kartikmosaicml commited on
Commit
30f1d55
1 Parent(s): 0d87f19

Initial commit for readme and model card

Browse files

Creating a draft readme. Need to update it further.

Files changed (1) hide show
  1. README.md +244 -0
README.md ADDED
@@ -0,0 +1,244 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ ---
2
+ license: apache-2.0
3
+ tags:
4
+ - Composer
5
+ - MosaicML
6
+ - llm-foundry
7
+ - StreamingDatasets
8
+ datasets:
9
+ - mc4
10
+ - c4
11
+ - togethercomputer/RedPajama-Data-1T
12
+ - bigcode/the-stack
13
+ - allenai/s2orc
14
+ - TODO
15
+ inference: false
16
+ ---
17
+
18
+ # MPT-30B
19
+
20
+ MPT-7B is a decoder-style transformer pretrained from scratch on 1T tokens of English text and code.
21
+ This model was trained by [MosaicML](https://www.mosaicml.com).
22
+
23
+ MPT-30B is part of the family of MosaicPretrainedTransformer (MPT) models, which use a modified transformer architecture optimized for efficient training and inference.
24
+
25
+ These architectural changes include performance-optimized layer implementations and the elimination of context length limits by replacing
26
+ positional embeddings with Attention with Linear Biases ([ALiBi](https://arxiv.org/abs/2108.12409)).
27
+ Thanks to these modifications, MPT models can be trained with high throughput efficiency and stable convergence.
28
+ MPT models can also be served efficiently with both standard HuggingFace pipelines and NVIDIA's [FasterTransformer](https://github.com/NVIDIA/FasterTransformer).
29
+
30
+ This model uses the MosaicML LLM codebase, which can be found in the [llm-foundry repository](https://github.com/mosaicml/llm-foundry). It was trained by MosaicML’s NLP team on the [MosaicML platform](https://www.mosaicml.com/training) for LLM pretraining, finetuning, and inference.
31
+
32
+ ### How is this model different?
33
+
34
+ MPT-30B is
35
+
36
+ * **Licensed for the possibility of commercial use** (unlike [LLaMA](https://arxiv.org/abs/2302.13971)).
37
+ * **Trained on a large amount of data** (1T tokens like [LLaMA](https://arxiv.org/abs/2302.13971) vs. 300B for [Pythia](https://github.com/EleutherAI/pythia), 300B for [OpenLLaMA](https://github.com/openlm-research/open_llama), and 800B for [StableLM](https://github.com/Stability-AI/StableLM)).
38
+ * **Prepared to handle extremely long inputs** thanks to [ALiBi](https://arxiv.org/abs/2108.12409) (TODO: talk about MPT-30B-instruct finetuned on 8k).
39
+ * **Capable of fast training and inference** (via [FlashAttention](https://arxiv.org/pdf/2205.14135.pdf) and [FasterTransformer](https://github.com/NVIDIA/FasterTransformer))
40
+ * **Equipped with highly efficient open-source training code** via the [llm-foundry repository](https://github.com/mosaicml/llm-foundry)
41
+
42
+ ### Models finetuned off MPT-30B:
43
+
44
+ The following models are finetuned on MPT-30B:
45
+
46
+ * [MPT-30B-Instruct](https://huggingface.co/mosaicml/mpt-30b-instruct): a model for short-form instruction following.
47
+ Built by finetuning MPT-30B on several carefully curated datasets.
48
+ * [Demo on Hugging Face Spaces](https://huggingface.co/spaces/mosaicml/mpt-30b-instruct) (TODO: will this be a thing?)
49
+
50
+ * [MPT-30B-Chat](https://huggingface.co/mosaicml/mpt-30b-chat): a chatbot-like model for dialogue generation.
51
+ Built by finetuning MPT-30B on [ShareGPT-Vicuna](https://huggingface.co/datasets/anon8231489123/ShareGPT_Vicuna_unfiltered), [Camel-AI](https://huggingface.co/camel-ai),
52
+ [GPTeacher](https://github.com/teknium1/GPTeacher), [Guanaco](https://huggingface.co/datasets/timdettmers/openassistant-guanaco), [Baize](https://github.com/project-baize/baize-chatbot) and some generated datasets.
53
+ * License: _CC-By-NC-SA-4.0_
54
+ * [Demo on Hugging Face Spaces](https://huggingface.co/spaces/mosaicml/mpt-30b-chat)
55
+
56
+ ## Model Date
57
+
58
+ TBD
59
+
60
+ ## Model License
61
+
62
+ Apache-2.0
63
+
64
+ ## Documentation
65
+
66
+ * [Blog post: Introducing MPT-30B: TBD: A New Standard for Open-Source, Commercially Usable LLMs](https://www.mosaicml.com/blog/mpt-30b)
67
+ * [Codebase (mosaicml/llm-foundry repo)](https://github.com/mosaicml/llm-foundry/)
68
+ * Questions: Feel free to contact us via the [MosaicML Community Slack](https://mosaicml.me/slack)!
