unsloth/gemma-3n-e4b-it-litert-preview-gguf overview

{
  "metadata": {},
  "card_data": {
    "base_model": "google/gemma-3n-E4B-it-litert-preview",
    "language": [
      "en"
    ],
    "pipeline_tag": "image-text-to-text",
    "library_name": "transformers",
    "license": "gemma",
    "tags": [
      "gemma3",
      "unsloth",
      "transformers",
      "gemma",
      "google"
    ],
    "frontmatter": {
      "base_model": "google/gemma-3n-E4B-it-litert-preview",
      "language": [
        "en"
      ],
      "pipeline_tag": "image-text-to-text",
      "library_name": "transformers",
      "license": "gemma",
      "tags": [
        "gemma3",
        "unsloth",
        "transformers",
        "gemma",
        "google"
      ]
    },
    "hero_image_url": "https://github.com/unslothai/unsloth/raw/main/images/unsloth%20new%20logo.png",
    "summary": "| Unsloth supports          |    Free Notebooks                                                                                           | Performance | Memory use | |-----------------|--------------------------------------------------------------------------------------------------------------------------|-------------|----------| | **Gemma-3n-E4B**      | ▶️ Start on Colab               | 2x faster | 60% less | | **GRPO with Gemma 3 (1B)**      | ▶️ Start on Colab-GRPO.ipynb)               | 2x faster | 80% less | | **Gemma 3 (4B) Vision**      | ▶️ Start on Colab-Vision.ipynb)               | 2x faster | 60% less | | **Qwen3 (14B)**      | ▶️ Start on Colab-Reasoning-Conversational.ipynb)               | 2x faster | 60% less | | **DeepSeek-R1-0528-Qwen3-8B (14B)**      | ▶️ Start on Colab_GRPO.ipynb)               | 2x faster | 80% less | | **Llama-3.2 (3B)**      | ▶️ Start on Colab-Conversational.ipynb)               | 2.4x faster | 58% less |  # Gemma-3n-E4B model card **Model Page**: Gemma 3n **Resources and Technical Documentation**: **Terms of Use**: Terms\\ **Authors**: Google DeepMind",
    "quick_links": [],
    "benchmark_table_html": "",
    "readme_markdown": "---\nbase_model: google/gemma-3n-E4B-it-litert-preview\nlanguage:\n- en\npipeline_tag: image-text-to-text\nlibrary_name: transformers\nlicense: gemma\ntags:\n- gemma3\n- unsloth\n- transformers\n- gemma\n- google\n---\n> [!NOTE]  \n>  Updated from google/gemma-3n-E4B-it\n>\n<div>\n  <p style=\"margin-bottom: 0; margin-top: 0;\">\n    <strong>Learn how to run & fine-tune Gemma 3n correctly - <a href=\"https://docs.unsloth.ai/basics/gemma-3n\">Read our Guide</a>.</strong>\n  </p>\n  <p style=\"margin-bottom: 0;\">\n    <em>See <a href=\"https://huggingface.co/collections/unsloth/gemma-3n-685d3874830e49e1c93f9339\">our collection</a> for all versions of Gemma 3n including GGUF, 4-bit & 16-bit formats.</em>\n  </p>\n<p style=\"margin-top: 0;margin-bottom: 0;\">\n    <em><a href=\"https://docs.unsloth.ai/basics/unsloth-dynamic-v2.0-gguf\">Unsloth Dynamic 2.0</a> achieves SOTA accuracy & performance versus other quants.</em>\n  </p>\n  <div style=\"display: flex; gap: 5px; align-items: center; \">\n    <a href=\"https://github.com/unslothai/unsloth/\">\n      <img src=\"https://github.com/unslothai/unsloth/raw/main/images/unsloth%20new%20logo.png\" width=\"133\">\n    </a>\n    <a href=\"https://discord.gg/unsloth\">\n      <img src=\"https://github.com/unslothai/unsloth/raw/main/images/Discord%20button.png\" width=\"173\">\n    </a>\n    <a href=\"https://docs.unsloth.ai/basics/gemma-3n\">\n      <img src=\"https://raw.githubusercontent.com/unslothai/unsloth/refs/heads/main/images/documentation%20green%20button.png\" width=\"143\">\n    </a>\n  </div>\n<h1 style=\"margin-top:0; margin-bottom: 0;\">✨ Gemma 3n Usage Guidelines</h1>\n</div>\n\n- Currently **only text** is supported.