id
int64 0
27
| text
stringlengths 635
1.86k
| source
listlengths 3
3
|
|---|---|---|
0
|
Introducing MPT-7B, the first entry in our MosaicML Foundation Series. MPT-7B is a transformer trained from scratch on 1T tokens of text and code. It is open source, available for commercial use, and matches the quality of LLaMA-7B. MPT-7B was trained on the MosaicML platform in 9.5 days with zero human intervention at a cost of ~$200k.
Large language models (LLMs) are changing the world, but for those outside well-resourced industry labs, it can be extremely difficult to train and deploy these models. This has led to a flurry of activity centered on open-source LLMs, such as the LLaMA series from Meta, the Pythia series from EleutherAI, the StableLM series from StabilityAI, and the OpenLLaMA model from Berkeley AI Research.
Today, we at MosaicML are releasing a new model series called MPT (MosaicML Pretrained Transformer) to address the limitations of the above models and finally provide a commercially-usable, open-source model that matches (and - in many ways - surpasses) LLaMA-7B. Now you can train, finetune, and deploy your own private MPT models, either starting from one of our checkpoints or training from scratch. For inspiration, we are also releasing three finetuned models in addition to the base MPT-7B: MPT-7B-Instruct, MPT-7B-Chat, and MPT-7B-StoryWriter-65k+, the last of which uses a context length of 65k tokens!
Our MPT model series is:
Licensed for commercial use (unlike LLaMA).
|
[
"https://www.mosaicml.com/blog/mpt-7b",
"https://stability.ai/blog/stability-ai-launches-the-first-of-its-stablelm-suite-of-language-models",
"https://lmsys.org/blog/2023-03-30-vicuna/"
] |
1
|
Our MPT model series is:
Licensed for 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 trained on up to 65k inputs and can handle up to 84k vs. 2k-4k for other open source models).
Optimized for fast training and inference (via FlashAttention and FasterTransformer)
Equipped with highly efficient open-source training code.
We rigorously evaluated MPT on a range of benchmarks, and MPT met the high quality bar set by LLaMA-7B.
Today, we are releasing the base MPT model and three other finetuned variants that demonstrate the many ways of building on this base model:
MPT-7B Base:
MPT-7B Base is a decoder-style transformer with 6.7B parameters. It was trained on 1T tokens of text and code that was curated by MosaicML's data team. This base model includes FlashAttention for fast training and inference and ALiBi for finetuning and extrapolation to long context lengths.
License: Apache-2.0
HuggingFace Link: https://huggingface.co/mosaicml/mpt-7b
MPT-7B-StoryWriter-65k+
|
[
"https://www.mosaicml.com/blog/mpt-7b",
"https://stability.ai/blog/stability-ai-launches-the-first-of-its-stablelm-suite-of-language-models",
"https://lmsys.org/blog/2023-03-30-vicuna/"
] |
2
|
MPT-7B-StoryWriter-65k+
MPT-7B-StoryWriter-65k+ is a model designed to read and write 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, and we have demonstrated generations as long as 84k tokens on a single node of A100-80GB GPUs.
License: Apache-2.0
HuggingFace Link: https://huggingface.co/mosaicml/mpt-7b-storywriter
MPT-7B-Instruct
MPT-7B-Instruct is a model for short-form instruction following. Built by finetuning MPT-7B on a dataset we also release, derived from Databricks Dolly-15k and Anthropic's Helpful and Harmless datasets.
License: CC-By-SA-3.0
HuggingFace Link: https://huggingface.co/mosaicml/mpt-7b-instruct
MPT-7B-Chat
MPT-7B-Chat is a chatbot-like model for dialogue generation. Built by finetuning MPT-7B on the ShareGPT-Vicuna, HC3, Alpaca, Helpful and Harmless, and Evol-Instruct datasets.
License: CC-By-NC-SA-4.0 (non-commercial use only)
HuggingFace Link: https://huggingface.co/mosaicml/mpt-7b-chat
|
[
"https://www.mosaicml.com/blog/mpt-7b",
"https://stability.ai/blog/stability-ai-launches-the-first-of-its-stablelm-suite-of-language-models",
"https://lmsys.org/blog/2023-03-30-vicuna/"
] |
3
|
We hope businesses and the open-source community will build on this effort: alongside the model checkpoints, we have open-sourced the entire codebase for pretraining, finetuning, and evaluating MPT via our new MosaicML LLM Foundry!
This release is more than just a model checkpoint: it's an entire framework for building great LLMs with MosaicML's usual emphasis on efficiency, ease-of-use, and rigorous attention to detail. These models were built by MosaicML's NLP team on the MosaicML platform with the exact same tools our customers use (just ask our customers, like Replit!).
e trained MPT-7B with ZERO human intervention from start to finish: over 9.5 days on 440 GPUs, the MosaicML platform detected and addressed 4 hardware failures and resumed the training run automatically, and - due to architecture and optimization improvements we made - there were no catastrophic loss spikes. Check out our empty training logbook for MPT-7B!
Training and Deploying Your Own Custom MPT
If you'd like to start building and deploying your own custom MPT models on the MosaicML platform, sign up here to get started.
For more engineering details on data, training, and inference, skip ahead to the section below.
For more information about our four new models, read on!
Introducing the Mosaic Pretrained Transformers (MPT)
MPT models are GPT-style decoder-only transformers with several improvements: performance-optimized layer implementations, architecture changes that provide greater training stability, and the elimination of context length limits by replacing positional embeddings with ALiBi. Thanks to these modifications, customers can train MPT models with efficiency (40-60% MFU) without diverging from loss spikes and can serve MPT models with both standard HuggingFace pipelines and FasterTransformer.
