function'), ChatCompletionMessageToolCall(id='chatcmpl-tool-7ffdcb98e59f4134a6171defe7f2e31b', function=Function(arguments='{\"query\": \"Donald Trump latest tariffs policy\"}', name='search'), type='function')]\n```\n\n### 4️⃣ 產生最終回答\n\n```python\nresponse = client.chat.completions.create(\n model=client.models.list().data[0].id,\n messages=[\n {\"role\": \"system\", \"content\": \"記住你的知識截止於 2024/12，今天是 2025/4/7\"},\n {\"role\": \"user\", \"content\": \"台北氣溫如何? 另外，告訴我川普最新關稅政策\"},\n {\n \"role\": \"assistant\",\n \"content\": \"\",\n \"tool_calls\": [\n {\n \"id\": response.choices[0].message.tool_calls[0].id,\n \"type\": \"function\",\n \"function\": {\n \"name\": response.choices[0].message.tool_calls[0].function.name,\n \"arguments\": response.choices[0].message.tool_calls[0].function.arguments\n }\n },\n {\n \"id\": response.choices[0].message.tool_calls[1].id,\n \"type\": \"function\",\n \"function\": {\n \"name\": response.choices[0].message.tool_calls[1].function.name,\n \"arguments\": response.choices[0].message.tool_calls[1].function.arguments\n }\n }\n ]\n },\n {\n \"role\": \"tool\",\n \"content\": search(**json.loads(response.choices[0].message.tool_calls[0].function.arguments)),\n \"tool_call_id\": response.choices[0].message.tool_calls[0].id # tool_call_id 必須要帶，才能正確配對 工具 及 tool_call\n },\n {\n \"role\": \"tool\",\n \"content\": get_weather(**json.loads(response.choices[0].message.tool_calls[1].function.arguments)),\n \"tool_call_id\": response.choices[0].message.tool_calls[1].id # tool_call_id 必須要帶，才能正確配對 工具 及 tool_call\n }\n ],\n max_tokens=1500,\n temperature=0.6,\n top_p=0.95,\n tools=tools,\n tool_choice=\"auto\"\n)\n\nprint(response.choices[0].message.reasoning_content)\nprint(response.choices[0].message.content)\n```\n\n#### 🧠 推理內容輸出（僅顯示部分）\n> 首先，我需要處理使用者的查詢，他們要求了解臺北市的當下氣溫以及川普最近的關稅政策… \n> 在呼叫了 `get_weather` 後得到了臺北市氣溫為 26 度（攝氏）… \n> 接著，使用 `search` 搜尋「川普最新關稅政策 2025」… \n> 整合後，我提供如下摘要：\n\n#### 📋 最終輸出內容\n\n```text\n以下是您請求的資訊：\n\n**臺北市氣溫**\n- 目前的氣溫為 **26°C**（攝氏）\n- 天候狀況：晴朗無風\n\n**川普最新關稅政策概述**\n1. **對等關稅政策** \n - 對 18 個經濟體課徵 50% 的對等關稅 \n - 自 4 月 5 日起，所有進口產品全面徵收 10% 基本關稅 \n\n2. **報復型對等關稅** \n - 日本 24%、歐盟 20% \n\n3. **對中國的高額關稅** \n - 增加至 54%（原有關稅 + 新增 34%） \n\n4. **特殊案例** \n - 加拿大與墨西哥不適用，但其他商品課徵 25% \n - 汽車與部分商品的免稅即將到期 \n\n5. **對台灣的影響** \n - 美國計畫對台灣課徵 32% 關稅，但晶片暫無額外課稅 \n\n6. **全球視角** \n - 歐盟與日本關稅比例相對較高\n```\n\n\n## Citation\n\n<!-- If there is a paper or blog post introducing the model, the APA and Bibtex information for that should go in this section. -->\n```yaml\n@misc{twinkleai2025llama3.2f1,\n title = {Llama-3.2-3B-F1-Reasoning-Instruct: A Traditional Chinese Instruction-Tuned Reasoning Language Model for Taiwan},\n author = {Huang, Liang Hsun and Chen, Min Yi and Lin, Wen Bin and Chuang, Chao Chun and Sung, Dave},\n year = {2025},\n howpublished = {\\url{https://huggingface.co/twinkle-ai/Llama-3.2-3B-F1-Instruct}},\n note = {Twinkle AI and APMIC. All authors contributed equally.}\n}\n```\n\n## Acknowledge\n- 特此感謝[國家高速網路與計算中心](https://www.nchc.org.tw/)的指導與 [APMIC](https://www.apmic.ai/) 的算力支援，才得以讓本專案訓利完成。\n- 特此致謝黃啟聖老師、許武龍（哈爸）、臺北市立第一女子高級中學物理科陳姿燁老師、[奈視科技](https://nanoseex.com/) CTO Howard、[AIPLUX Technology](https://aiplux.com/)、郭家嘉老師以及所有在資料集製作過程中提供寶貴協助的夥伴。\n\n## Model Card Authors\n\n[Twinkle AI](https://huggingface.co/twinkle-ai)\n\n## Model Card Contact\n\n[Twinkle AI](https://huggingface.co/twinkle-ai)",
"related_quantizations": []
},
"tags": [
"transformers",
"gguf",
"llama",
"text-generation",
"Taiwan",
"R.O.C",
"zhtw",
"SLM",
"Llama-32",
"en",
"zh",
"dataset:lianghsun/tw-reasoning-instruct",
"dataset:minyichen/tw-instruct-R1-200k",
"dataset:minyichen/tw_mm_R1",
"base_model:lianghsun/Llama-3.2-3B-F1-Base",
"base_model:quantized:lianghsun/Llama-3.2-3B-F1-Base",
"license:llama3.2",
"model-index",
"endpoints_compatible",
"region:us",
"conversational"
],
"likes": 1,
"downloads": 101,
"gated": false,
"private": false,
"last_modified": "2025-05-06T17:42:26.000Z",
"created_at": "2025-05-06T17:42:12.000Z",
"pipeline_tag": "text-generation",
"library_name": "transformers"
}
Source payload excerpt (from Hugging Face API)
{
"_id": "681a49f4c6eed6d87d484a41",
"id": "ThanatosDi/Llama-3.2-3B-F1-Reasoning-Instruct-GGUF",
"modelId": "ThanatosDi/Llama-3.2-3B-F1-Reasoning-Instruct-GGUF",
"sha": "06cd6bf54ea06182dd5cb241e892872c427c5ce1",
"createdAt": "2025-05-06T17:42:12.000Z",
"lastModified": "2025-05-06T17:42:26.000Z",
"author": "ThanatosDi",
"downloads": 101,
"likes": 1,
"gated": false,
"private": false,
"pipeline_tag": "text-generation",
"library_name": "transformers",
"siblings_count": 6
}