--- license: apache-2.0 datasets: - tiiuae/falcon-refinedweb pipeline_tag: text-generation library_name: openlm tags: - mamba - linear language: - en model-index: - name: mamba-7b results: - task: type: text-generation dataset: type: MMLU name: MMLU metrics: - name: accuracy type: accuracy value: 33.3 verified: false - task: type: text-generation dataset: type: HellaSwag name: HellaSwag metrics: - name: accuracy type: accuracy value: 77.9 verified: false - task: type: text-generation dataset: type: PIQA name: PIQA metrics: - name: accuracy type: accuracy value: 81.0 verified: false - task: type: text-generation dataset: type: Winogrande name: Winogrande metrics: - name: accuracy type: accuracy value: 71.8 verified: false - task: type: text-generation dataset: type: ai2_arc name: ARC-E metrics: - name: accuracy type: accuracy value: 77.5 verified: false - task: type: text-generation dataset: type: ai2_arc name: ARC-C metrics: - name: accuracy type: accuracy value: 46.7 verified: false --- # Mamba-7B (insert cool midjourney pic here?)
This is a 7B parameter model with the [Mamba](https://arxiv.org/abs/2312.00752) architecture, trained on 1.2T tokens of the [RefinedWeb](https://huggingface.co/datasets/tiiuae/falcon-refinedweb) dataset. Mamba is a state-space model that does not use self-attention unlike the standard transformer architecture. It has shown strong performance on various natural language benchmarks. To date, the largest publicly released pure-Mamba pretrain is [Mamba-2.8B](https://huggingface.co/state-spaces/mamba-2.8b). We follow their training recipe and release our version of Mamba-7B. ## Model Details - **Developed by**: [Toyota Research Institute](https://www.tri.global/our-work/robotics) - **Model Type**: This is an auto-regressive language model based on the [Mamba](https://arxiv.org/abs/2312.00752) architecture. - **Dataset**: Trained on 1.2T tokens of [RefinedWeb](https://huggingface.co/datasets/tiiuae/falcon-refinedweb) - **Tokenizer**: `EleutherAI/gpt-neox-20b` - **Library**: [OpenLM](https://github.com/mlfoundations/open_lm/) - **License**: This model is licensed under [Apache License, Version 2.0](https://www.apache.org/licenses/LICENSE-2.0). | Parameters | Hidden Size | Layers | Vocab Size | Sequence Length | |------------|-------------|--------| ---------- | --------------- | | 7B | 4096 | 64 | 50432 | 2048 | ## Training Details - Mamba-7B was trained using AWS SageMaker on 128 H100 80GB GPUs. - Training began in March 2023 and lasted around 3 weeks (some down time due to crashes and loss spikes) | **Hyperparameter** | **Value** | |--------------------|------------| | Precision | `bfloat16` | | Optimizer | AdamW | | Learning rate | 3e-4 | | LR cooldown end | 1e-5 | | QK-norm | False | | Warmup steps | 2000 | | Z-loss | 1e-4 | | Batch size | 2M | ## Usage (Not functional yet. Not sure if this is the right flow. Will work on this.)
This model was trained using [OpenLM](https://github.com/mlfoundations/open_lm/). To use HuggingFace models trained with OpenLM, first install the OpenLM package ```bash pip install openlm ``` Importing from `openlm_hf` will automatically import the necessary classes. ```python from openlm_hf import * # registers the Auto* classes from transformers import AutoTokenizer, AutoModelForCausalLM tokenizer = AutoTokenizer.from_pretrained("tri-ml/mamba-7b-rw") model = AutoModelForCausalLM.from_pretrained("tri-ml/mamba-7b-rw").cuda() inputs = tokenizer(["A beautiful flower"], return_tensors="pt") gen_kwargs = {"max_length": 128, "top_p": 0.8, "temperature": 0.8, "do_sample": True, "repetition_penalty": 1.1} output = model.generate(inputs['input_ids'], **gen_kwargs) output = tokenizer.decode(output[0].tolist(), skip_special_tokens=True) print(output) # A beautiful flower box made of white rose wood. It is a perfect gift for weddings, birthdays and anniversaries. # All the roses are from our farm Roses Flanders. Therefor you know that these flowers last much longer than those in store or online! ``` ## Performance Evaluation Our evaluations were done using the [Eleuther LM Eval Harness](https://github.com/EleutherAI/lm-evaluation-harness) repo. Below we report the performance of Mamba 7B compared to other base models.
| | MMLU (5-shot) | HellaSwag | PIQA | Winogrande | ARC-E | ARC-C | | ----------------- | ---------------- | ------------- | -------- | -------------- | --------- | --------- | | Mamba-1.4B | 25.2 | 59.0 | 73.9 | 61.4 | 65.5 | 32.9 | | Mamba-2.8B | 26.3 | 66.2 | 75.8 | 63.4 | 69.7 | 36.3 | | Llama2-7B | 45.9 | 76.0 | 79.1 | 69.1 | 76.3 | 46.3 | | Gemma-7B | 62.9 | 80.7 | 81.9 | 73.7 | 81.1 | 53.2 | | Mistral-7B | 62.4 | 81.0 | 82.1 | 74.0 | 80.9 | 53.8 | | **Mamba-7B** | 33.3 | 77.9 | 81.0 | 71.8 | 77.5 | 46.7 |
## How to Cite If you use this model, please cite our paper on Linearizing Large Language Models. ``` @article{Mercat2024Linearizing, title={Linearizing Large Language Models}, author={Jean Mercat and Igor Vasiljevic and Sedrick Keh and Kushal Arora and Achal Dave and Adrien Gaidon and Thomas Kollar}, journal={ArXiv}, year={2024}, volume={}, } ``` ## Citations Mamba ``` @article{mamba, title={Mamba: Linear-Time Sequence Modeling with Selective State Spaces}, author={Gu, Albert and Dao, Tri}, journal={arXiv preprint arXiv:2312.00752}, year={2023} } ``` OpenLM ``` @misc{open_lm, author = {Gururangan, Suchin and Wortsman, Mitchell and Gadre, Samir Yitzhak and Dave, Achal and Kilian, Maciej and Shi, Weijia and Mercat, Jean and Smyrnis, Georgios and Ilharco, Gabriel and Jordan, Matt and Heckel, Reinhard and Dimakis, Alex and Farhadi, Ali and Shankar, Vaishaal and Schmidt, Ludwig}, title = {{open_lm}: a minimal but performative language modeling (LM) repository}, year = {2023}, note = {GitHub repository}, url = {https://github.com/mlfoundations/open_lm/} } ```