richarderkhov/aisingapore_-_sea-lion-7b-gguf overview

{
  "metadata": {},
  "card_data": {
    "frontmatter": {},
    "hero_image_url": "",
    "summary": "SEA-LION is a collection of Large Language Models (LLMs) which has been pretrained and instruct-tuned for the Southeast Asia (SEA) region. The size of the models range from 3 billion to 7 billion parameters. This is the card for the SEA-LION 7B base model. SEA-LION stands for Southeast Asian Languages In One Network.",
    "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\nsea-lion-7b - GGUF\n- Model creator: https://huggingface.co/aisingapore/\n- Original model: https://huggingface.co/aisingapore/sea-lion-7b/\n\n\n| Name | Quant method | Size |\n| ---- | ---- | ---- |\n| [sea-lion-7b.Q2_K.gguf](https://huggingface.co/RichardErkhov/aisingapore_-_sea-lion-7b-gguf/blob/main/sea-lion-7b.Q2_K.gguf) | Q2_K | 3.07GB |\n| [sea-lion-7b.IQ3_XS.gguf](https://huggingface.co/RichardErkhov/aisingapore_-_sea-lion-7b-gguf/blob/main/sea-lion-7b.IQ3_XS.gguf) | IQ3_XS | 3.35GB |\n| [sea-lion-7b.IQ3_S.gguf](https://huggingface.co/RichardErkhov/aisingapore_-_sea-lion-7b-gguf/blob/main/sea-lion-7b.IQ3_S.gguf) | IQ3_S | 3.42GB |\n| [sea-lion-7b.Q3_K_S.gguf](https://huggingface.co/RichardErkhov/aisingapore_-_sea-lion-7b-gguf/blob/main/sea-lion-7b.Q3_K_S.gguf) | Q3_K_S | 3.42GB |\n| [sea-lion-7b.IQ3_M.gguf](https://huggingface.co/RichardErkhov/aisingapore_-_sea-lion-7b-gguf/blob/main/sea-lion-7b.IQ3_M.gguf) | IQ3_M | 3.72GB |\n| [sea-lion-7b.Q3_K.gguf](https://huggingface.co/RichardErkhov/aisingapore_-_sea-lion-7b-gguf/blob/main/sea-lion-7b.Q3_K.gguf) | Q3_K | 3.97GB |\n| [sea-lion-7b.Q3_K_M.gguf](https://huggingface.co/RichardErkhov/aisingapore_-_sea-lion-7b-gguf/blob/main/sea-lion-7b.Q3_K_M.gguf) | Q3_K_M | 3.97GB |\n| [sea-lion-7b.Q3_K_L.gguf](https://huggingface.co/RichardErkhov/aisingapore_-_sea-lion-7b-gguf/blob/main/sea-lion-7b.Q3_K_L.gguf) | Q3_K_L | 4.27GB |\n| [sea-lion-7b.IQ4_XS.gguf](https://huggingface.co/RichardErkhov/aisingapore_-_sea-lion-7b-gguf/blob/main/sea-lion-7b.IQ4_XS.gguf) | IQ4_XS | 4.07GB |\n| [sea-lion-7b.Q4_0.gguf](https://huggingface.co/RichardErkhov/aisingapore_-_sea-lion-7b-gguf/blob/main/sea-lion-7b.Q4_0.gguf) | Q4_0 | 4.22GB |\n| [sea-lion-7b.IQ4_NL.gguf](https://huggingface.co/RichardErkhov/aisingapore_-_sea-lion-7b-gguf/blob/main/sea-lion-7b.IQ4_NL.gguf) | IQ4_NL | 4.25GB |\n| [sea-lion-7b.Q4_K_S.gguf](https://huggingface.co/RichardErkhov/aisingapore_-_sea-lion-7b-gguf/blob/main/sea-lion-7b.Q4_K_S.gguf) | Q4_K_S | 4.25GB |\n| [sea-lion-7b.Q4_K.gguf](https://huggingface.co/RichardErkhov/aisingapore_-_sea-lion-7b-gguf/blob/main/sea-lion-7b.Q4_K.gguf) | Q4_K | 4.67GB |\n| [sea-lion-7b.Q4_K_M.gguf](https://huggingface.co/RichardErkhov/aisingapore_-_sea-lion-7b-gguf/blob/main/sea-lion-7b.Q4_K_M.gguf) | Q4_K_M | 4.67GB |\n| [sea-lion-7b.Q4_1.gguf](https://huggingface.co/RichardErkhov/aisingapore_-_sea-lion-7b-gguf/blob/main/sea-lion-7b.Q4_1.gguf) | Q4_1 | 4.6GB |\n| [sea-lion-7b.Q5_0.gguf](https://huggingface.co/RichardErkhov/aisingapore_-_sea-lion-7b-gguf/blob/main/sea-lion-7b.Q5_0.gguf) | Q5_0 | 4.97GB |\n| [sea-lion-7b.Q5_K_S.gguf](https://huggingface.co/RichardErkhov/aisingapore_-_sea-lion-7b-gguf/blob/main/sea-lion-7b.Q5_K_S.gguf) | Q5_K_S | 4.97GB |\n| [sea-lion-7b.Q5_K.gguf](https://huggingface.co/RichardErkhov/aisingapore_-_sea-lion-7b-gguf/blob/main/sea-lion-7b.Q5_K.gguf) | Q5_K | 5.3GB |\n| [sea-lion-7b.Q5_K_M.gguf](https://huggingface.co/RichardErkhov/aisingapore_-_sea-lion-7b-gguf/blob/main/sea-lion-7b.Q5_K_M.gguf) | Q5_K_M | 5.3GB |\n| [sea-lion-7b.Q5_1.gguf](https://huggingface.co/RichardErkhov/aisingapore_-_sea-lion-7b-gguf/blob/main/sea-lion-7b.Q5_1.gguf) | Q5_1 | 5.35GB |\n| [sea-lion-7b.Q6_K.gguf](https://huggingface.co/RichardErkhov/aisingapore_-_sea-lion-7b-gguf/blob/main/sea-lion-7b.Q6_K.gguf) | Q6_K | 5.77GB |\n| [sea-lion-7b.Q8_0.gguf](https://huggingface.co/RichardErkhov/aisingapore_-_sea-lion-7b-gguf/blob/main/sea-lion-7b.Q8_0.gguf) | Q8_0 | 7.46GB |\n\n\n\n\nOriginal model description:\n---\nlicense: mit\nlanguage:\n- en\n- zh\n- id\n- ms\n- th\n- vi\n- fil\n- ta\n- my\n- km\n- lo\n---\n# SEA-LION\n\nSEA-LION is a collection of Large Language Models (LLMs) which has been pretrained and instruct-tuned for the Southeast Asia (SEA) region.