avi-skowron commited on
Commit
555c484
1 Parent(s): 539109a

updated model card

Browse files
Files changed (1) hide show
  1. README.md +269 -0
README.md ADDED
@@ -0,0 +1,269 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ ---
2
+ language:
3
+ - en
4
+ tags:
5
+ - pytorch
6
+ - causal-lm
7
+ - pythia
8
+ license: apache-2.0
9
+ datasets:
10
+ - the_pile
11
+ ---
12
+
13
+ The *Pythia Scaling Suite* is a collection of models developed to facilitate
14
+ interpretability research. It contains two sets of eight models of sizes
15
+ 70M, 160M, 410M, 1B, 1.4B, 2.8B, 6.9B, and 12B. For each size, there are two
16
+ models: one trained on the Pile, and one trained on the Pile after the dataset
17
+ has been globally deduplicated. All 8 model sizes are trained on the exact
18
+ same data, in the exact same order. All Pythia models are available
19
+ [on Hugging Face](https://huggingface.co/models?other=pythia).
20
+
21
+ The Pythia model suite was deliberately designed to promote scientific
22
+ research on large language models, especially interpretability research.
23
+ Despite not centering downstream performance as a design goal, we find the
24
+ models <a href="#evaluations">match or exceed</a> the performance of
25
+ similar and same-sized models, such as those in the OPT and GPT-Neo suites.
26
+
27
+ Previously, we released an early version of the Pythia suite to the public.
28
+ However, we decided to retrain the model suite to address a few hyperparameter
29
+ discrepancies. This model card <a href="#changelog">lists the changes</a>;
30
+ see appendix B in the Pythia paper for further discussion. We found no
31
+ difference in benchmark performance between the two Pythia versions.
32
+ The old Pythia models are
33
+ [still available](https://huggingface.co/models?other=pythia_v0); we suggest
34
+ using the retrained suite if you are just starting to use Pythia.<br>
35
+ **This is the current release.**
36
+
37
+ Please note that all models in the *Pythia* suite were renamed in January
38
+ 2023. For clarity, a <a href="#naming-convention-and-parameter-count">table
39
+ comparing the old and new names</a> is provided in this model card, together
40
+ with exact parameter counts.
41
+
42
+ # Pythia-12B
43
+
44
+ ## Model Details
45
+
46
+ - Developed by: [EleutherAI](http://eleuther.ai)
47
+ - Model type: Transformer-based Language Model
48
+ - Language: English
49
+ - Learn more: [Pythia's GitHub repository](https://github.com/EleutherAI/pythia)
50
+ for training procedure, config files, and details on how to use.
51
+ - Library: [GPT-NeoX](https://github.com/EleutherAI/gpt-neox)
52
+ - License: Apache 2.0
53
+ - Contact: to ask questions about this model, join the [EleutherAI
54
+ Discord](https://discord.gg/zBGx3azzUn), and post them in `#release-discussion`.
55
+ Please read the existing *Pythia* documentation before asking about it in the
56
+ EleutherAI Discord. For general correspondence: [contact@eleuther.
57
+ ai](mailto:contact@eleuther.ai).
