silma-ai

SILMA.AI is a leading Generative AI startup dedicated to empowering Arabic speakers with state-of-the-art AI solutions. SILMA 1.0 was the TOP-RANKED open-weights Arabic LLM (Until February 2025) with an impressive 9 billion parameter size, surpassing models that are over seven times larger 🏆 Important Tip: 💡 For RAG use-cases please use SILMA Kashif v1.0 as it has been specifically trained for Question Answering tasks. SIMLA is a small language model outperforming 72B models in most arabic language tasks, thus more practical for business use-cases SILMA is built over the robust foundational models of Google Gemma, combining the strengths of both to provide you with unparalleled performance SILMA is an open-weight model, free to use in accordance with our open license We are a team of seasoned Arabic AI experts who understand the nuances of the language and cultural considerations, enabling us to build solutions that truly resonate with Arabic users. Below we share some code snippets on how to get quickly started with running the model. First, install the Transformers library with: Then, copy the snippet from the section that is relevant for your usecase. You can ensure the correct chat template is applied by using `tokenizer.applychattemplate` as follows: Torch compile is a method for speeding-up the inference of PyTorch modules. The Silma model can be run up to 6x faster by leveraging torch compile. Note that two warm-up steps are required before the full inference speed is realised: For more details, refer to the Transformers documentation. The instruction-tuned models use a chat template that must be adhered to for conversational use. The easiest way to apply it is using the tokenizer's built-in chat template, as shown in the following snippet. Let's load the model and apply the chat template to a conversation. In this example, we'll start with a single user interaction: At this point, the prompt contains the following text: As you can see, each turn is preceded by a ` ` delimiter and then the role of the entity (either `user`, for content supplied by the user, or `model` for LLM responses). Turns finish with the ` ` token. You can follow this format to build the prompt manually, if you need to do it without the tokenizer's chat template. After the prompt is ready, generation can be performed like this: Input: Text string, such as a question, a prompt, or a document to be summarized. Output: Generated Arabic or English text in response to the input, such as an answer to a question, or a summary of a document. The following are the minimum/recommended GPU requirements for running inference: Recommended At least one GPU with a minimum of 48 GB of GPU memory Examples: Nvidia A40, L40, RTX A6000 At least one GPU with 16-24 GB of GPU memory Examples: Nvidia RTX 4090, RTX 4000, L4 Assuming that the model is loaded in either 8-bit or 4-bit Quantization mode These models have certain limitations that users should be aware of. Open Large Language Models (LLMs) have a wide range of applications across various industries and domains. The following list of potential uses is not comprehensive. The purpose of this list is to provide contextual information about the possible use-cases that the model creators considered as part of model training and development. Content Creation and Communication Text Generation: These models can be used to generate creative text formats such as poems, scripts, code, marketing copy, and email drafts. Chatbots and Conversational AI: Power conversational interfaces for customer service, virtual assistants, or interactive applications. Text Summarization: Generate concise summaries of a text corpus, research papers, or reports. Research and Education Natural Language Processing (NLP) Research: These models can serve as a foundation for researchers to experiment with NLP techniques, develop algorithms, and contribute to the advancement of the field. Language Learning Tools: Support interactive language learning experiences, aiding in grammar correction or providing writing practice. Knowledge Exploration: Assist researchers in exploring large bodies of text by generating summaries or answering questions about specific topics. Training Data The quality and diversity of the training data significantly influence the model's capabilities. Biases or gaps in the training data can lead to limitations in the model's responses. The scope of the training dataset determines the subject areas the model can handle effectively. Context and Task Complexity LLMs are better at tasks that can be framed with clear prompts and instructions. Open-ended or highly complex tasks might be challenging. A model's performance can be influenced by the amount of context provided (longer context generally leads to better outputs, up to a certain point). Language Ambiguity and Nuance Natural language is inherently complex. LLMs might struggle to grasp subtle nuances, sarcasm, or figurative language. Factual Accuracy LLMs generate responses based on information they learned from their training datasets, but they are not knowledge bases. They may generate incorrect or outdated factual statements. Common Sense LLMs rely on statistical patterns in language. They might lack the ability to apply common sense reasoning in certain situations. The development of large language models (LLMs) raises several ethical concerns. In creating an open model, we have carefully considered the following: Bias and Fairness LLMs trained on large-scale, real-world text data can reflect socio-cultural biases embedded in the training material. Misinformation and Misuse LLMs can be misused to generate text that is false, misleading, or harmful. Guidelines are provided for responsible use with the model, see the [Responsible Generative AI Toolkit][rai-toolkit]. Transparency and Accountability: This model card summarizes details on the models' architecture, capabilities, limitations, and evaluation processes. A responsibly developed open model offers the opportunity to share innovation by making LLM technology accessible to developers and researchers across the AI ecosystem. Perpetuation of biases: It's encouraged to perform continuous monitoring (using evaluation metrics, human review) and the exploration of de-biasing techniques during model training, fine-tuning, and other use cases. Generation of harmful content: Mechanisms and guidelines for content safety are essential. Developers are encouraged to exercise caution and implement appropriate content safety safeguards based on their specific product policies and application use cases. Privacy violations: Models were trained on data filtered for removal of PII (Personally Identifiable Information). Developers are encouraged to adhere to privacy regulations with privacy-preserving techniques.