\nThe size of the models range from 3 billion to 7 billion parameters.\nThis is the card for the SEA-LION 7B base model.\n\nSEA-LION stands for <i>Southeast Asian Languages In One Network</i>.\n\n\n## Model Details\n\n### Model Description\n\nThe SEA-LION model is a significant leap forward in the field of Natural Language Processing,\nspecifically trained to understand the SEA regional context.\n\nSEA-LION is built on the robust MPT architecture and has a vocabulary size of 256K.\n\nFor tokenization, the model employs our custom SEABPETokenizer, which is specially tailored for SEA languages, ensuring optimal model performance.\n\nThe training data for SEA-LION encompasses 980B tokens.\n\n- **Developed by:** Products Pillar, AI Singapore\n- **Funded by:** Singapore NRF\n- **Model type:** Decoder\n- **Languages:** English, Chinese, Indonesian, Malay, Thai, Vietnamese, Filipino, Tamil, Burmese, Khmer, Lao\n- **License:** MIT License\n\n### Performance Benchmarks\n\nSEA-LION has an average performance on general tasks in English (as measured by Hugging Face's LLM Leaderboard):\n\n| Model       | ARC   | HellaSwag | MMLU  | TruthfulQA | Average |\n|-------------|:-----:|:---------:|:-----:|:----------:|:-------:|\n| SEA-LION 7B | 39.93 | 68.51     | 26.87 |      35.09 | 42.60   |\n\n## Training Details\n\n### Data\n\nSEA-LION was trained on 980B tokens of the following data:\n\n| Data Source               | Unique Tokens | Multiplier | Total Tokens | Percentage |\n|---------------------------|:-------------:|:----------:|:------------:|:----------:|\n| RefinedWeb - English      |        571.3B |          1 |       571.3B |     58.20% |\n| mC4 - Chinese             |         91.2B |          1 |        91.2B |      9.29% |\n| mC4 - Indonesian          |         3.68B |          4 |        14.7B |      1.50% |\n| mC4 - Malay               |         0.72B |          4 |         2.9B |      0.29% |\n| mC4 - Filipino            |         1.32B |          4 |         5.3B |      0.54% |\n| mC4 - Burmese             |          1.2B |          4 |         4.9B |      0.49% |\n| mC4 - Vietnamese          |         63.4B |          1 |        63.4B |      6.46% |\n| mC4 - Thai                |          5.8B |          2 |        11.6B |      1.18% |\n| WangChanBERTa - Thai      |            5B |          2 |          10B |      1.02% |\n| mC4 - Lao                 |         0.27B |          4 |         1.1B |      0.12% |\n| mC4 - Khmer               |         0.97B |          4 |         3.9B |      0.40% |\n| mC4 - Tamil               |         2.55B |          4 |        10.2B |      1.04% |\n| the Stack - Python        |         20.9B |          2 |        41.8B |      4.26% |\n| the Stack - Javascript    |         55.6B |          1 |        55.6B |      5.66% |\n| the Stack - Shell         |         1.2B5 |          2 |         2.5B |      0.26% |\n| the Stack - SQL           |         6.4B  |          2 |        12.8B |      1.31% |\n| the Stack - Markdown      |         26.6B |          1 |        26.6B |      2.71% |\n| RedPajama - StackExchange |         21.2B |          1 |        21.2B |      2.16% |\n| RedPajama - ArXiv         |         30.6B |          1 |        30.6B |      3.12% |\n\n### Infrastructure\n\nSEA-LION was trained using [MosaicML