MPT-7B (Base Model)
|
[
"https://www.mosaicml.com/blog/mpt-7b",
"https://stability.ai/blog/stability-ai-launches-the-first-of-its-stablelm-suite-of-language-models",
"https://lmsys.org/blog/2023-03-30-vicuna/"
] |
4
|
MPT-7B (Base Model)
MPT-7B matches the quality of LLaMA-7B and outperforms other open source 7B - 20B models on standard academic tasks. To evaluate model quality, we compiled 11 open-source benchmarks commonly used for in-context learning (ICL) and formatted and evaluated them in an industry-standard manner. We also added our own self-curated Jeopardy benchmark to evaluate the model's ability to produce factually correct answers to challenging questions.
See Table 1 for a comparison of zero-shot performance between MPT and other models:
To ensure apples-to-apples comparisons, we fully re-evaluated each model: the model checkpoint was run through our open source LLM Foundry eval framework with the same (empty) prompt strings and no model-specific prompt tuning. For full details on the evaluation, see the Appendix. In previous benchmarks, our setup is 8x faster than other eval frameworks on a single GPU and seamlessly achieves linear scaling with multiple GPUs. Built-in support for FSDP makes it possible to evaluate large models and use larger batch sizes for further acceleration.
We invite the community to use our evaluation suite for their own model evaluations and to submit pull requests with additional datasets and ICL task types so we can ensure the most rigorous possible evaluation.
MPT-7B-StoryWriter-65k+
Most open-source language models can only handle sequences with up to a few thousand tokens (see Figure 1). But with the MosaicML platform and a single node of 8xA100-80GB, you can easily finetune MPT-7B to handle context lengths up to 65k! The ability to handle such extreme context length adaptation comes from ALiBi, one of the key architectural choices in MPT-7B.
|
[
"https://www.mosaicml.com/blog/mpt-7b",
"https://stability.ai/blog/stability-ai-launches-the-first-of-its-stablelm-suite-of-language-models",
"https://lmsys.org/blog/2023-03-30-vicuna/"
] |
5
|
To show off this capability and to get you thinking about what you could do with a 65k context window, we are releasing MPT-7B-StoryWriter-65k+. StoryWriter was finetuned from MPT-7B for 2500 steps on 65k-token excerpts of fiction books contained in the books3 corpus. Like pretraining, this finetuning process used a next-token-prediction objective. Once we prepared the data, all that was needed for training was Composer with FSDP, activation checkpointing, and a microbatch size of 1.
As it turns out, the full text of The Great Gatsby weighs in at just under 68k tokens. So, naturally, we had StoryWriter read The Great Gatsby and generate an epilogue. One of the epilogues we generated is in Figure 2. StoryWriter took in The Great Gatsby in about 20 seconds (about 150k words-per-minute). Due to the long sequence length, its "typing" speed is slower than our other MPT-7B models, about 105 words-per-minute.
Even though StoryWriter was fine-tuned with a 65k context length, ALiBi makes it possible for the model to extrapolate to even longer inputs than it was trained on: 68k tokens in the case of The Great Gatsby, and up to 84k tokens in our testing.
MPT-7B-Instruct
|
[
"https://www.mosaicml.com/blog/mpt-7b",
"https://stability.ai/blog/stability-ai-launches-the-first-of-its-stablelm-suite-of-language-models",
"https://lmsys.org/blog/2023-03-30-vicuna/"
] |
6
|
MPT-7B-Instruct
LLM pretraining teaches the model to continue generating text based on the input it was provided. But in practice, we expect LLMs to treat the input as instructions to follow. Instruction finetuning is the process of training LLMs to perform instruction-following in this way. By reducing the reliance on clever prompt engineering, instruction finetuning makes LLMs more accessible, intuitive, and immediately usable. The progress of instruction finetuning has been driven by open-source datasets like FLAN, Alpaca, and the Dolly-15k dataset.
We created a commercially-usable instruction-following variant of our model called MPT-7B-Instruct. We liked the commercial license of Dolly, but wanted more data, so we augmented Dolly with a subset of Anthropic's Helpful & Harmless dataset, quadrupling the dataset size while maintaining a commercial license.
This new aggregate dataset, released here, was used to finetune MPT-7B, resulting in MPT-7B-Instruct, which is commercially usable. Anecdotally, we find MPT-7B-Instruct to be an effective instruction-follower. (See Figure 3 for an example interaction.) With its extensive training on 1 trillion tokens, MPT-7B-Instruct should be competitive with the larger dolly-v2-12b, whose base model, Pythia-12B, was only trained on 300 billion tokens.
We are releasing the code, weights, and an online demo of MPT-7B-Instruct. We hope that the small size, competitive performance, and commercial license of MPT-7B-Instruct will make it immediately valuable to the community.
MPT-7B-Chat
|
[
"https://www.mosaicml.com/blog/mpt-7b",
"https://stability.ai/blog/stability-ai-launches-the-first-of-its-stablelm-suite-of-language-models",
"https://lmsys.org/blog/2023-03-30-vicuna/"
] |
7
|
MPT-7B-Chat
A multi-turn conversation with the chat model in which it suggests high-level approaches to solving a problem (using AI to protect endangered wildlife) and then proposes an implementation of one of them in Python using Keras.
We have also developed MPT-7B-Chat, a conversational version of MPT-7B. MPT-7B-Chat has been finetuned using ShareGPT-Vicuna, HC3, Alpaca, Helpful and Harmless, and Evol-Instruct, ensuring that it is well-equipped for a wide array of conversational tasks and applications. It uses the ChatML format, which provides a convenient and standardized way to pass the model system messages and helps prevent malicious prompt injection.
