mosaicml

MPT-7B is a decoder-style transformer pretrained from scratch on 1T tokens of English text and code. This model was trained by MosaicML. MPT-7B is part of the family of MosaicPretrainedTransformer (MPT) models, which use a modified transformer architecture optimized for efficient training and inference. These architectural changes include performance-optimized layer implementations and the elimination of context length limits by replacing positional embeddings with Attention with Linear Biases (ALiBi). Thanks to these modifications, MPT models can be trained with high throughput efficiency and stable convergence. MPT models can also be served efficiently with both standard HuggingFace pipelines and NVIDIA's FasterTransformer. This model uses the MosaicML LLM codebase, which can be found in the llm-foundry repository. It was trained by MosaicML’s NLP team on the MosaicML platform for LLM pretraining, finetuning, and inference. Licensed for the possibility of commercial use (unlike LLaMA). Trained on a large amount of data (1T tokens like LLaMA vs. 300B for Pythia, 300B for OpenLLaMA, and 800B for StableLM). Prepared to handle extremely long inputs thanks to ALiBi (we finetuned MPT-7B-StoryWriter-65k+ on up to 65k inputs and can handle up to 84k vs. 2k-4k for other open source models). Capable of fast training and inference (via FlashAttention and FasterTransformer) Equipped with highly efficient open-source training code via the llm-foundry repository MPT-7B-StoryWriter-65k+: a model designed to read and write fictional stories with super long context lengths. Built by finetuning MPT-7B with a context length of 65k tokens on a filtered fiction subset of the books3 dataset. At inference time, thanks to ALiBi, MPT-7B-StoryWriter-65k+ can extrapolate even beyond 65k tokens. We demonstrate generations as long as 80k tokens on a single A100-80GB GPU in our blogpost. License: Apache 2.0 MPT-7B-Instruct: a model for short-form instruction following. Built by finetuning MPT-7B on a dataset we also release, derived from the Databricks Dolly-15k and the Anthropic Helpful and Harmless (HH-RLHF) datasets. License: Apache 2.0 MPT-7B-Chat: a chatbot-like model for dialogue generation. Built by finetuning MPT-7B on the ShareGPT-Vicuna, HC3, Alpaca, HH-RLHF, and Evol-Instruct datasets. License: CC-By-NC-SA-4.0 Blog post: Introducing MPT-7B: A New Standard for Open-Source, Commercially Usable LLMs Codebase (mosaicml/llm-foundry repo) Questions: Feel free to contact us via the MosaicML Community Slack! This model is best used with the MosaicML llm-foundry repository for training and finetuning. Note: This model requires that `trustremotecode=True` be passed to the `frompretrained` method. This is because we use a custom `MPT` model architecture that is not yet part of the Hugging Face `transformers` package. `MPT` includes options for many training efficiency features such as FlashAttention, ALiBi, QK LayerNorm, and more. To use the optimized triton implementation of FlashAttention, you can load the model on GPU (`cuda:0`) with `attnimpl='triton'` and with `bfloat16` precision: Although the model was trained with a sequence length of 2048, ALiBi enables users to increase the maximum sequence length during finetuning and/or inference. For example: This model was trained with the EleutherAI/gpt-neox-20b tokenizer. The model can then be used, for example, within a text-generation pipeline. Note: when running Torch modules in lower precision, it is best practice to use the torch.autocast context manager. The architecture is a modification of a standard decoder-only transformer. The model has been modified from a standard transformer in the following ways: It uses FlashAttention It uses ALiBi (Attention with Linear Biases) and does not use positional embeddings It does not use biases | Hyperparameter | Value | |----------------|-------| |nparameters | 6.7B | |nlayers | 32 | | nheads | 32 | | dmodel | 4096 | | vocab size | 50432 | | sequence length | 2048 | Data was formatted using the MosaicML StreamingDataset library to host our data in object storage and efficiently stream it to our compute cluster during training. StreamingDataset obviates the need to download the whole dataset before starting training, and allows instant resumption of training from any point in the dataset. The model was trained for 1T tokens (with batch size 1760 and sequence length 2048). It was trained on the following data mix: | Data Source | Number of Tokens in Source | Proportion | Effective Number of Tokens | Epochs | |-------------|----------------------------|------------|----------------------------|--------| | mC4 3.1.0 - English | 417.99 B | 0.33 | 330 B | 0.14 | | C4 - English - SemDedup 80% | 100.42 B | 0.299 | 299 B | 2.98 | | RedPajama - CommonCrawl | 878.45 B | 0.1 | 100 B | 0.11 | | The Stack - Selected Languages | 463.78 B | 0.1 | 100 B | 0.22 | | RedPajama - Wikipedia - En | 4.87 B | 0.04 | 40 B | 8.21 | | The Stack - Markdown | 107.07 B | 0.035 | 35 B | 0.33 | | S2ORC | 48.85 B | 0.033 | 33 B | 0.68 | | RedPajama - Books | 26.02 B | 0.03 | 30B | 1.15 | | RedPajama - arXiv | 28.10 B | 0.019 | 19 B | 0.68 | | RedPajama - StackExchange | 20.54 B | 0.014 | 14 B |0.68 | Samples for each batch were selected from one of the datasets with the probability specified above. The examples were shuffled within each dataset, and each example was constructed from as many sequences from that dataset as were necessary to fill the 2048 sequence length. The data was tokenized using the EleutherAI/gpt-neox-20b tokenizer. This BPE tokenizer has a number of desirable characteristics, most of which are relevant for tokenizing code: (1) It was trained on a diverse mix of data that includes code (The Pile) (2) It applies consistent space delimitation, unlike the GPT2 tokenizer which tokenizes inconsistently depending on the presence of prefix spaces (3) It contains tokens for repeated space characters, which allows superior compression of text with large amounts of repeated space characters. The model vocabulary size of 50432 was set to be a multiple of 128 (as in MEGATRON-LM), model flop utilization (MFU) increased by up to four percentage points. This model was trained on 440 A100-40GBs for about 9.5 days using the MosaicML Platform. The model was trained with sharded data parallelism using FSDP and used the LION optimizer. The following language is modified from EleutherAI's GPT-NeoX-20B MPT-7B (Base) is not intended for deployment without finetuning. It should not be used for human-facing interactions without further guardrails and user consent. MPT-7B can produce factually incorrect output, and should not be relied on to produce factually accurate information. MPT-7B was trained on various public datasets. While great efforts have been taken to clean the pretraining data, it is possible that this model could generate lewd, biased or otherwise offensive outputs. If you're interested in training and deploying your own MPT or LLMs on the MosaicML Platform, sign up here. The license on this model does not constitute legal advice. We are not responsible for the actions of third parties who use this model. Please cosult an attorney before using this model for commercial purposes.