58
+
59
+ <figure>
60
+
61
+ | Pythia model | Non-Embedding Params | Layers | Model Dim | Heads | Batch Size | Learning Rate | Equivalent Models |
62
+ | -----------: | -------------------: | :----: | :-------: | :---: | :--------: | :-------------------: | :--------------------: |
63
+ | 70M | 18,915,328 | 6 | 512 | 8 | 2M | 1.0 x 10<sup>-3</sup> | — |
64
+ | 160M | 85,056,000 | 12 | 768 | 12 | 4M | 6.0 x 10<sup>-4</sup> | GPT-Neo 125M, OPT-125M |
65
+ | 410M | 302,311,424 | 24 | 1024 | 16 | 4M | 3.0 x 10<sup>-4</sup> | OPT-350M |
66
+ | 1.0B | 805,736,448 | 16 | 2048 | 8 | 2M | 3.0 x 10<sup>-4</sup> | — |
67
+ | 1.4B | 1,208,602,624 | 24 | 2048 | 16 | 4M | 2.0 x 10<sup>-4</sup> | GPT-Neo 1.3B, OPT-1.3B |
68
+ | 2.8B | 2,517,652,480 | 32 | 2560 | 32 | 2M | 1.6 x 10<sup>-4</sup> | GPT-Neo 2.7B, OPT-2.7B |
69
+ | 6.9B | 6,444,163,072 | 32 | 4096 | 32 | 2M | 1.2 x 10<sup>-4</sup> | OPT-6.7B |
70
+ | 12B | 11,327,027,200 | 36 | 5120 | 40 | 2M | 1.2 x 10<sup>-4</sup> | — |
71
+ <figcaption>Engineering details for the <i>Pythia Suite</i>. Deduped and
72
+ non-deduped models of a given size have the same hyperparameters. “Equivalent”
73
+ models have <b>exactly</b> the same architecture, and the same number of
74
+ non-embedding parameters.</figcaption>
75
+ </figure>
76
+
77
+ ## Uses and Limitations
78
+
79
+ ### Intended Use
80
+
81
+ The primary intended use of Pythia is research on the behavior, functionality,
82
+ and limitations of large language models. This suite is intended to provide
83
+ a controlled setting for performing scientific experiments. To enable the
84
+ study of how language models change over the course of training, we provide
85
+ 143 evenly spaced intermediate checkpoints per model. These checkpoints are
86
+ hosted on Hugging Face as branches. Note that branch `143000` corresponds
87
+ exactly to the model checkpoint on the `main` branch of each model.
88
+
89
+ You may also further fine-tune and adapt Pythia-12B for deployment,
90
+ as long as your use is in accordance with the Apache 2.0 license. Pythia
91
+ models work with the Hugging Face [Transformers
92
+ Library](https://huggingface.co/docs/transformers/index). If you decide to use
93
+ pre-trained Pythia-12B as a basis for your fine-tuned model, please
94
+ conduct your own risk and bias assessment.
95
+
96
+ ### Out-of-scope use
97
+
98
+ The Pythia Suite is **not** intended for deployment. It is not a in itself
99
+ a product and cannot be used for human-facing interactions. For example,
100
+ the model may generate harmful or offensive text. Please evaluate the risks
101
+ associated with your particular use case.
102
+
103
+ Pythia models are English-language only, and are not suitable for translation
104
+ or generating text in other languages.
105
+
106
+ Pythia-12B has not been fine-tuned for downstream contexts in which
107
+ language models are commonly deployed, such as writing genre prose,
108
+ or commercial chatbots. This means Pythia-12B will **not**
109
+ respond to a given prompt the way a product like ChatGPT does. This is because,
110
+ unlike this model, ChatGPT was fine-tuned using methods such as Reinforcement
111
+ Learning from Human Feedback (RLHF) to better “understand” human instructions.
112
+
113
+ ### Limitations and biases
114
+
115
+ The core functionality of a large language model is to take a string of text
116
+ and predict the next token. The token used by the model need not produce the
117
+ most “accurate” text. Never rely on Pythia-12B to produce factually accurate
118
+ output.
119
+
120
+ This model was trained on [the Pile](https://pile.eleuther.ai/), a dataset
121
+ known to contain profanity and texts that are lewd or otherwise offensive.
122
+ See [Section 6 of the Pile paper](https://arxiv.org/abs/2101.00027) for a
123
+ discussion of documented biases with regards to gender, religion, and race.
124
+ Pythia-12B may produce socially unacceptable or undesirable text, *even if*
125
+ the prompt itself does not include anything explicitly offensive.
126
+
127
+ If you plan on using text generated through, for example, the Hosted Inference
128
+ API, we recommend having a human curate the outputs of this language model
129
+ before presenting it to other people. Please inform your audience that the
130
+ text was generated by Pythia-12B.