While MPT-7B-Instruct focuses on delivering a more natural and intuitive interface for instruction-following, MPT-7B-Chat aims to provide seamless, engaging multi-turn interactions for users. (See Figure 4 for an example interaction.)
As with MPT-7B and MPT-7B-Instruct, we are releasing the code, weights, and an online demo for MPT-7B-Chat.
How we built these models on the MosaicML platform
The models released today were built by the MosaicML NLP team, but the tools we used are the exact same ones available to every customer of MosaicML.
Think of MPT-7B as a demonstration – our small team was able to build these models in only a few weeks, including the data preparation, training, finetuning, and deployment (and writing this blog!). Let's take a look at the process of building MPT-7B with MosaicML:
Data
|
[
"https://www.mosaicml.com/blog/mpt-7b",
"https://stability.ai/blog/stability-ai-launches-the-first-of-its-stablelm-suite-of-language-models",
"https://lmsys.org/blog/2023-03-30-vicuna/"
] |
8
|
Data
We wanted MPT-7B to be a high-quality standalone model and a useful jumping off point for diverse downstream uses. Accordingly, our pretraining data came from a MosaicML-curated mix of sources, which we summarize in Table 2 and describe in detail in the Appendix. Text was tokenized using the EleutherAI GPT-NeoX-20B tokenizer and the model was pretrained on 1 trillion tokens. This dataset emphasizes English natural language text and diversity for future uses (e.g., code or scientific models), and includes elements of the recently-released RedPajama dataset so that the web crawl and Wikipedia portions of the dataset contain up-to-date information from 2023.
Tokenizer
We used EleutherAI's GPT-NeoX 20B tokenizer. This BPE tokenizer has a number of desirable characteristics, most of which are relevant for tokenizing code:
Trained on a diverse mix of data that includes code (The Pile)
Applies consistent space delimitation, unlike the GPT2 tokenizer which tokenizes inconsistently depending on the presence of prefix spaces
Contains tokens for repeated space characters, which allows superior compression of text with large amounts of repeated space characters.
The tokenizer has a vocabulary size of 50257, but we set the model vocabulary size to 50432. The reasons for this were twofold: First, to make it a multiple of 128 (as in Shoeybi et al.), which we found improved MFU by up to four percentage points in initial experiments. Second, to leave tokens available that can be used in subsequent UL2 training.
Efficient Data Streaming
We leveraged MosaicML's StreamingDataset to host our data in a standard cloud object store and efficiently stream it to our compute cluster during training. StreamingDataset provides a number of advantages:
Obviates the need to download the whole dataset before starting training.
|
[
"https://www.mosaicml.com/blog/mpt-7b",
"https://stability.ai/blog/stability-ai-launches-the-first-of-its-stablelm-suite-of-language-models",
"https://lmsys.org/blog/2023-03-30-vicuna/"
] |
9
|
Obviates the need to download the whole dataset before starting training.
Allows instant resumption of training from any point in the dataset. A paused run can be resumed without fast-forwarding the dataloader from the start.
Is fully deterministic. Samples are read in the same order regardless of the number of GPUs, nodes, or CPU workers.
Allows arbitrary mixing of data sources in: simply enumerate the your data sources and desired proportions of the total training data, and StreamingDataset handles the rest. This made it extremely easy to run preparatory experiments on different data mixes.
Check out the StreamingDataset blog for more details!
Training Compute
All MPT-7B models were trained on the MosaicML platform with the following tools:
Compute: A100-40GB and A100-80GB GPUs from Oracle Cloud
Orchestration and Fault Tolerance: MCLI and MosaicML platform
Data: OCI Object Storage and StreamingDataset
Training software: Composer, PyTorch FSDP, and LLM Foundry
As shown in Table 3, nearly all of the training budget was spent on the base MPT-7B model, which took ~9.5 days to train on 440xA100-40GB GPUs, and cost ~$200k. The finetuned models took much less compute and were much cheaper – ranging between a few hundred and few thousand dollars each.
Each of these training recipes can be fully customized. For example, if you'd like to start from our open source MPT-7B and finetune it on proprietary data with a long context length, you can do that today on the MosaicML platform.
|
[
"https://www.mosaicml.com/blog/mpt-7b",
"https://stability.ai/blog/stability-ai-launches-the-first-of-its-stablelm-suite-of-language-models",
"https://lmsys.org/blog/2023-03-30-vicuna/"
] |
10
|
As another example, to train a new model from scratch on a custom domain (e.g. on biomedical text or code), simply reserve short-term large blocks of compute with MosaicML's hero cluster offering. Just pick the desired model size and token budget, upload your data to an object store like S3, and launch an MCLI job. You will have your very own custom LLM in just days!
Check out our earlier LLM blog post for guidance on the times and costs to train different LLMs. Find the latest throughput data for specific model configurations here. In line with our previous work, all MPT-7B models were trained with Pytorch FullyShardedDataParallelism (FSDP) and without tensor- or pipeline- parallelism.
Training Stability
As many teams have documented, training LLMs with billions of parameters on hundreds-to-thousands of GPUs is incredibly challenging. Hardware will fail frequently and in creative and unexpected ways. Loss spikes will derail training. Teams must "babysit" the training run 24/7 in case of failures and apply manual interventions when things go wrong. Check out the OPT logbook for a candid example of the many perils awaiting anyone training an LLM.
