--- pipeline_tag: text-generation inference: false license: apache-2.0 # datasets: # metrics: # - code_eval library_name: transformers tags: - language - granite-3.0 model-index: - name: granite-3.0-3b-a800m-instruct results: - task: type: text-generation dataset: type: human-exams name: MMLU metrics: - name: pass@1 type: pass@1 value: veriefied: false - task: type: text-generation dataset: type: human-exams name: MMLU-Pro metrics: - name: pass@1 type: pass@1 value: veriefied: false - task: type: text-generation dataset: type: human-exams name: AGI-Eval metrics: - name: pass@1 type: pass@1 value: veriefied: false - task: type: text-generation dataset: type: commonsense name: WinoGrande metrics: - name: pass@1 type: pass@1 value: veriefied: false - task: type: text-generation dataset: type: commonsense name: OBQA metrics: - name: pass@1 type: pass@1 value: veriefied: false - task: type: text-generation dataset: type: commonsense name: SIQA metrics: - name: pass@1 type: pass@1 value: veriefied: false - task: type: text-generation dataset: type: commonsense name: PIQA metrics: - name: pass@1 type: pass@1 value: veriefied: false - task: type: text-generation dataset: type: commonsense name: Hellaswag metrics: - name: pass@1 type: pass@1 value: veriefied: false - task: type: text-generation dataset: type: commonsense name: TruthfulQA metrics: - name: pass@1 type: pass@1 value: veriefied: false - task: type: text-generation dataset: type: reading-comprehension name: BoolQ metrics: - name: pass@1 type: pass@1 value: veriefied: false - task: type: text-generation dataset: type: reading-comprehension name: SQuAD v2 metrics: - name: pass@1 type: pass@1 value: veriefied: false - task: type: text-generation dataset: type: reasoning name: ARC-C metrics: - name: pass@1 type: pass@1 value: veriefied: false - task: type: text-generation dataset: type: reasoning name: GPQA metrics: - name: pass@1 type: pass@1 value: veriefied: false - task: type: text-generation dataset: type: reasoning name: BBH metrics: - name: pass@1 type: pass@1 value: veriefied: false - task: type: text-generation dataset: type: code name: HumanEval metrics: - name: pass@1 type: pass@1 value: veriefied: false - task: type: text-generation dataset: type: code name: MBPP metrics: - name: pass@1 type: pass@1 value: veriefied: false - task: type: text-generation dataset: type: math name: GSM8K metrics: - name: pass@1 type: pass@1 value: veriefied: false - task: type: text-generation dataset: type: math name: MATH metrics: - name: pass@1 type: pass@1 value: veriefied: false - task: type: text-generation dataset: type: multilingual name: MGSM metrics: - name: pass@1 type: pass@1 value: veriefied: false --- # Granite-3.0-3B-A800M-Instruct ## Model Summary **Granite-3.0-3B-A800M-Instruct** is a lightweight and open-source 3B parameter model fine tuned from *Granite-3.0-3B-A800M-Base-4K* on a combination of open-source and proprietary instruction data with a **permissively licensed**. This language model is designed to excel in instruction following tasks such as summarization, problem-solving, text translation, reasoning, code tasks, funcion-calling, and more. - **Developers:** IBM Research - **GitHub Repository:** [ibm-granite/granite-language-models](https://github.com/ibm-granite/granite-language-models) - **Website**: [Granite Docs](https://www.ibm.com/granite/docs/) - **Paper:** [Granite Language Models](https://) - **Release Date**: October 21st, 2024 - **License:** [Apache 2.0](https://www.apache.org/licenses/LICENSE-2.0). ## Supported Languages English, German, Spanish, French, Japanese, Portuguese, Arabic, Czech, Italian, Korean, Dutch, Chinese (Simplified) ## Usage ### Intended use The model is designed to respond to general instructions and can be used to build AI assistants for multiple domains, including bussiness applications. ### Capabilities * Summarization * Text classification * Text extraction * Question-answering * Retrieval Augmented Generation (RAG) * Code related * Function-calling * Multilingual dialog use cases ### Generation This is a simple example of how to use **Granite-3.0-3B-A800M-Instruct** model. Install the following libraries: ```shell pip install torch torchvision torchaudio pip install accelerate pip install transformers ``` Then, copy the snippet from the section that is relevant for your usecase. ```python import torch from transformers import AutoModelForCausalLM, AutoTokenizer device = "auto" model_path = "ibm-granite/granite-3.0-3b-a800m-instruct" tokenizer = AutoTokenizer.from_pretrained(model_path) # drop device_map if running on CPU model = AutoModelForCausalLM.from_pretrained(model_path, device_map=device) model.eval() # change input text as desired chat = [ { "role": "user", "content": "Please list one IBM Research laboratory located in the United States. You should only output its name and location." }, ] chat = tokenizer.apply_chat_template(chat, tokenize=False, add_generation_prompt=True) # tokenize the text input_tokens = tokenizer(chat, return_tensors="pt").to(device) # generate output tokens