\n- Ollama: `ollama run hf.co/unsloth/gemma-3n-E4B-it-GGUF:Q4_K_XL` - auto-sets correct chat template and settings\n- Set temperature = 1.0, top_k = 64, top_p = 0.95, min_p = 0.0\n- Gemma 3n max tokens (context length): 32K. Gemma 3n chat template:\n```\n<bos><start_of_turn>user\\nHello!<end_of_turn>\\n<start_of_turn>model\\nHey there!<end_of_turn>\\n<start_of_turn>user\\nWhat is 1+1?<end_of_turn>\\n<start_of_turn>model\\n\n```\n- For complete detailed instructions, see our [step-by-step guide](https://docs.unsloth.ai/basics/gemma-3n).\n\n# 🦥 Fine-tune Gemma 3n with Unsloth\n\n- Fine-tune Gemma 3n (4B) for free using our Google [Colab notebook here](https://docs.unsloth.ai/get-started/unsloth-notebooks)!\n- Read our Blog about Gemma 3n support: [unsloth.ai/blog/gemma-3n](https://unsloth.ai/blog/gemma-3n)\n- View the rest of our notebooks in our [docs here](https://docs.unsloth.ai/get-started/unsloth-notebooks).\n\n| Unsloth supports          |    Free Notebooks                                                                                           | Performance | Memory use |\n|-----------------|--------------------------------------------------------------------------------------------------------------------------|-------------|----------|\n| **Gemma-3n-E4B**      | [▶️ Start on Colab](https://docs.unsloth.ai/get-started/unsloth-notebooks)               | 2x faster | 60% less |\n| **GRPO with Gemma 3 (1B)**      | [▶️ Start on Colab](https://colab.research.google.com/github/unslothai/notebooks/blob/main/nb/Gemma3_(1B)-GRPO.ipynb)               | 2x faster | 80% less |\n| **Gemma 3 (4B) Vision**      | [▶️ Start on Colab](https://colab.research.google.com/github/unslothai/notebooks/blob/main/nb/Gemma3_(4B)-Vision.ipynb)               | 2x faster | 60% less |\n| **Qwen3 (14B)**      | [▶️ Start on Colab](https://colab.research.google.com/github/unslothai/notebooks/blob/main/nb/Qwen3_(14B)-Reasoning-Conversational.ipynb)               | 2x faster | 60% less |\n| **DeepSeek-R1-0528-Qwen3-8B (14B)**      | [▶️ Start on Colab](https://colab.research.google.com/github/unslothai/notebooks/blob/main/nb/DeepSeek_R1_0528_Qwen3_(8B)_GRPO.ipynb)               | 2x faster | 80% less |\n| **Llama-3.2 (3B)**      | [▶️ Start on Colab](https://colab.research.google.com/github/unslothai/notebooks/blob/main/nb/Llama3.2_(1B_and_3B)-Conversational.ipynb)               | 2.4x faster | 58% less |\n<br>\n\n# Gemma-3n-E4B model card\n**Model Page**: [Gemma 3n](https://ai.google.dev/gemma/docs/gemma-3n)\n\n**Resources and Technical Documentation**:\n\n-   [Responsible Generative AI Toolkit](https://ai.google.dev/responsible)\n-   [Gemma on Kaggle](https://www.kaggle.com/models/google/gemma-3n)\n-   [Gemma on HuggingFace](https://huggingface.co/collections/google/gemma-3n-685065323f5984ef315c93f4)\n-   [Gemma on Vertex Model Garden](https://console.cloud.google.com/vertex-ai/publishers/google/model-garden/gemma3n)\n\n**Terms of Use**: [Terms](https://ai.google.dev/gemma/terms)\\\n**Authors**: Google DeepMind\n\n## Model Information\n\nSummary description and brief definition of inputs and outputs.\n\n### Description\n\nGemma is a family of lightweight, state-of-the-art open models from Google,\nbuilt from the same research and technology used to create the Gemini models.\nGemma 3n models are designed for efficient execution on low-resource devices.\nThey are capable of multimodal input, handling text, image, video, and audio\ninput, and generating text outputs, with open weights for pre-trained and\ninstruction-tuned variants. These models were trained with data in over 140\nspoken languages.\n\nGemma 3n models use selective parameter activation technology to reduce resource\nrequirements. This technique allows the models to operate at an effective size\nof 2B and 4B parameters, which is lower than the total number of parameters they\ncontain. For more information on Gemma 3n's efficient parameter management\ntechnology, see the\n[Gemma 3n](https://ai.google.dev/gemma/docs/gemma-3n#parameters)\npage.