At MosaicML, our research and engineering teams have worked tirelessly over the last 6 months to eliminate these issues. As a result, our MPT-7B training logbook (Figure 5) is very boring! We trained MPT-7B on 1 trillion tokens from start to finish with no human intervention. No loss spikes, no mid-stream learning rate changes, no data skipping, automatic handling of dead GPUs, etc.
|
[
"https://www.mosaicml.com/blog/mpt-7b",
"https://stability.ai/blog/stability-ai-launches-the-first-of-its-stablelm-suite-of-language-models",
"https://lmsys.org/blog/2023-03-30-vicuna/"
] |
11
|
How did we do this? First, we addressed convergence stability with architecture and optimization improvements. Our MPT models use ALiBi rather than positional embeddings, which we found to improve resilience to loss spikes. We also train our MPT models with the Lion optimizer rather than AdamW, which provides stable update magnitudes and cuts optimizer state memory in half.
Second, we used the MosaicML platform's NodeDoctor feature to monitor for and resolve hardware failures and the JobMonitor feature to resume runs after these failures were resolved. These features enabled us to train MPT-7B with no human intervention from start to finish despite 4 hardware failures during the run. See Figure 6 for a closeup view of what autoresumption looks like on the MosaicML platform.
Inference
MPT is designed to be fast, easy, and cheap to deploy for inference. To begin with, all MPT models are subclassed from the HuggingFace PretrainedModel base class, which means that they are fully compatible with the HuggingFace ecosystem. You can upload MPT models to the HuggingFace Hub, generate outputs with standard pipelines like `model.generate(...)`, build HuggingFace Spaces (see some of ours here!), and more.
What about performance? With MPT's optimized layers (including FlashAttention and low precision layernorm), the out-of-the-box performance of MPT-7B when using `model.generate(...)` is 1.5x-2x faster than other 7B models like LLaMa-7B. This makes it easy to build fast and flexible inference pipelines with just HuggingFace and PyTorch.
But what if you really need the best performance? In that case, directly port MPT weights to FasterTransformer or ONNX. Check out the LLM Foundry's inference folder for scripts and instructions.
|
[
"https://www.mosaicml.com/blog/mpt-7b",
"https://stability.ai/blog/stability-ai-launches-the-first-of-its-stablelm-suite-of-language-models",
"https://lmsys.org/blog/2023-03-30-vicuna/"
] |
12
|
Finally, for the best hosting experience, deploy your MPT models directly on MosaicML's Inference service. Start with our managed endpoints for models like MPT-7B-Instruct, and/or deploy your own custom model endpoints for optimal cost and data privacy.
What's Next?
This MPT-7B release is the culmination of two years of work at MosaicML building and battle-testing open-source software (Composer, StreamingDataset, LLM Foundry) and proprietary infrastructure (MosaicML Training and Inference) that makes it possible for customers to train LLMs on any compute provider, with any data source, with efficiency, privacy and cost transparency - and to have things go right the first time.
We believe MPT, the MosaicML LLM Foundry, and the MosaicML platform are the best starting point for building custom LLMs for private, commercial, and community use, whether you want to finetune our checkpoints or train your own from scratch. We look forward to seeing how the community builds on these tools and artifacts.
Importantly, today's MPT-7B models are just the beginning! To help our customers address more challenging tasks and continually improve their products, MosaicML will continue to produce foundation models of higher and higher quality. Exciting follow-on models are already training. Expect to hear more about them soon!
Acknowledgements
We are grateful to our friends at AI2 for helping us to curate our pretraining dataset, choose a great tokenizer, and for many other helpful conversations along the way ⚔️
Appendix
Data
mC4
Multilingual C4 (mC4) 3.1.0 is an update of the original mC4 by Chung et al., which contains sources through August 2022. We selected the English subset, and then applied the following filtering criteria to each document:
|
[
"https://www.mosaicml.com/blog/mpt-7b",
"https://stability.ai/blog/stability-ai-launches-the-first-of-its-stablelm-suite-of-language-models",
"https://lmsys.org/blog/2023-03-30-vicuna/"
] |
13
|
The most common character must be alphabetic.
≥ 92% of characters must be alphanumeric.
If the document is > 500 words, the most common word cannot constitute > 7.5% of the total word count; If the document is ≤ 500 words, the most common word cannot constitute > 30% of the total word count.
The document must be ≥ 200 words and ≤ 50000 words.
The first three filtering criteria were used to improve sample quality, and the final filtering criterion (documents must be ≥200 words and ≤50000 words) was used to increase the mean sequence length of the pretraining data.
mC4 was released as part of the continued effort from Dodge et al..
C4
Colossal Cleaned Common Crawl (C4) is an English Common Crawl corpus introduced by Raffel et al.. We applied Abbas et al.'s Semantic Deduplication process to remove the 20% most similar documents within C4, as internal experiments showed that this is a Pareto improvement for models trained on C4.
RedPajama
We included a number of subsets of the RedPajama dataset, which is Together's attempt to replicate LLaMA's training data. Specifically, we used the CommonCrawl, arXiv, Wikipedia, Books, and StackExchange subsets.
The Stack
We wanted our model to be capable of code generation, so we turned to The Stack, a 6.4TB corpus of code data. We used The Stack Dedup, a variant of the stack that has been approximately deduplicated (via MinHashLSH) to 2.9TB. We selected a subset of 18 of The Stack's 358 programming languages in order to reduce dataset size and increase relevance:
C
C-Sharp
C++
Common Lisp
F-Sharp
Fortran
Go
Haskell
Java
|
[
"https://www.mosaicml.com/blog/mpt-7b",
"https://stability.ai/blog/stability-ai-launches-the-first-of-its-stablelm-suite-of-language-models",
"https://lmsys.org/blog/2023-03-30-vicuna/"
] |
14
|
C++
Common Lisp
F-Sharp
Fortran
Go
Haskell
Java
Ocaml
Perl
Python
Ruby
Rust
Scala
Scheme
Shell
Tex
We chose to have code constitute 10% of the pretraining tokens, as internal experiments showed that we could train on up to 20% code (and 80% natural language) with no negative impact on natural language evaluation.