69
+
70
+
71
+ ## How to Use
72
+
73
+ This model is best used with the MosaicML [llm-foundry repository](https://github.com/mosaicml/llm-foundry) for training and finetuning.
74
+
75
+ ```python
76
+ import transformers
77
+ model = transformers.AutoModelForCausalLM.from_pretrained(
78
+ 'mosaicml/mpt-30b',
79
+ trust_remote_code=True
80
+ )
81
+ ```
82
+ Note: This model requires that `trust_remote_code=True` be passed to the `from_pretrained` method.
83
+ This is because we use a custom `MPT` model architecture that is not yet part of the Hugging Face `transformers` package.
84
+ `MPT` includes options for many training efficiency features such as [FlashAttention](https://arxiv.org/pdf/2205.14135.pdf), [ALiBi](https://arxiv.org/abs/2108.12409), [QK LayerNorm](https://arxiv.org/abs/2010.04245), and more.
85
+
86
+ To use the optimized [triton implementation](https://github.com/openai/triton) of FlashAttention, you can load the model on GPU (`cuda:0`) with `attn_impl='triton'` and with `bfloat16` precision:
87
+ ```python
88
+ import torch
89
+ import transformers
90
+
91
+ name = 'mosaicml/mpt-30b'
92
+
93
+ config = transformers.AutoConfig.from_pretrained(name, trust_remote_code=True)
94
+ config.attn_config['attn_impl'] = 'triton'
95
+ config.init_device = 'cuda:0' # For fast initialization directly on GPU!
96
+
97
+ model = transformers.AutoModelForCausalLM.from_pretrained(
98
+ name,
99
+ config=config,
100
+ torch_dtype=torch.bfloat16, # Load model weights in bfloat16
101
+ trust_remote_code=True
102
+ )
103
+ ```
104
+
105
+ Although the model was trained with a sequence length of 8192, ALiBi enables users to increase the maximum sequence length during finetuning and/or inference. For example:
106
+ TODO: check if we want to advertise this.
107
+ ```python
108
+ import transformers
109
+
110
+ name = 'mosaicml/mpt-30b'
111
+
112
+ config = transformers.AutoConfig.from_pretrained(name, trust_remote_code=True)
113
+ config.max_seq_len = 16384 # (input + output) tokens can now be up to 16384
114
+
115
+ model = transformers.AutoModelForCausalLM.from_pretrained(
116
+ name,
117
+ config=config,
118
+ trust_remote_code=True
119
+ )
120
+ ```
121
+
122
+ This model was trained with the [EleutherAI/gpt-neox-20b](https://huggingface.co/EleutherAI/gpt-neox-20b) tokenizer.
123
+
124
+ ```python
125
+ from transformers import AutoTokenizer
126
+ tokenizer = AutoTokenizer.from_pretrained('EleutherAI/gpt-neox-20b')
127
+ ```
128
+
129
+ The model can then be used, for example, within a text-generation pipeline.
130
+ Note: when running Torch modules in lower precision, it is best practice to use the [torch.autocast context manager](https://pytorch.org/docs/stable/amp.html).
131
+
132
+ ```python
133
+ from transformers import pipeline
134
+
135
+ pipe = pipeline('text-generation', model=model, tokenizer=tokenizer, device='cuda:0')
136
+
137
+ with torch.autocast('cuda', dtype=torch.bfloat16):
138
+ print(
139
+ pipe('Here is a recipe for vegan banana bread:\n',
140
+ max_new_tokens=100,
141
+ do_sample=True,
142
+ use_cache=True))
143
+ ```
144
+
145
+ ## Model Description
146
+
147
+ The architecture is a modification of a standard decoder-only transformer.
148
+
149
+ The model has been modified from a standard transformer in the following ways:
150
+ * It uses [FlashAttention](https://arxiv.org/pdf/2205.14135.pdf)
151
+ * It uses [ALiBi (Attention with Linear Biases)](https://arxiv.org/abs/2108.12409) and does not use positional embeddings
152
+ * It does not use biases
153
+
154
+
155
+ | Hyperparameter | Value |
156
+ |----------------|-------|
157
+ |n_parameters | 29.95B |
158
+ |n_layers | 48 |
159
+ | n_heads | 64 |
160
+ | d_model | 7168 |
161
+ | vocab size | 50432 |
162
+ | sequence length | 8192 |
163
+
164
+
165
+
166
+ ## Training Data
167
+
168
+ ### Streaming Datasets
169
+
170
+ Data was formatted using the MosaicML [StreamingDataset](https://github.com/mosaicml/streaming) library to host our data in object storage and efficiently stream it to our compute cluster during training.