\n\n### Inputs and outputs\n\n-   **Input:**\n    -   Text string, such as a question, a prompt, or a document to be\n        summarized\n    -   Images, normalized to 256x256, 512x512, or 768x768 resolution\n        and encoded to 256 tokens each\n    -   Audio data encoded to 6.25 tokens per second from a single channel\n    -   Total input context of 32K tokens\n-   **Output:**\n    -   Generated text in response to the input, such as an answer to a\n        question, analysis of image content, or a summary of a document\n    -   Total output length up to 32K tokens, subtracting the request\n        input tokens\n### Usage\n\nBelow, there are some code snippets on how to get quickly started with running\nthe model. First, install the Transformers library. Gemma 3n is supported\nstarting from transformers 4.53.0.\n\n```sh\n$ pip install -U transformers\n```\n\nThen, copy the snippet from the section that is relevant for your use case.\n\n#### Running with the `pipeline` API\n\nYou can initialize the model and processor for inference with `pipeline` as\nfollows.\n\n```python\nfrom transformers import pipeline\nimport torch\npipe = pipeline(\n    \"image-text-to-text\",\n    model=\"google/gemma-3n-e4b-it\",\n    device=\"cuda\",\n    torch_dtype=torch.bfloat16,\n)\n```\n\nWith instruction-tuned models, you need to use chat templates to process our\ninputs first. Then, you can pass it to the pipeline.\n\n```python\nmessages = [\n    {\n        \"role\": \"system\",\n        \"content\": [{\"type\": \"text\", \"text\": \"You are a helpful assistant.\"}]\n    },\n    {\n        \"role\": \"user\",\n        \"content\": [\n            {\"type\": \"image\", \"url\": \"https://huggingface.co/datasets/huggingface/documentation-images/resolve/main/p-blog/candy.JPG\"},\n            {\"type\": \"text\", \"text\": \"What animal is on the candy?\"}\n        ]\n    }\n]\noutput = pipe(text=messages, max_new_tokens=200)\nprint(output[0][\"generated_text\"][-1][\"content\"])\n# Okay, let's take a look!\n# Based on the image, the animal on the candy is a **turtle**.\n# You can see the shell shape and the head and legs.\n```\n\n#### Running the model on a single GPU\n\n```python\nfrom transformers import AutoProcessor, Gemma3nForConditionalGeneration\nfrom PIL import Image\nimport requests\nimport torch\nmodel_id = \"google/gemma-3n-e4b-it\"\nmodel = Gemma3nForConditionalGeneration.from_pretrained(model_id, device_map=\"auto\", torch_dtype=torch.bfloat16,).eval()\nprocessor = AutoProcessor.from_pretrained(model_id)\nmessages = [\n    {\n        \"role\": \"system\",\n        \"content\": [{\"type\": \"text\", \"text\": \"You are a helpful assistant.\"}]\n    },\n    {\n        \"role\": \"user\",\n        \"content\": [\n            {\"type\": \"image\", \"image\": \"https://huggingface.co/datasets/huggingface/documentation-images/resolve/main/bee.jpg\"},\n            {\"type\": \"text\", \"text\": \"Describe this image in detail.\"}\n        ]\n    }\n]\ninputs = processor.apply_chat_template(\n    messages,\n    add_generation_prompt=True,\n    tokenize=True,\n    return_dict=True,\n    return_tensors=\"pt\",\n).to(model.device)\ninput_len = inputs[\"input_ids\"].shape[-1]\nwith torch.inference_mode():\n    generation = model.generate(**inputs, max_new_tokens=100, do_sample=False)\n    generation = generation[0][input_len:]\ndecoded = processor.decode(generation, skip_special_tokens=True)\nprint(decoded)\n# **Overall Impression:** The image is a close-up shot of a vibrant garden scene,\n# focusing on a cluster of pink cosmos flowers and a busy bumblebee.\n# It has a slightly soft, natural feel, likely captured in daylight.