We also extracted the Markdown component of The Stack Dedup and treated this as an independent pretraining data subset (i.e. not counted towards the 10% code tokens). Our motivation for this is that markup language documents are predominantly natural language, and as such should count towards our natural language token budget.
Semantic Scholar ORC
The Semantic Scholar Open Research Corpus (S2ORC) is a corpus of English-language academic papers, which we consider to be a high-quality data source. The following quality filtering criteria were applied:
The paper is open access.
The paper has a title and abstract.
The paper is in English (as assessed using cld3).
The paper has at least 500 words and 5 paragraphs.
The paper was published after 1970 and before 2022-12-01.
The most frequent word in the paper consists of alpha characters only, and it appears in less than 7.5% of the document.
This yielded 9.9M papers. Instructions to obtain the latest dataset version are available here, and the original publication is here. The filtered version of the dataset was kindly provided to us by AI2.
Evaluation Tasks
Lambada: 5153 samples of text curated from the books corpus. Consists of a several hundred word paragraph in which the model is expected to predict the next word.
|
[
"https://www.mosaicml.com/blog/mpt-7b",
"https://stability.ai/blog/stability-ai-launches-the-first-of-its-stablelm-suite-of-language-models",
"https://lmsys.org/blog/2023-03-30-vicuna/"
] |
15
|
PIQA: 1838 samples of physical intuitive binary multiple choice questions, e.g. "Question: How can I easily carry clothes on hangers when I move?", "Answer: "Take a couple of empty heavy duty clothes hangers, then hook several hangers of clothes on Those hangers and carry them all at once."
COPA: 100 sentences of the form XYZ therefore/because TUV. Framed as binary multiple choice questions where the model has a choice of two possible ways to follow the therefore/because. e.g. {"query": "The woman was in a bad mood, therefore", "gold": 1, "choices": ["she engaged in small talk with her friend.", "she told her friend to leave her alone."]}
BoolQ: 3270 yes/no questions based on some passage which contains relevant information. Question topics range from pop culture to science, law, history, etc. e.g. {"query": "Passage: Kermit the Frog is a Muppet character and Jim Henson's most well-known creation. Introduced in 1955, Kermit serves as the straight man protagonist of numerous Muppet productions, most notably Sesame Street and The Muppet Show, as well as in other television series, films, specials, and public service announcements through the years. Henson originally performed Kermit until his death in 1990; Steve Whitmire performed Kermit from that time up until his dismissal from the role in 2016. Kermit is currently performed by Matt Vogel. He was also voiced by Frank Welker in Muppet Babies and occasionally in other animation projects, and is voiced by Matt Danner in the 2018 reboot of Muppet Babies.\nQuestion: has kermit the frog been on sesame street?\n", "choices": ["no", "yes"], "gold": 1}
Arc-Challenge: 1172 challenging four-choice multiple choice questions about science
|
[
"https://www.mosaicml.com/blog/mpt-7b",
"https://stability.ai/blog/stability-ai-launches-the-first-of-its-stablelm-suite-of-language-models",
"https://lmsys.org/blog/2023-03-30-vicuna/"
] |
16
|
Arc-Challenge: 1172 challenging four-choice multiple choice questions about science
Arc-Easy: 2376 easy four choice multiple choice science questions
HellaSwag: 10042 four choice multiple choice questions in which a real life scenario is presented and the model must choose the most likely conclusion to the scenario.
Jeopardy: 2117 Jeopardy questions from five categories: science, world history, us history, word origins, and literature. The model must provide the exact correct answer
MMLU: 14,042 multiple choice questions from 57 diverse academic categories
TriviaQA: 11313 free response pop culture trivia questions
Winograd: 273 schema questions where the model must resolve which referent of a pronoun is most likely.
Winogrande: 1,267 schema questions where the model must resolve which ambiguous sentence is more logically likely (both versions of the sentence are syntactically valid)
MPT Hugging Face Spaces Privacy Policy
Please see our MPT Hugging Face Spaces Privacy Policy.
Related posts
June 22, 2023MPT-30B: Raising the bar for open-source foundation models
July 18, 2023Announcing MPT-7B-8K: 8K Context Length for Document Understanding
March 9, 2023MosaicBERT: Pretraining BERT from Scratch for $20
Stability AI Launches the First of its Stable LM Suite of Language Models
Product
19 Apr
Written By
Guest User
“A Stochastic Parrot, flat design, vector art” — Stable Diffusion XL.
|
[
"https://www.mosaicml.com/blog/mpt-7b",
"https://stability.ai/blog/stability-ai-launches-the-first-of-its-stablelm-suite-of-language-models",
"https://lmsys.org/blog/2023-03-30-vicuna/"
] |
17
|
“A Stochastic Parrot, flat design, vector art” — Stable Diffusion XL.
Today, Stability AI released a new open source language model, Stable LM. The Alpha version of the model is available in 3 billion and 7 billion parameters, with 15 billion to 65 billion parameter models to follow. Developers can freely inspect, use, and adapt our Stable LM base models for commercial or research purposes, subject to the terms of the CC BY-SA-4.0 license.