171
+ StreamingDataset obviates the need to download the whole dataset before starting training, and allows instant resumption of training from any point in the dataset.
172
+
173
+
174
+ ### Data Mix
175
+
176
+ The model was trained for 1T tokens (with batch size TBD and sequence length TBD). It was trained on the following data mix:
177
+
178
+ (TODO: update this)
179
+ | Data Source | Number of Tokens in Source | Proportion | Effective Number of Tokens | Epochs |
180
+ |-------------|----------------------------|------------|----------------------------|--------|
181
+ | mC4 3.1.0 - English | 417.99 B | 0.33 | 330 B | 0.14 |
182
+ | C4 - English - SemDedup 80% | 100.42 B | 0.299 | 299 B | 2.98 |
183
+ | RedPajama - CommonCrawl | 878.45 B | 0.1 | 100 B | 0.11 |
184
+ | The Stack - Selected Languages | 463.78 B | 0.1 | 100 B | 0.22 |
185
+ | RedPajama - Wikipedia - En | 4.87 B | 0.04 | 40 B | 8.21 |
186
+ | The Stack - Markdown | 107.07 B | 0.035 | 35 B | 0.33 |
187
+ | S2ORC | 48.85 B | 0.033 | 33 B | 0.68 |
188
+ | RedPajama - Books | 26.02 B | 0.03 | 30B | 1.15 |
189
+ | RedPajama - arXiv | 28.10 B | 0.019 | 19 B | 0.68 |
190
+ | RedPajama - StackExchange | 20.54 B | 0.014 | 14 B |0.68 |
191
+
192
+ Samples for each batch were selected from one of the datasets with the probability specified above.
193
+ (TODO: check with @sam whether only FT was on 8k)
194
+ 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 8192 sequence length.
195
+
196
+ The data was tokenized using the [EleutherAI/gpt-neox-20b](https://huggingface.co/EleutherAI/gpt-neox-20b) tokenizer. This BPE tokenizer has a number of desirable characteristics,
197
+ most of which are relevant for tokenizing code:
198
+ (1) It was trained on a diverse mix of data that includes code (The Pile)
199
+ (2) It applies consistent space delimitation, unlike the GPT2 tokenizer which tokenizes inconsistently depending on the presence of prefix spaces
200
+ (3) It contains tokens for repeated space characters, which allows superior compression of text with large amounts of repeated space characters.
201
+
202
+ The model vocabulary size of 50432 was set to be a multiple of 128 (as in [MEGATRON-LM](https://arxiv.org/abs/1909.08053)), model flop utilization (MFU) increased by up to four percentage points.
203
+
204
+ ### Training Configuration
205
+
206
+ (TODO: get these details from @abhi and the others)
207
+ This model was trained on 440 A100-40GBs for about 9.5 days using the [MosaicML Platform](https://www.mosaicml.com/platform).
208
+ The model was trained with sharded data parallelism using [FSDP](https://pytorch.org/docs/stable/fsdp.html) and used the [LION](https://arxiv.org/abs/2302.06675) optimizer.
209
+
210
+ ## Limitations and Biases
211
+
212
+ _The following language is modified from [EleutherAI's GPT-NeoX-20B](https://huggingface.co/EleutherAI/gpt-neox-20b)_
213
+
214
+ MPT-30B (Base) is **not** intended for deployment without finetuning.
215
+ It should not be used for human-facing interactions without further guardrails and user consent.
216
+
217
+ MPT-30B can produce factually incorrect output, and should not be relied on to produce factually accurate information.
218
+ MPT-30B was trained on various public datasets.
219
+ 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.
220
+
221
+
222
+ ## MosaicML Platform
223
+
224
+ If you're interested in [training](https://www.mosaicml.com/training) and [deploying](https://www.mosaicml.com/inference) your own MPT or LLMs on the MosaicML Platform, [sign up here](https://forms.mosaicml.com/demo?utm_source=huggingface&utm_medium=referral&utm_campaign=mpt-30b).
225
+
226
+ ## Disclaimer
227
+
228
+ 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.
229
+
230
+ ## Citation
231
+
232
+ Please cite this model using the following format:
233
+
234
+ ```
235
+ @online{MosaicML2023Introducing,
236
+ author = {MosaicML NLP Team},
237
+ title = {Introducing MPT-30B: A New Standard for Open-Source,
238
+ ly Usable LLMs},
239
+ year = {2023},
240
+ url = {www.mosaicml.com/blog/mpt-30b},
241
+ note = {Accessed: TBD}, % change this date
242
+ urldate = {TBD} % change this date
243
+ }
244
+ ```