\n```\n\n### Citation\n\n```\n@article{gemma_3n_2025,\n    title={Gemma 3n},\n    url={https://ai.google.dev/gemma/docs/gemma-3n},\n    publisher={Google DeepMind},\n    author={Gemma Team},\n    year={2025}\n}\n```\n\n## Model Data\n\nData used for model training and how the data was processed.\n\n### Training Dataset\n\nThese models were trained on a dataset that includes a wide variety of sources\ntotalling approximately 11 trillion tokens. The knowledge cutoff date for the\ntraining data was June 2024. Here are the key components:\n\n-   **Web Documents**: A diverse collection of web text ensures the model\n    is exposed to a broad range of linguistic styles, topics, and vocabulary.\n    The training dataset includes content in over 140 languages.\n-   **Code**: Exposing the model to code helps it to learn the syntax and\n    patterns of programming languages, which improves its ability to generate\n    code and understand code-related questions.\n-   **Mathematics**: Training on mathematical text helps the model learn\n    logical reasoning, symbolic representation, and to address mathematical queries.\n-   **Images**: A wide range of images enables the model to perform image\n    analysis and visual data extraction tasks.\n-   Audio: A diverse set of sound samples enables the model to recognize\n    speech, transcribe text from recordings, and identify information in audio data.\nThe combination of these diverse data sources is crucial for training a\npowerful multimodal model that can handle a wide variety of different tasks and\ndata formats.\n\n### Data Preprocessing\n\nHere are the key data cleaning and filtering methods applied to the training\ndata:\n\n-   **CSAM Filtering**: Rigorous CSAM (Child Sexual Abuse Material)\n    filtering was applied at multiple stages in the data preparation process to\n    ensure the exclusion of harmful and illegal content.\n-   **Sensitive Data Filtering**: As part of making Gemma pre-trained models\n    safe and reliable, automated techniques were used to filter out certain\n    personal information and other sensitive data from training sets.\n-   **Additional methods**: Filtering based on content quality and safety in\n    line with\n    [our policies](https://ai.google/static/documents/ai-responsibility-update-published-february-2025.pdf).\n## Implementation Information\n\nDetails about the model internals.\n\n### Hardware\n\nGemma was trained using [Tensor Processing Unit\n(TPU)](https://cloud.google.com/tpu/docs/intro-to-tpu) hardware (TPUv4p, TPUv5p\nand TPUv5e). Training generative models requires significant computational\npower. TPUs, designed specifically for matrix operations common in machine\nlearning, offer several advantages in this domain:\n\n-   **Performance**: TPUs are specifically designed to handle the massive\n    computations involved in training generative models. They can speed up\n    training considerably compared to CPUs.\n-   **Memory**: TPUs often come with large amounts of high-bandwidth memory,\n    allowing for the handling of large models and batch sizes during training.\n    This can lead to better model quality.\n-   **Scalability**: TPU Pods (large clusters of TPUs) provide a scalable\n    solution for handling the growing complexity of large foundation models.\n    You can distribute training across multiple TPU devices for faster and more\n    efficient processing.\n-   **Cost-effectiveness**: In many scenarios, TPUs can provide a more\n    cost-effective solution for training large models compared to CPU-based\n    infrastructure, especially when considering the time and resources saved\n    due to faster training.\nThese advantages are aligned with\n[Google's commitments to operate sustainably](https://sustainability.google/operating-sustainably/).\n\n### Software\n\nTraining was done using [JAX](https://github.com/jax-ml/jax) and\n[ML Pathways](https://blog.google/technology/ai/introducing-pathways-next-generation-ai-architecture/).