In 2022, Stability AI drove the public release of Stable Diffusion, a revolutionary image model representing a transparent, open, and scalable alternative to proprietary AI. With the launch of the Stable LM suite of models, Stability AI is continuing to make foundational AI technology accessible to all. Our Stable LM models can generate text and code and will power various downstream applications. They demonstrate how small and efficient models can deliver high performance with appropriate training.
The release of Stable LM builds on our experience in open sourcing earlier language models with EleutherAI, a nonprofit research hub. These language models include GPT-J, GPT-NeoX, and the Pythia suite, trained on The Pile open source dataset. Many recent open source language models continue to build on these efforts, including Cerebras-GPT and Dolly-2.
Stable LM is trained on a new experimental dataset built on The Pile, but three times larger with 1.5 trillion content tokens. We will release details on the dataset in due course. The richness of this dataset gives Stable LM surprisingly high performance in conversational and coding tasks despite its small size of 3 to 7 billion parameters (by comparison, GPT-3 has 175 billion parameters).
|
[
"https://www.mosaicml.com/blog/mpt-7b",
"https://stability.ai/blog/stability-ai-launches-the-first-of-its-stablelm-suite-of-language-models",
"https://lmsys.org/blog/2023-03-30-vicuna/"
] |
18
|
We are also releasing a set of research models that are fine-tuned instructions. Initially, these fine-tuned models will use a combination of five recent open source datasets for conversational agents: Alpaca, GPT4All, Dolly, ShareGPT, and HH. These fine-tuned models are intended for research use only and are released under a noncommercial CC BY-NC-SA 4.0 license, in line with Stanford’s Alpaca license.
Check out some examples below, produced by our 7 billion parameters fine-tuned model:
View fullsize
View fullsize
View fullsize
Language models will form the backbone of our digital economy, and we want everyone to have a voice in their design. Models like Stable LM demonstrate our commitment to AI technology that is transparent, accessible, and supportive:
Transparent. We open source our models to promote transparency and foster trust. Researchers can “look under the hood” to verify performance, work on interpretability techniques, identify potential risks, and help develop safeguards. Organizations across the public and private sectors can adapt (“fine-tune”) these open source models for their own applications without sharing their sensitive data or giving up control of their AI capabilities.
Accessible. We design for the edge so everyday users can run our models on local devices. Using these models, developers can build independent applications compatible with widely available hardware instead of relying on proprietary services from one or two companies. In this way, the economic benefits of AI are shared by a broad community of users and developers. Open, fine-grained access to our models allows the broad research and academic community to develop interpretability and safety techniques beyond what is possible with closed models.
|
[
"https://www.mosaicml.com/blog/mpt-7b",
"https://stability.ai/blog/stability-ai-launches-the-first-of-its-stablelm-suite-of-language-models",
"https://lmsys.org/blog/2023-03-30-vicuna/"
] |
19
|
Supportive. We build models to support our users, not replace them. We are focused on efficient, specialized, and practical AI performance – not a quest for god-like intelligence. We develop tools that help everyday people and everyday firms use AI to unlock creativity, boost their productivity, and open up new economic opportunities.
The models are now available in our GitHub repository. We will publish a full technical report in the near future and look forward to ongoing collaboration with developers and researchers as we roll out the Stable LM suite. In addition, we will be kicking off our crowd-sourced RLHF program and working with community efforts such as Open Assistant to create an open source dataset for AI assistants.
We will be releasing more models soon and are growing our team. If you are passionate about democratizing access to this technology and experienced in LLMs, please apply
here!
Guest User
Previous
Previous
Stability AI releases its Image Upscaling API
Next
Next
Stability AI Partners with Iconic Artist Peter Gabriel to Launch Series of AI Animation Challenges titled #DiffuseTogether
LMSYS ORG
Projects
Blog
About
Donations
Chatbot Arena
Projects
Blog
About
Donations
Chatbot Arena
Vicuna: An Open-Source Chatbot Impressing GPT-4 with 90%* ChatGPT Quality
by: The Vicuna Team, Mar 30, 2023
|
[
"https://www.mosaicml.com/blog/mpt-7b",
"https://stability.ai/blog/stability-ai-launches-the-first-of-its-stablelm-suite-of-language-models",
"https://lmsys.org/blog/2023-03-30-vicuna/"
] |
20
|
by: The Vicuna Team, Mar 30, 2023
We introduce Vicuna-13B, an open-source chatbot trained by fine-tuning LLaMA on user-shared conversations collected from ShareGPT. Preliminary evaluation using GPT-4 as a judge shows Vicuna-13B achieves more than 90%* quality of OpenAI ChatGPT and Google Bard while outperforming other models like LLaMA and Stanford Alpaca in more than 90%* of cases. The cost of training Vicuna-13B is around $300. The code and weights, along with an online demo, are publicly available for non-commercial use.
Vicuna (generated by stable diffusion 2.1)
According to a fun and non-scientific evaluation with GPT-4. Further rigorous evaluation is needed.
How Good is Vicuna?
After fine-tuning Vicuna with 70K user-shared ChatGPT conversations, we discover that Vicuna becomes capable of generating more detailed and well-structured answers compared to Alpaca (see examples below), with the quality on par with ChatGPT.
However, evaluating chatbots is never a simple task.
With recent advancements in GPT-4, we are curious whether its capabilities have reached a human-like level that could enable an automated evaluation framework for benchmark generation and performance assessments.
Our initial finding indicates that GPT-4 can produce highly consistent ranks and detailed assessment when comparing chatbots’ answers (see above example of GPT-4 judgment).
Preliminary evaluations based on GPT-4, summarized in Figure 1, show that Vicuna achieves 90%* capability of Bard/ChatGPT.
While this proposed framework shows a potential to automate chatbot assessment, it is not yet a rigorous approach.