MPT-7B-StoryWriter-65k+ is a model designed to read and write fictional stories with super long context lengths. It was built by finetuning MPT-7B with a context length of 65k tokens on a filtered fiction subset of the books3 dataset. At inference time, thanks to ALiBi, MPT-7B-StoryWriter-65k+ can extrapolate even beyond 65k tokens. We demonstrate generations as long as 84k tokens on a single node of 8 A100-80GB GPUs in our blogpost. License: Apache 2.0 This model was trained by MosaicML and follows a modified decoder-only transformer architecture. Blog post: Introducing MPT-7B: A New Standard for Open-Source, Commercially Usable LLMs Codebase (mosaicml/llm-foundry repo) Questions: Feel free to contact us via the MosaicML Community Slack! Note: This model requires that `trustremotecode=True` be passed to the `frompretrained` method. This is because we use a custom model architecture that is not yet part of the `transformers` package. It includes options for many training efficiency features such as FlashAttention (Dao et al. 2022), ALiBi, QK LayerNorm, and more. To use the optimized triton implementation of FlashAttention, you can load the model on GPU (`cuda:0`) with `attnimpl='triton'` and with `bfloat16` precision: Although the model was trained with a sequence length of 2048 and finetuned with a sequence length of 65536, ALiBi enables users to increase the maximum sequence length during finetuning and/or inference. For example: This model was trained with the EleutherAI/gpt-neox-20b tokenizer. The model can then be used, for example, within a text-generation pipeline. Note: when running Torch modules in lower precision, it is best practice to use the torch.autocast context manager. These were not created by MosaicML, but you may find them useful. These links are not an endorsement of the creators or their content. - Oobabooga Running MPT-7B-Storywriter - NEW MPT-7B-StoryWriter CRUSHES GPT-4! - Has a long section on running locally using Oobabooga The full text of the The Great Gatsby (67873 tokens) was fed to the model, followed by the text "EPILOGUE" > ... > > So we beat on, boats against the current, borne back ceaselessly into the past > > EPILOGUE >I have lived so long in the shadow of this dream that I may never see the bright sunlight of reality. I may look up some morning and find that I am six feet under, and that Gatsby, who gave me the dream and everything else, will have vanished behind me. > >And so it was. I had hardly crossed the threshold of the old world when my eye caught the headlines of a newspaper, and I knew at once that my dream was over. I knew that Gatsby was dead. > >And I knew that day that I didn't love her. > >That night I took the last train for West Egg, and when I reached home I went directly to my attic and brought down all Gatsby's letters. > >I have sorted them by date and placed them in the order in which they were written. They are very few in number. > >They are dated between April of nineteen eighteen and January of nineteen nineteen, and the last one I have here is the one that told me he was dead. > >There is something in that last letter that frightens me. It is the sense of a finality, of a completeness. I never heard of him after that, and I was glad. > >But a few months ago I read somewhere that he had been seen in New York. He had become a great man. > >And I knew that he had not changed at all. The architecture is a modification of a standard decoder-only transformer. The model has been modified from a standard transformer in the following ways: It uses FlashAttention It uses ALiBi (Attention with Linear Biases) and does not use positional embeddings It does not use biases | Hyperparameter | Value | |----------------|-------| |nparameters | 6.7B | |nlayers | 32 | | nheads | 32 | | dmodel | 4096 | | vocab size | 50432 | | sequence length | 65536 | For more details on the pretraining process, see MPT-7B. The data was tokenized using the EleutherAI/gpt-neox-20b tokenizer. This model was trained on 8 A100-80GBs for about 2 days using the MosaicML Platform. The model was trained with sharded data parallelism using FSDP and used the LION optimizer. The following language is modified from EleutherAI's GPT-NeoX-20B MPT-7B-StoryWriter can produce factually incorrect output, and should not be relied on to produce factually accurate information. MPT-7B-StoryWriter was trained on various public datasets. While great efforts have been taken to clean the pretraining data, it is possible that this model could generate lewd, biased or otherwise offensive outputs. This model was finetuned by Alex Trott and the MosaicML NLP team If you're interested in training and deploying your own MPT or LLMs on the MosaicML Platform, sign up here. The license on this model does not constitute legal advice. We are not responsible for the actions of third parties who use this model. Please cosult an attorney before using this model for commercial purposes.