\nJAX allows researchers to take advantage of the latest generation of hardware,\nincluding TPUs, for faster and more efficient training of large models. ML\nPathways is Google's latest effort to build artificially intelligent systems\ncapable of generalizing across multiple tasks. This is specially suitable for\nfoundation models, including large language models like these ones.\n\nTogether, JAX and ML Pathways are used as described in the\n[paper about the Gemini family of models](https://goo.gle/gemma2report):\n*\"the 'single controller' programming model of Jax and Pathways allows a single\nPython process to orchestrate the entire training run, dramatically simplifying\nthe development workflow.\"*\n\n## Evaluation\n\nModel evaluation metrics and results.\n\n### Benchmark Results\n\nThese models were evaluated at full precision (float32) against a large\ncollection of different datasets and metrics to cover different aspects of\ncontent generation. Evaluation results marked with **IT** are for\ninstruction-tuned models. Evaluation results marked with **PT** are for\npre-trained models.\n\n#### Reasoning and factuality\n\n| Benchmark                      | Metric         | n-shot   |  E2B PT  |  E4B PT  |\n| ------------------------------ |----------------|----------|:--------:|:--------:|\n| [HellaSwag][hellaswag]         | Accuracy       | 10-shot  |   72.2   |   78.6   |\n| [BoolQ][boolq]                 | Accuracy       | 0-shot   |   76.4   |   81.6   |\n| [PIQA][piqa]                   | Accuracy       | 0-shot   |   78.9   |   81.0   |\n| [SocialIQA][socialiqa]         | Accuracy       | 0-shot   |   48.8   |   50.0   |\n| [TriviaQA][triviaqa]           | Accuracy       | 5-shot   |   60.8   |   70.2   |\n| [Natural Questions][naturalq]  | Accuracy       | 5-shot   |   15.5   |   20.9   |\n| [ARC-c][arc]                   | Accuracy       | 25-shot  |   51.7   |   61.6   |\n| [ARC-e][arc]                   | Accuracy       | 0-shot   |   75.8   |   81.6   |\n| [WinoGrande][winogrande]       | Accuracy       | 5-shot   |   66.8   |   71.7   |\n| [BIG-Bench Hard][bbh]          | Accuracy       | few-shot |   44.3   |   52.9   |\n| [DROP][drop]                   | Token F1 score | 1-shot   |   53.9   |   60.8   |\n\n[hellaswag]: https://arxiv.org/abs/1905.07830\n[boolq]: https://arxiv.org/abs/1905.10044\n[piqa]: https://arxiv.org/abs/1911.11641\n[socialiqa]: https://arxiv.org/abs/1904.09728\n[triviaqa]: https://arxiv.org/abs/1705.03551\n[naturalq]: https://github.com/google-research-datasets/natural-questions\n[arc]: https://arxiv.org/abs/1911.01547\n[winogrande]: https://arxiv.org/abs/1907.10641\n[bbh]: https://paperswithcode.com/dataset/bbh\n[drop]: https://arxiv.org/abs/1903.00161\n\n#### Multilingual\n\n| Benchmark                           | Metric                  | n-shot   |  E2B IT  |  E4B IT  |\n| ------------------------------------|-------------------------|----------|:--------:|:--------:|\n| [MGSM][mgsm]                        | Accuracy                |  0-shot  |   53.1   |   60.7   |\n| [WMT24++][wmt24pp] (ChrF)           | Character-level F-score |  0-shot  |   42.7   |   50.1   |\n| [Include][include]                  | Accuracy                |  0-shot  |   38.6   |   57.2   |\n| [MMLU][mmlu] (ProX)                 | Accuracy                |  0-shot  |    8.1   |   19.9   |\n| [OpenAI MMLU][openai-mmlu]          | Accuracy                |  0-shot  |   22.3   |   35.6   |\n| [Global-MMLU][global-mmlu]          | Accuracy                |  0-shot  |   55.1   |   60.3   |\n| [ECLeKTic][eclektic]                | ECLeKTic score          |  0-shot  |    2.5   |    1.9   |\n\n[mgsm]: https://arxiv.org/abs/2210.03057\n[wmt24pp]: https://arxiv.org/abs/2502.12404v1\n[include]:https://arxiv.org/abs/2411.19799\n[mmlu]: https://arxiv.org/abs/2009.03300\n[openai-mmlu]: https://huggingface.co/datasets/openai/MMMLU\n[global-mmlu]: https://huggingface.co/datasets/CohereLabs/Global-MMLU\n[eclektic]: https://arxiv.org/abs/2502.21228\n\n#### STEM and code\n\n| Benchmark                           | Metric                   | n-shot   |  E2B IT  |  E4B IT  |\n| ------------------------------------|--------------------------|----------|:--------:|:--------:|\n| [GPQA][gpqa] Diamond                | RelaxedAccuracy/accuracy |  0-shot  |   24.8   |   23.7   |\n| [LiveCodeBench][lcb] v5             | pass@1                   |  0-shot  |   18.6   |   25.7   |\n| Codegolf v2.2                       | pass@1                   |  0-shot  |   11.0   |   16.8   |\n| [AIME 2025][aime-2025]              | Accuracy                 |  0-shot  |    6.7   |   11.6   |\n\n[gpqa]: https://arxiv.org/abs/2311.12022\n[lcb]: https://arxiv.org/abs/2403.07974\n[aime-2025]: https://www.vals.ai/benchmarks/aime-2025-05-09\n\n#### Additional benchmarks\n\n| Benchmark                            | Metric     | n-shot   |  E2B IT  |  E4B IT  |\n| ------------------------------------ |------------|----------|:--------:|:--------:|\n| [MMLU][mmlu]                         |  Accuracy  |  0-shot  |   60.1   |   64.9   |\n| [MBPP][mbpp]                         |  pass@1    |  3-shot  |   56.6   |   63.6   |\n| [HumanEval][humaneval]               |  pass@1    |  0-shot  |   66.5   |   75.0   |\n| [LiveCodeBench][lcb]                 |  pass@1    |  0-shot  |   13.2   |   13.2   |\n| HiddenMath                           |  Accuracy  |  0-shot  |   27.7   |   37.7   |\n| [Global-MMLU-Lite][global-mmlu-lite] |  Accuracy  |  0-shot  |   59.0   |   64.5   |\n| [MMLU][mmlu] (Pro)                   |  Accuracy  |  0-shot  |   40.5   |   50.6   |\n\n[gpqa]: https://arxiv.org/abs/2311.12022\n[mbpp]: https://arxiv.org/abs/2108.07732\n[humaneval]: https://arxiv.org/abs/2107.03374\n[lcb]: https://arxiv.org/abs/2403.07974\n[global-mmlu-lite]: https://huggingface.co/datasets/CohereForAI/Global-MMLU-Lite\n\n## Ethics and Safety\n\nEthics and safety evaluation approach and results.\n\n### Evaluation Approach\n\nOur evaluation methods include structured evaluations and internal red-teaming\ntesting of relevant content policies. Red-teaming was conducted by a number of\ndifferent teams, each with different goals and human evaluation metrics. These\nmodels were evaluated against a number of different categories relevant to\nethics and safety, including:\n\n-   **Child Safety**: Evaluation of text-to-text and image to text prompts\n    covering child safety policies, including child sexual abuse and\n    exploitation.\n-   **Content Safety:** Evaluation of text-to-text and image to text prompts\n    covering safety policies including, harassment, violence and gore, and hate\n    speech.\n-   **Representational Harms**: Evaluation of text-to-text and image to text\n    prompts covering safety policies including bias, stereotyping, and harmful\n    associations or inaccuracies.\nIn addition to development level evaluations, we conduct \"assurance\nevaluations\" which are our 'arms-length' internal evaluations for responsibility\ngovernance decision making. They are conducted separately from the model\ndevelopment team, to inform decision making about release. High level findings\nare fed back to the model team, but prompt sets are held-out to prevent\noverfitting and preserve the results' ability to inform decision making. Notable\nassurance evaluation results are reported to our Responsibility & Safety Council\nas part of release review.\n\n### Evaluation Results\n\nFor all areas of safety testing, we saw safe levels of performance across the\ncategories of child safety, content safety, and representational harms relative\nto previous Gemma models. All testing was conducted without safety filters to\nevaluate the model capabilities and behaviors. For text-to-text,  image-to-text,\nand audio-to-text, and across all model sizes, the model produced minimal policy\nviolations, and showed significant improvements over previous Gemma models'\nperformance with respect to high severity violations. A limitation of our\nevaluations was they included primarily English language prompts.