Composer](https://github.com/mosaicml/composer)\non the following hardware:\n\n| Training Details     | SEA-LION 7B  |\n|----------------------|:------------:|\n| AWS EC2 p4d.24xlarge | 32 instances |\n| Nvidia A100 40GB GPU | 256          |\n| Training Duration    | 22 days      |\n\n\n### Configuration\n\n| HyperParameter    | SEA-LION 7B        |\n|-------------------|:------------------:|\n| Precision         | bfloat16           |\n| Optimizer         | decoupled_adamw    |\n| Scheduler         | cosine_with_warmup |\n| Learning Rate     | 6.0e-5             |\n| Global Batch Size | 2048               |\n| Micro Batch Size  | 4                  |\n\n\n## Technical Specifications\n\n### Model Architecture and Objective\n\nSEA-LION is a decoder model using the MPT architecture.\n\n| Parameter       | SEA-LION 7B |\n|-----------------|:-----------:|\n| Layers          | 32          |\n| d_model         | 4096        |\n| head_dim        | 32          |\n| Vocabulary      | 256000      |\n| Sequence Length | 2048        |\n\n\n### Tokenizer Details\n\nWe sample 20M lines from the training data to train the tokenizer.<br>\nThe framework for training is [SentencePiece](https://github.com/google/sentencepiece).<br>\nThe tokenizer type is Byte-Pair Encoding (BPE).\n\n\n\n## The Team\n\nLam Wen Zhi Clarence<br>\nLeong Wei Qi<br>\nLi Yier<br>\nLiu Bing Jie Darius<br>\nLovenia Holy<br>\nMontalan Jann Railey<br>\nNg Boon Cheong Raymond<br>\nNgui Jian Gang<br>\nNguyen Thanh Ngan<br>\nOng Tat-Wee David<br>\nRengarajan Hamsawardhini<br>\nSusanto Yosephine<br>\nTai Ngee Chia<br>\nTan Choon Meng<br>\nTeo Jin Howe<br>\nTeo Eng Sipp Leslie<br>\nTeo Wei Yi<br>\nTjhi William<br>\nYeo Yeow Tong<br>\nYong Xianbin<br>\n\n## Acknowledgements\n\nAI Singapore is a national programme supported by the National Research Foundation, Singapore and hosted by the National University of Singapore.\nAny opinions, findings and conclusions or recommendations expressed in this material are those of the author(s) and do not reflect the views of National Research Foundation, Singapore.\n\n## Contact\n\nFor more info, please contact us using this [SEA-LION Inquiry Form](https://forms.gle/sLCUVb95wmGf43hi6)\n\n[Link to SEA-LION's GitHub repository](https://github.com/aisingapore/sealion)\n\n\n## Disclaimer\n\nThis the repository for the base model.\nThe model has _not_ been aligned for safety.\nDevelopers and users should perform their own safety fine-tuning and related security measures.\nIn no event shall the authors be held liable for any claim, damages, or other liability\narising from the use of the released weights and codes.\n\n\n## References\n\n```bibtex\n@misc{lowphansirikul2021wangchanberta,\n    title={WangchanBERTa: Pretraining transformer-based Thai Language Models},\n    author={Lalita Lowphansirikul and Charin Polpanumas and Nawat Jantrakulchai and Sarana Nutanong},\n    year={2021},\n    eprint={2101.09635},\n    archivePrefix={arXiv},\n    primaryClass={cs.CL}\n}\n```\n\n",
    "related_quantizations": []
  },
  "tags": [
    "gguf",
    "arxiv:2101.09635",
    "endpoints_compatible",
    "region:us"
  ],
  "likes": 0,
  "downloads": 193,
  "gated": false,
  "private": false,
  "last_modified": "2024-05-31T09:07:58.000Z",
  "created_at": "2024-05-31T06:05:08.000Z",
  "pipeline_tag": "",
  "library_name": ""
}

{
  "_id": "665968945776d589f29b8870",
  "id": "RichardErkhov/aisingapore_-_sea-lion-7b-gguf",
  "modelId": "RichardErkhov/aisingapore_-_sea-lion-7b-gguf",
  "sha": "01297b18374145b5e21ac255ccb7b261e62b3a26",
  "createdAt": "2024-05-31T06:05:08.000Z",
  "lastModified": "2024-05-31T09:07:58.000Z",
  "author": "RichardErkhov",
  "downloads": 193,
  "likes": 0,
  "gated": false,
  "private": false,
  "pipeline_tag": "",
  "library_name": "",
  "siblings_count": 24
}