Building an evaluation system for chatbots remains an open question requiring further research. More details are provided in the evaluation section.
|
[
"https://www.mosaicml.com/blog/mpt-7b",
"https://stability.ai/blog/stability-ai-launches-the-first-of-its-stablelm-suite-of-language-models",
"https://lmsys.org/blog/2023-03-30-vicuna/"
] |
21
|
Figure 1. Relative Response Quality Assessed by GPT-4*
Online Demo
Try the Vicuna-13B demo here!
Overview
The rapid advancement of large language models (LLMs) has revolutionized chatbot systems, resulting in unprecedented levels of intelligence as seen in OpenAI's ChatGPT. However, despite its impressive performance, the training and architecture details of ChatGPT remain unclear, hindering research and open-source innovation in this field. Inspired by the Meta LLaMA and Stanford Alpaca project, we introduce Vicuna-13B, an open-source chatbot backed by an enhanced dataset and an easy-to-use, scalable infrastructure. By fine-tuning a LLaMA base model on user-shared conversations collected from ShareGPT.com, Vicuna-13B has demonstrated competitive performance compared to other open-source models like Stanford Alpaca. This blog post provides a preliminary evaluation of Vicuna-13B's performance and describes its training and serving infrastructure. We also invite the community to interact with our online demo to test the capabilities of this chatbot.
Figure 2. Workflow Overview
|
[
"https://www.mosaicml.com/blog/mpt-7b",
"https://stability.ai/blog/stability-ai-launches-the-first-of-its-stablelm-suite-of-language-models",
"https://lmsys.org/blog/2023-03-30-vicuna/"
] |
22
|
Figure 2. Workflow Overview
Figure 2 provides an overview of our work. To begin, we collected around 70K conversations from ShareGPT.com, a website where users can share their ChatGPT conversations. Next, we enhanced the training scripts provided by Alpaca to better handle multi-turn conversations and long sequences. The training was done with PyTorch FSDP on 8 A100 GPUs in one day. For serving the demo, we implemented a lightweight distributed serving system. We conducted a preliminary evaluation of the model quality by creating a set of 80 diverse questions and utilizing GPT-4 to judge the model outputs. To compare two different models, we combine the outputs from each model into a single prompt for each question. The prompts are then sent to GPT-4, which assesses which model provides better responses. A detailed comparison of LLaMA, Alpaca, ChatGPT, and Vicuna is shown in Table 1 below.
Table 1. Comparison between several notable models
Model Name LLaMA Alpaca Vicuna Bard/ChatGPT Dataset Publicly available datasets (1T token) Self-instruct from davinci-003 API (52K samples) User-shared conversations (70K samples) N/A Training code N/A Available Available N/A Evaluation metrics Academic benchmark Author evaluation GPT-4 assessment Mixed Training cost (7B) 82K GPU-hours $500 (data) + $100 (training) $140 (training) N/A Training cost (13B) 135K GPU-hours N/A $300 (training) N/A
Training
|
[
"https://www.mosaicml.com/blog/mpt-7b",
"https://stability.ai/blog/stability-ai-launches-the-first-of-its-stablelm-suite-of-language-models",
"https://lmsys.org/blog/2023-03-30-vicuna/"
] |
23
|
Training
Vicuna is created by fine-tuning a LLaMA base model using approximately 70K user-shared conversations gathered from ShareGPT.com with public APIs. To ensure data quality, we convert the HTML back to markdown and filter out some inappropriate or low-quality samples. Additionally, we divide lengthy conversations into smaller segments that fit the model's maximum context length.
Our training recipe builds on top of Stanford’s alpaca with the following improvements.
Multi-turn conversations: We adjust the training loss to account for multi-turn conversations and compute the fine-tuning loss solely on the chatbot's output.
Memory Optimizations: To enable Vicuna's understanding of long context, we expand the max context length from 512 in alpaca to 2048, which substantially increases GPU memory requirements. We tackle the memory pressure by utilizing gradient checkpointing and flash attention.
Cost Reduction via Spot Instance: The 40x larger dataset and 4x sequence length for training poses a considerable challenge in training expenses. We employ SkyPilot managed spot to reduce the cost by leveraging the cheaper spot instances with auto-recovery for preemptions and auto zone switch. This solution slashes costs for training the 7B model from $500 to around $140 and the 13B model from around $1K to $300.
Serving
We build a serving system that is capable of serving multiple models with distributed workers. It supports flexible plug-in of GPU workers from both on-premise clusters and the cloud. By utilizing a fault-tolerant controller and managed spot feature in SkyPilot, this serving system can work well with cheaper spot instances from multiple clouds to reduce the serving costs. It is currently a lightweight implementation and we are working on integrating more of our latest research into it.
How To Evaluate a Chatbot?
|
[
"https://www.mosaicml.com/blog/mpt-7b",
"https://stability.ai/blog/stability-ai-launches-the-first-of-its-stablelm-suite-of-language-models",
"https://lmsys.org/blog/2023-03-30-vicuna/"
] |
24
|
How To Evaluate a Chatbot?
Evaluating AI chatbots is a challenging task, as it requires examining language understanding, reasoning, and context awareness. With AI chatbots becoming more advanced, current open benchmarks may no longer suffice. For instance, the evaluation dataset used in Stanford’s Alpaca, self-instruct, can be effectively answered by SOTA chatbots, making it difficult for humans to discern differences in performance. More limitations include training/test data contamination and the potentially high cost of creating new benchmarks. To tackle these issues, we propose an evaluation framework based on GPT-4 to automate chatbot performance assessment.
