richarderkhov/p208p2002_-_llama-3-zhtw-8b-gguf overview

{
  "metadata": {},
  "card_data": {
    "frontmatter": {},
    "hero_image_url": "",
    "summary": "在 Llama 3 上試驗中文 Continue Pretraining (CP)，共計訓練 800M tokens。 由於中文預訓練語料品質還有改進空間，CP 後表現未能超越原版 Llama 3，我們比較幾個開源社群訓練的中文 Llama 3 也有類似狀況。 在英文方面 LLaMA 3 zhtw 使用 FineWeb，使得 MMLU 表現高於其他中文CP模型，能力與原版 LLaMA 3 持平。",
    "quick_links": [],
    "benchmark_table_html": "",
    "readme_markdown": "Quantization made by Richard Erkhov.\n\n[Github](https://github.com/RichardErkhov)\n\n[Discord](https://discord.gg/pvy7H8DZMG)\n\n[Request more models](https://github.com/RichardErkhov/quant_request)\n\n\nllama-3-zhtw-8B - GGUF\n- Model creator: https://huggingface.co/p208p2002/\n- Original model: https://huggingface.co/p208p2002/llama-3-zhtw-8B/\n\n\n| Name | Quant method | Size |\n| ---- | ---- | ---- |\n| [llama-3-zhtw-8B.Q2_K.gguf](https://huggingface.co/RichardErkhov/p208p2002_-_llama-3-zhtw-8B-gguf/blob/main/llama-3-zhtw-8B.Q2_K.gguf) | Q2_K | 2.96GB |\n| [llama-3-zhtw-8B.IQ3_XS.gguf](https://huggingface.co/RichardErkhov/p208p2002_-_llama-3-zhtw-8B-gguf/blob/main/llama-3-zhtw-8B.IQ3_XS.gguf) | IQ3_XS | 3.28GB |\n| [llama-3-zhtw-8B.IQ3_S.gguf](https://huggingface.co/RichardErkhov/p208p2002_-_llama-3-zhtw-8B-gguf/blob/main/llama-3-zhtw-8B.IQ3_S.gguf) | IQ3_S | 3.43GB |\n| [llama-3-zhtw-8B.Q3_K_S.gguf](https://huggingface.co/RichardErkhov/p208p2002_-_llama-3-zhtw-8B-gguf/blob/main/llama-3-zhtw-8B.Q3_K_S.gguf) | Q3_K_S | 3.41GB |\n| [llama-3-zhtw-8B.IQ3_M.gguf](https://huggingface.co/RichardErkhov/p208p2002_-_llama-3-zhtw-8B-gguf/blob/main/llama-3-zhtw-8B.IQ3_M.gguf) | IQ3_M | 3.52GB |\n| [llama-3-zhtw-8B.Q3_K.gguf](https://huggingface.co/RichardErkhov/p208p2002_-_llama-3-zhtw-8B-gguf/blob/main/llama-3-zhtw-8B.Q3_K.gguf) | Q3_K | 3.74GB |\n| [llama-3-zhtw-8B.Q3_K_M.gguf](https://huggingface.co/RichardErkhov/p208p2002_-_llama-3-zhtw-8B-gguf/blob/main/llama-3-zhtw-8B.Q3_K_M.gguf) | Q3_K_M | 3.74GB |\n| [llama-3-zhtw-8B.Q3_K_L.gguf](https://huggingface.co/RichardErkhov/p208p2002_-_llama-3-zhtw-8B-gguf/blob/main/llama-3-zhtw-8B.Q3_K_L.gguf) | Q3_K_L | 4.03GB |\n| [llama-3-zhtw-8B.IQ4_XS.gguf](https://huggingface.co/RichardErkhov/p208p2002_-_llama-3-zhtw-8B-gguf/blob/main/llama-3-zhtw-8B.IQ4_XS.gguf) | IQ4_XS | 4.18GB |\n| [llama-3-zhtw-8B.Q4_0.gguf](https://huggingface.co/RichardErkhov/p208p2002_-_llama-3-zhtw-8B-gguf/blob/main/llama-3-zhtw-8B.Q4_0.gguf) | Q4_0 | 4.34GB |\n| [llama-3-zhtw-8B.IQ4_NL.gguf](https://huggingface.co/RichardErkhov/p208p2002_-_llama-3-zhtw-8B-gguf/blob/main/llama-3-zhtw-8B.IQ4_NL.gguf) | IQ4_NL | 4.38GB |\n| [llama-3-zhtw-8B.Q4_K_S.gguf](https://huggingface.co/RichardErkhov/p208p2002_-_llama-3-zhtw-8B-gguf/blob/main/llama-3-zhtw-8B.Q4_K_S.gguf) | Q4_K_S | 4.37GB |\n| [llama-3-zhtw-8B.Q4_K.gguf](https://huggingface.co/RichardErkhov/p208p2002_-_llama-3-zhtw-8B-gguf/blob/main/llama-3-zhtw-8B.Q4_K.gguf) | Q4_K | 4.58GB |\n| [llama-3-zhtw-8B.Q4_K_M.gguf](https://huggingface.co/RichardErkhov/p208p2002_-_llama-3-zhtw-8B-gguf/blob/main/llama-3-zhtw-8B.Q4_K_M.gguf) | Q4_K_M | 4.58GB |\n| [llama-3-zhtw-8B.Q4_1.gguf](https://huggingface.co/RichardErkhov/p208p2002_-_llama-3-zhtw-8B-gguf/blob/main/llama-3-zhtw-8B.Q4_1.gguf) | Q4_1 | 4.78GB |\n| [llama-3-zhtw-8B.Q5_0.gguf](https://huggingface.co/RichardErkhov/p208p2002_-_llama-3-zhtw-8B-gguf/blob/main/llama-3-zhtw-8B.Q5_0.gguf) | Q5_0 | 5.21GB |\n| [llama-3-zhtw-8B.Q5_K_S.gguf](https://huggingface.co/RichardErkhov/p208p2002_-_llama-3-zhtw-8B-gguf/blob/main/llama-3-zhtw-8B.Q5_K_S.gguf) | Q5_K_S | 5.21GB |\n| [llama-3-zhtw-8B.Q5_K.gguf](https://huggingface.co/RichardErkhov/p208p2002_-_llama-3-zhtw-8B-gguf/blob/main/llama-3-zhtw-8B.Q5_K.gguf) | Q5_K | 5.34GB |\n| [llama-3-zhtw-8B.Q5_K_M.gguf](https://huggingface.co/RichardErkhov/p208p2002_-_llama-3-zhtw-8B-gguf/blob/main/llama-3-zhtw-8B.Q5_K_M.gguf) | Q5_K_M | 5.34GB |\n| [llama-3-zhtw-8B.Q5_1.gguf](https://huggingface.co/RichardErkhov/p208p2002_-_llama-3-zhtw-8B-gguf/blob/main/llama-3-zhtw-8B.Q5_1.gguf) | Q5_1 | 5.65GB |\n| [llama-3-zhtw-8B.Q6_K.gguf](https://huggingface.co/RichardErkhov/p208p2002_-_llama-3-zhtw-8B-gguf/blob/main/llama-3-zhtw-8B.Q6_K.gguf) | Q6_K | 6.14GB |\n| [llama-3-zhtw-8B.Q8_0.gguf](https://huggingface.co/RichardErkhov/p208p2002_-_llama-3-zhtw-8B-gguf/blob/main/llama-3-zhtw-8B.Q8_0.gguf) | Q8_0 | 7.95GB |\n\n\n\n\nOriginal model description:\n---\ndatasets:\n- HuggingFaceFW/fineweb\n- erhwenkuo/c4-chinese-zhtw\n- erhwenkuo/wikipedia-zhtw\n- p208p2002/wudao\n- p208p2002/NDLTD-T10-90-111\n- codeparrot/github-code-clean\nlanguage:\n- en\n- zh\nlicense: llama3\n---\n# Llama 3 zhtw\n\n在 Llama 3 上試驗中文 Continue Pretraining (CP)，共計訓練 800M tokens。\n\n由於中文預訓練語料品質還有改進空間，CP 後表現未能超越原版 Llama 3，我們比較幾個開源社群訓練的中文 Llama 3 也有類似狀況。\n\n在英文方面 LLaMA 3 zhtw 使用 FineWeb，使得 MMLU 表現高於其他中文CP模型，能力與原版 LLaMA 3 持平。\n\n## Benchmarks\n| Models                       |     | ↑ TMMLU+ (ACC) | CMMLU (ACC)   | MMLU (ACC)    |\n| ---------------------------- | --- | -------------- | ------------- | ------------- |\n|                              |     | TC, Knowledge  | CN, Knowledge | EN, Knowledge |\n|                              |     | 5 shot         | 5 shot        | 5 shot        |\n| Yi-6B                        | 6B  | 49.63          | 75.53         | 65.35         |\n| Qwen-7B                      | 7B  | 42.84          | 73.1          | 61.00         |\n| Meta-Llama-3-8B              | 8B  | 41.97          | 50.8          | 65.17         |\n| **p208p2002/llama-3-zhtw-8B** | 8B  | 41.84          | 50.6          | 65.31         |\n| Breeze-7B-Base-v0_1          | 7B  | 40.35          | 44.05         | 61.63         |\n| hfl/llama-3-chinese-8b       | 8B  | 39.64          | 50.9          | 61.1          |\n\n## Recipe\n\n### Datasets\n| Dataset        | Lang        | Weight |\n|----------------|-------------|--------|\n| FineWeb        | en          | 0.35   |\n| Wudao          | zh-cn       | 0.1    |\n| C4Tw           | zh-tw       | 0.1    |\n| WikiZhTw       | zh-tw       | 0.15   |\n| NdltdT10       | zh-tw       | 0.1    |\n| GitHubMarkDown | code        | 0.1    |\n| GitHubPython   | code        | 0.1    |\n\n### Hyper Parameters\n\n- Learning Rate: 1e-7\n- Global Batch Size: 60\n- Sequence Length: 8192\n\n",
    "related_quantizations": []
  },
  "tags": [
    "gguf",
    "endpoints_compatible",
    "region:us",
    "conversational"
  ],
  "likes": 0,
  "downloads": 266,
  "gated": false,
  "private": false,
  "last_modified": "2024-10-08T02:49:22.000Z",
  "created_at": "2024-10-07T23:42:49.000Z",
  "pipeline_tag": "",
  "library_name": ""
}

{
  "_id": "670471f9eb57ed8f3871974d",
  "id": "RichardErkhov/p208p2002_-_llama-3-zhtw-8B-gguf",
  "modelId": "RichardErkhov/p208p2002_-_llama-3-zhtw-8B-gguf",
  "sha": "b82c28d0bbc6ce831cf61173848cdb3bc2015378",
  "createdAt": "2024-10-07T23:42:49.000Z",
  "lastModified": "2024-10-08T02:49:22.000Z",
  "author": "RichardErkhov",
  "downloads": 266,
  "likes": 0,
  "gated": false,
  "private": false,
  "pipeline_tag": "",
  "library_name": "",
  "siblings_count": 24
}