SILMA Kashif 2B Instruct v1.0 is the premier release within the SILMA Kashif Family of models, specifically designed for RAG (Retrieval-Augmented Generation) tasks Kashif excels in a specific task, answering questions based on contextual pieces in both Arabic and English. In addition, the model is also capable of performing Entity Extraction tasks as a minor skill SILMA Kashif 2B v1.0 stands out as the top-performing open model in RAG within the 3-9 billion parameter range based on our evaluations using SILMA RAGQA Benchmark SILMA Kashif is built on the powerful foundational models of Google Gemma, merging their strengths to deliver unmatched performance for users Kashif is an open-weight model, free to use in accordance with our open license Finally, the model comes with a context length of 12k Important note: Kashif is a specialized model which should ONLY be used in RAG setups. If you are looking for a general purpose model, please refer to SILMA 9B Instruct v1.0 The model underwent intensive training to master a wide range of tasks and excel in performance. - The ability to answer questions in Arabic and English - The ability to deal with short and long contexts - The ability to provide short and long answers effectively - The ability to answer complex numerical questions - The ability to answer questions based on tabular data - Answering multi-hop questions: The ability to answer a single question using pieces of data from multiple paragraphs - Negative rejection: The ability to identify and exclude inaccurate answers, and provide a more accurate statement such as "The answer cannot be found in the given context" - Multi-domains: The ability to answer questions based on texts from different fields such as finance, medical, legal, etc. - The ability to deal with ambiguous contexts - The ability to extract entities from text - Ability to deal with diverse and complex prompts |Dataset | exactmatch | rouge1 | bleu | bertscore| |---|---|---|---|---| |ragbench-finqa-en-test | 0.000 | 0.587 | 0.321 | 0.760| |ragbench-tatqa-ar-test | 0.000 | 0.484 | 0.130 | 0.774| |ragbench-tatqa-en-test | 0.059 | 0.646 | 0.423 | 0.808| |rag-instruct-benchmark-tester-en | 0.370 | 0.683 | 0.196 | 0.791| |ragbench-expertqa-en-test |0.000 | 0.465 | 0.151 | 0.677| |ragbench-msmarco-ar-test |0.000 | 0.144 | 0.096 | 0.781| |sciq-ar-test |0.170 | 0.000 | 0.000 | 0.753| |ragbench-covidqa-en-test |0.020 | 0.521 | 0.242 | 0.734| |ragbench-emanual-ar-test |0.000 | 0.237 | 0.159 | 0.806| |ragbench-finqa-ar-test |0.000 | 0.377 | 0.109 | 0.780| |xquad-r-validation-en |0.120 | 0.326 | 0.041 | 0.603| |ragbench-emanual-en-test |0.000 | 0.565 | 0.288 | 0.722| |xquad-r-ar-validation |0.070 | 0.130 | 0.042 | 0.698| |boolq-ar-test |0.450 | 0.000 | 0.000 | 0.700| |ragbench-hotpotqa-en-test |0.060 | 0.732 | 0.503 | 0.837| |ragbench-covidqa-ar-test |0.000 | 0.179 | 0.104 | 0.783| |ragbench-msmarco-en-test |0.020 | 0.491 | 0.207 | 0.729| |### Benchmark Average Scores |0.079 | 0.386 | 0.177 | 0.749| silma.ai is a leading GenAI startup that excels in building and tailoring cutting-edge Large Language Models (LLMs) and AI technologies for the Arabic language Below we share some code snippets on how to get quickly started with running the model. First, install the Transformers library with: Note: for advanced usage examples such as multi-gpu, quantization or chat templates, please refer to SILMA v1.0 examples Here is a recommended way to prompt the model. You can modify the prompt based on your specific requirements, but if you encounter any challenges, following the format below in which we used to train the model may be helpful. The following are the minimum/recommended GPU requirements for running inference: Recommended At least one GPU with a minimum of 24 GB of GPU memory Examples: Nvidia RTX 4090 At least one GPU with 8 GB of GPU memory Examples: Nvidia RTX 3070, RTX 3080 or T4 We have seen 2.6% drop in score (to 0.338) for the same model quantized 4bit The model should only be used in question answering use-cases such as RAG The model can also be used to extract entities from text Because it has few parameters, we've noticed that the model isn't very effective for handling complex numerical and financial reasoning, such as solving tricky calculations The model has been trained specifically for text-based question answering, which may limit its ability to perform tasks beyond this scope, including simple tasks