\n\n## Usage and Limitations\n\nThese models have certain limitations that users should be aware of.\n\n### Intended Usage\n\nOpen generative models have a wide range of applications across various\nindustries and domains. The following list of potential uses is not\ncomprehensive. The purpose of this list is to provide contextual information\nabout the possible use-cases that the model creators considered as part of model\ntraining and development.\n\n-   Content Creation and Communication\n    -   **Text Generation**: Generate creative text formats such as\n        poems, scripts, code, marketing copy, and email drafts.\n    -   **Chatbots and Conversational AI**: Power conversational\n        interfaces for customer service, virtual assistants, or interactive\n        applications.\n    -   **Text Summarization**: Generate concise summaries of a text\n        corpus, research papers, or reports.\n    -   **Image Data Extraction**: Extract, interpret, and summarize\n        visual data for text communications.\n    -   **Audio Data Extraction**: Transcribe spoken language, translate speech\n        to text in other languages, and analyze sound-based data.\n-   Research and Education\n    -   **Natural Language Processing (NLP) and generative model\n        Research**: These models can serve as a foundation for researchers to\n        experiment with generative models and NLP techniques, develop\n        algorithms, and contribute to the advancement of the field.\n    -   **Language Learning Tools**: Support interactive language\n        learning experiences, aiding in grammar correction or providing writing\n        practice.\n    -   **Knowledge Exploration**: Assist researchers in exploring large\n        bodies of data by generating summaries or answering questions about\n        specific topics.\n### Limitations\n\n-   Training Data\n    -   The quality and diversity of the training data significantly\n        influence the model's capabilities. Biases or gaps in the training data\n        can lead to limitations in the model's responses.\n    -   The scope of the training dataset determines the subject areas\n        the model can handle effectively.\n-   Context and Task Complexity\n    -   Models are better at tasks that can be framed with clear\n        prompts and instructions. Open-ended or highly complex tasks might be\n        challenging.\n    -   A model's performance can be influenced by the amount of context\n        provided (longer context generally leads to better outputs, up to a\n        certain point).\n-   Language Ambiguity and Nuance\n    -   Natural language is inherently complex. Models might struggle\n        to grasp subtle nuances, sarcasm, or figurative language.\n-   Factual Accuracy\n    -   Models generate responses based on information they learned\n        from their training datasets, but they are not knowledge bases. They\n        may generate incorrect or outdated factual statements.\n-   Common Sense\n    -   Models rely on statistical patterns in language. They might\n        lack the ability to apply common sense reasoning in certain situations.\n### Ethical Considerations and Risks\n\nThe development of generative models raises several ethical concerns. In\ncreating an open model, we have carefully considered the following:\n\n-   Bias and Fairness\n    -   Generative models trained on large-scale, real-world text and image data\n        can reflect socio-cultural biases embedded in the training material.\n        These models underwent careful scrutiny, input data pre-processing\n        described and posterior evaluations reported in this card.