output = model.generate(**input_tokens, max_new_tokens=100) # decode output tokens into text output = tokenizer.batch_decode(output) # print output print(output) ``` ## Model Architeture **Granite-3.0-3B-A800M-Instruct** is based on a decoder-only sparse Mixture of Experts(MoE) transformer architecture. Core components of this architecture are: Fine-grained Experts, Dropless Token Routing, and Load Balancing Loss. | Model | 2B Dense | 8B Dense | 1B MoE | 3B MoE | | :-------- | :--------| :--------| :--------| :-------- | | Embedding size | 2048 | 4096 | 1024 | **1536** | | Number of layers | 40 | 40 | 24 | **32** | | Attention head size | 64 | 128 | 64 | **64** | | Number of attention heads | 32 | 32 | 16 | **24** | | Number of KV heads | 8 | 8 | 8 | **8** | | MLP hidden size | 8192 | 12800 | 512 | **512** | | MLP activation | SwiGLU | SwiGLU | SwiGLU | **SwiGLU** | | Number of Experts | — | — | 32 | **40** | | MoE TopK | — | — | 8 | **8** | | Initialization std | 0.1 | 0.1 | 0.1 | **0.1** | | Sequence Length | 4096 | 4096 | 4096 | **4096** | | Position Embedding | RoPE | RoPE | RoPE | **RoPE** | | # Paremeters | 2.5B | 8.1B | 1.3B | **3.3B** | | # Active Parameters | 2.5B | 8.1B | 400M | **800M** | | # Training tokens | 12T | 12T | 10T | **10T** | ## Training Data Granite Language Instruct models are trained on a collection of publicly available datasets with non-restrictive license, as well as an IBM collection of synthetic datasets. We annotated and filtered these datasets to only include high-quality instances from each of them in our final mixture. This dataset selection is representative of the following domains: * English datasets: [Open-Platypus](https://huggingface.co/datasets/garage-bAInd/Open-Platypus), [WebInstructSub](https://huggingface.co/datasets/TIGER-Lab/WebInstructSub), [OASST-OctoPack](https://huggingface.co/datasets/bigcode/oasst-octopack), [Daring-Anteater](https://huggingface.co/datasets/nvidia/Daring-Anteater), [SoftAge-Multiturn](https://huggingface.co/datasets/SoftAge-AI/multi-turn_dataset), [Glaive-RAG-v1 ](https://huggingface.co/datasets/glaiveai/RAG-v1 ), [EvolKit-20k](https://huggingface.co/datasets/arcee-ai/EvolKit-20k ), [Magpie-Phi3-Pro-300K-Filtered](https://huggingface.co/datasets/Magpie-Align/Magpie-Phi3-Pro-300K-Filtered). * Multilingual datasets: [Aya Dataset](https://huggingface.co/datasets/CohereForAI/aya_dataset) and IBM Synthetic datasets (e.g., Blue Multilingual, Daring Anteater Translated). * Code datasets: [Glaive Code Assistant V3](https://huggingface.co/datasets/glaiveai/glaive-code-assistant-v3), [SQL Create Context Instruction](https://huggingface.co/datasets/bugdaryan/sql-create-context-instruction), and [Self-OSS-Instruct-SC2](https://huggingface.co/datasets/bigcode/self-oss-instruct-sc2-exec-filter-50k). Single and multi-turn IBM synthetic datasets, including a set of datasets generated via the evol-instruct method. * Math: [MetaMathQA](https://huggingface.co/datasets/meta-math/MetaMathQA), [StackMathQA](https://huggingface.co/datasets/math-ai/StackMathQA ), and [MathInstruct](https://huggingface.co/datasets/TIGER-Lab/MathInstruct) * Tools: [xlam-function-calling](https://huggingface.co/datasets/Salesforce/xlam-function-calling-60k), [Glaive Function Calling V2](https://huggingface.co/datasets/glaiveai/glaive-function-calling-v2), [Hermes Function Calling V1](https://huggingface.co/datasets/NousResearch/hermes-function-calling-v1), and IBM Synthetic API data. * Safety: [SimpleSafetyTests](https://huggingface.co/datasets/Bertievidgen/SimpleSafetyTests), [HarmBench Behaviors](https://github.com/centerforaisafety/HarmBench/blob/main/data/behavior_datasets/harmbench_behaviors_text_all.csv), [Strong Reject](https://github.com/alexandrasouly/strongreject/blob/main/strongreject_dataset/strongreject_dataset.csv), [AdvBench](https://huggingface.co/datasets/walledai/AdvBench), [MistralGuard](https://huggingface.co/datasets/natolambert/xstest-v2-copy), [Do-Not-Answer](https://huggingface.co/datasets/LibrAI/do-not-answer), and IBM Synthetic data for safety. ## Infrastructure We train the Granite Language models using IBM's super computing cluster, Blue Vela, which is outfitted with NVIDIA H100 GPUs. This cluster provides a scalable and efficient infrastructure for training our models over thousands of GPUs. ## Ethical Considerations and Limitations Granite instruct models are primarily finetuned using instruction-response pairs mostly in English, but also in German, Spanish, French, Japanese, Portuguese, Arabic, Czech, Italian, Korean, Dutch, and Chinese (Simplified). As this model has been exposed to multilingual data, it can handle multilingual dialog use cases with a limited performance in non-English tasks. In such case, introducing a small number of examples (few-shot) can help the model in generating more accurate outputs. The model also inherits ethical considerations and limitations from its base model. For more information, please refer to *[Granite-3.0-3B-A800M-Base-4K](https://huggingface.co/ibm-granite/granite-3.0-3b-a800m-base)* model card.