First, we devised eight question categories, such as Fermi problems, roleplay scenarios, and coding/math tasks, to test various aspects of a chatbot's performance. Through careful prompt engineering, GPT-4 is able to generate diverse, challenging questions that baseline models struggle with. We select ten questions per category and collect answers from five chatbots: LLaMA, Alpaca, ChatGPT, Bard, and Vicuna. We then ask GPT-4 to rate the quality of their answers based on helpfulness, relevance, accuracy, and detail. We discover that GPT-4 can produce not only relatively consistent scores but also detailed explanations on why such scores are given (detailed examples link). However, we also notice that GPT-4 is not very good at judging coding/math tasks.
Figure 3. Response Comparison Assessed by GPT-4
|
[
"https://www.mosaicml.com/blog/mpt-7b",
"https://stability.ai/blog/stability-ai-launches-the-first-of-its-stablelm-suite-of-language-models",
"https://lmsys.org/blog/2023-03-30-vicuna/"
] |
25
|
Figure 3. Response Comparison Assessed by GPT-4
Figure 3 displays the comparison results between all baselines and Vicuna. GPT-4 prefers Vicuna over state-of-the-art open-source models (LLaMA, Alpaca) in more than 90% of the questions, and it achieves competitive performance against proprietary models (ChatGPT, Bard). In 45% of the questions, GPT-4 rates Vicuna's response as better or equal to ChatGPT's.
As GPT-4 assigns a quantitative score to each response on a scale of 10, we calculate the total score for each (baseline, Vicuna) comparison pair by adding up the scores obtained by each model on 80 questions. As shown in Table 2, Vicuna’s total score is 92% of ChatGPT’s. Despite recent advancements, these chatbots still face limitations, such as struggling with basic math problems or having limited coding ability.
Table 2. Total Scores Assessed by GPT-4.
Baseline Baseline Score Vicuna Score LLaMA-13B 513.0 694.0 Alpaca-13B 583.0 704.0 Bard 664.0 655.5 ChatGPT 693.0 638.0
While this proposed evaluation framework demonstrates the potential for assessing chatbots, it is not yet a rigorous or mature approach, as large language models are prone to hallucinate. Developing a comprehensive, standardized evaluation system for chatbots remains an open question requiring further research.
Edited: After this blog post, we conducted a deeper study on this GPT4-based evaluation approach. You are welcome to read our new Judging LLM-as-a-judge paper and try the new evaluation tool.
Limitations
|
[
"https://www.mosaicml.com/blog/mpt-7b",
"https://stability.ai/blog/stability-ai-launches-the-first-of-its-stablelm-suite-of-language-models",
"https://lmsys.org/blog/2023-03-30-vicuna/"
] |
26
|
Limitations
We have noticed that, similar to other large language models, Vicuna has certain limitations. For instance, it is not good at tasks involving reasoning or mathematics, and it may have limitations in accurately identifying itself or ensuring the factual accuracy of its outputs. Additionally, it has not been sufficiently optimized to guarantee safety or mitigate potential toxicity or bias. To address the safety concerns, we use the OpenAI moderation API to filter out inappropriate user inputs in our online demo. Nonetheless, we anticipate that Vicuna can serve as an open starting point for future research to tackle these limitations.
Release
In our first release, we will share the training, serving, and evaluation code on a GitHub repo: https://github.com/lm-sys/FastChat.
We also released the Vicuna-13B model weights.
There is no plan to release the dataset. Join our Discord server and follow our Twitter to get the latest updates.
License
The online demo is a research preview intended for non-commercial use only, subject to the model License of LLaMA, Terms of Use of the data generated by OpenAI, and Privacy Practices of ShareGPT. Please contact us If you find any potential violation.
The code is released under the Apache License 2.0.
Acknowledgment
We would like to thank Xinyang Geng, Hao Liu, and Eric Wallace from BAIR; Xuecheng Li, and Tianyi Zhang from Stanford Alpaca team for their insightful discussion and feedback; Qirong Ho from MBZUAI for providing support on the serving cluster. Please check out a blog post from BAIR about a concurrent effort on their chatbot, Koala.
The Team
This is a joint effort with collaborators from multiple institutions, including UC Berkeley, CMU, Stanford, UC San Diego, and MBZUAI.
|
[
"https://www.mosaicml.com/blog/mpt-7b",
"https://stability.ai/blog/stability-ai-launches-the-first-of-its-stablelm-suite-of-language-models",
"https://lmsys.org/blog/2023-03-30-vicuna/"
] |
27
|
Students (alphabetical order): Wei-Lin Chiang, Zhuohan Li, Zi Lin, Ying Sheng, Zhanghao Wu, Hao Zhang (✉), Lianmin Zheng (✉), Siyuan Zhuang, Yonghao Zhuang
Advisors (alphabetical order): Joseph E. Gonzalez, Ion Stoica, Eric P. Xing
✉ Correspondence to: Lianmin Zheng (lianminzheng@gmail.com), Hao Zhang (sjtu.haozhang@gmail.com), or LMSYS (lmsys.org@gmail.com).
Citation
After this blog post, we extended our idea of GPT-4 based evaluation and wrote a more formal paper that systematically studies this "LLM-as-a-judge" approach.
You are welcome to read and cite this paper:
Judging LLM-as-a-judge with MT-Bench and Chatbot Arena.
|
[
"https://www.mosaicml.com/blog/mpt-7b",
"https://stability.ai/blog/stability-ai-launches-the-first-of-its-stablelm-suite-of-language-models",
"https://lmsys.org/blog/2023-03-30-vicuna/"
] |
Subsets and Splits
No community queries yet
The top public SQL queries from the community will appear here once available.