\n-   Misinformation and Misuse\n    -   Generative models can be misused to generate text that is\n        false, misleading, or harmful.\n    -   Guidelines are provided for responsible use with the model, see the\n        [Responsible Generative AI Toolkit](https://ai.google.dev/responsible).\n-   Transparency and Accountability:\n    -   This model card summarizes details on the models' architecture,\n        capabilities, limitations, and evaluation processes.\n    -   A responsibly developed open model offers the opportunity to\n        share innovation by making generative model technology accessible to\n        developers and researchers across the AI ecosystem.\nRisks identified and mitigations:\n\n-   **Perpetuation of biases**: It's encouraged to perform continuous monitoring\n    (using evaluation metrics, human review) and the exploration of de-biasing\n    techniques during model training, fine-tuning, and other use cases.\n-   **Generation of harmful content**: Mechanisms and guidelines for content\n    safety are essential. Developers are encouraged to exercise caution and\n    implement appropriate content safety safeguards based on their specific\n    product policies and application use cases.\n-   **Misuse for malicious purposes**: Technical limitations and developer\n    and end-user education can help mitigate against malicious applications of\n    generative models. Educational resources and reporting mechanisms for users\n    to flag misuse are provided. Prohibited uses of Gemma models are outlined\n    in the\n    [Gemma Prohibited Use Policy](https://ai.google.dev/gemma/prohibited_use_policy).\n-   **Privacy violations**: Models were trained on data filtered for removal of\n    certain personal information and other sensitive data. Developers are\n    encouraged to adhere to privacy regulations with privacy-preserving\n    techniques.\n### Benefits\n\nAt the time of release, this family of models provides high-performance open\ngenerative model implementations designed from the ground up for responsible AI\ndevelopment compared to similarly sized models.\n\nUsing the benchmark evaluation metrics described in this document, these models\nhave shown to provide superior performance to other, comparably-sized open model\nalternatives.",
    "related_quantizations": []
  },
  "tags": [
    "transformers",
    "gguf",
    "gemma3",
    "unsloth",
    "gemma",
    "google",
    "image-text-to-text",
    "en",
    "arxiv:1905.07830",
    "arxiv:1905.10044",
    "arxiv:1911.11641",
    "arxiv:1904.09728",
    "arxiv:1705.03551",
    "arxiv:1911.01547",
    "arxiv:1907.10641",
    "arxiv:1903.00161",
    "arxiv:2210.03057",
    "arxiv:2502.12404",
    "arxiv:2411.19799",
    "arxiv:2009.03300",
    "arxiv:2502.21228",
    "arxiv:2311.12022",
    "arxiv:2403.07974",
    "arxiv:2108.07732",
    "arxiv:2107.03374",
    "base_model:google/gemma-3n-E4B-it-litert-preview",
    "base_model:quantized:google/gemma-3n-E4B-it-litert-preview",
    "license:gemma",
    "endpoints_compatible",
    "region:us",
    "conversational"
  ],
  "likes": 4,
  "downloads": 3607,
  "gated": false,
  "private": false,
  "last_modified": "2025-07-16T09:03:27.000Z",
  "created_at": "2025-07-08T08:33:58.000Z",
  "pipeline_tag": "image-text-to-text",
  "library_name": "transformers"
}

{
  "_id": "686cd7f6180f078067906d75",
  "id": "unsloth/gemma-3n-E4B-it-litert-preview-GGUF",
  "modelId": "unsloth/gemma-3n-E4B-it-litert-preview-GGUF",
  "sha": "60a01054165a834c9c47b79dd94de15a8f636ab6",
  "createdAt": "2025-07-08T08:33:58.000Z",
  "lastModified": "2025-07-16T09:03:27.000Z",
  "author": "unsloth",
  "downloads": 3607,
  "likes": 4,
  "gated": false,
  "private": false,
  "pipeline_tag": "image-text-to-text",
  "library_name": "transformers",
  "siblings_count": 29
}