131
+
132
+ ### Quickstart
133
+
134
+ Pythia models can be loaded and used via the following code, demonstrated here
135
+ for the third `pythia-70m-deduped` checkpoint:
136
+
137
+ ```python
138
+ from transformers import GPTNeoXForCausalLM, AutoTokenizer
139
+
140
+ model = GPTNeoXForCausalLM.from_pretrained(
141
+ "EleutherAI/pythia-70m-deduped",
142
+ revision="step3000",
143
+ cache_dir="./pythia-70m-deduped/step3000",
144
+ )
145
+
146
+ tokenizer = AutoTokenizer.from_pretrained(
147
+ "EleutherAI/pythia-70m-deduped",
148
+ revision="step3000",
149
+ cache_dir="./pythia-70m-deduped/step3000",
150
+ )
151
+
152
+ inputs = tokenizer("Hello, I am", return_tensors="pt")
153
+ tokens = model.generate(**inputs)
154
+ tokenizer.decode(tokens[0])
155
+ ```
156
+
157
+ Revision/branch `step143000` corresponds exactly to the model checkpoint on
158
+ the `main` branch of each model.<br>
159
+ For more information on how to use all Pythia models, see [documentation on
160
+ GitHub](https://github.com/EleutherAI/pythia).
161
+
162
+ ## Training
163
+
164
+ ### Training data
165
+
166
+ [The Pile](https://pile.eleuther.ai/) is a 825GiB general-purpose dataset in
167
+ English. It was created by EleutherAI specifically for training large language
168
+ models. It contains texts from 22 diverse sources, roughly broken down into
169
+ five categories: academic writing (e.g. arXiv), internet (e.g. CommonCrawl),
170
+ prose (e.g. Project Gutenberg), dialogue (e.g. YouTube subtitles), and
171
+ miscellaneous (e.g. GitHub, Enron Emails). See [the Pile
172
+ paper](https://arxiv.org/abs/2101.00027) for a breakdown of all data sources,
173
+ methodology, and a discussion of ethical implications. Consult [the
174
+ datasheet](https://arxiv.org/abs/2201.07311) for more detailed documentation
175
+ about the Pile and its component datasets. The Pile can be downloaded from
176
+ the [official website](https://pile.eleuther.ai/), or from a [community
177
+ mirror](https://the-eye.eu/public/AI/pile/).<br>
178
+ The Pile was **not** deduplicated before being used to train Pythia-12B.
179
+
180
+ ### Training procedure
181
+
182
+ All models were trained on the exact same data, in the exact same order. Each
183
+ model saw 299,892,736,000 tokens during training, and 143 checkpoints for each
184
+ model are saved every 2,097,152,000 tokens, spaced evenly throughout training.
185
+ This corresponds to training for just under 1 epoch on the Pile for
186
+ non-deduplicated models, and about 1.5 epochs on the deduplicated Pile.
187
+
188
+ All *Pythia* models trained for 143000 steps at a batch size
189
+ of 2M (2,097,152 tokens).<br>
190
+ See [GitHub](https://github.com/EleutherAI/pythia) for more details on training
191
+ procedure, including [how to reproduce
192
+ it](https://github.com/EleutherAI/pythia/blob/main/README.md#reproducing-training).<br>
193
+ Pythia uses the same tokenizer as [GPT-NeoX-
194
+ 20B](https://huggingface.co/EleutherAI/gpt-neox-20b).
195
+
196
+ ## Evaluations
197
+
198
+ All 16 *Pythia* models were evaluated using the [LM Evaluation
199
+ Harness](https://github.com/EleutherAI/lm-evaluation-harness). You can access
200
+ the results by model and step at `results/json/*` in the [GitHub
201
+ repository](https://github.com/EleutherAI/pythia/tree/main/results/json/v1.1-evals).<br>
202
+ Expand the sections below to see plots of evaluation results for all
203
+ Pythia and Pythia-deduped models compared with OPT and BLOOM.
204
+
205
+ <details>
206
+ <summary>LAMBADA – OpenAI</summary>
207
+ <img src="/EleutherAI/pythia-12b/resolve/main/eval_plots/lambada_openai_v1.png" style="width:auto"/>
208
+ </details>
209
+
210
+ <details>
211
+ <summary>Physical Interaction: Question Answering (PIQA)</summary>
212
+ <img src="/EleutherAI/pythia-12b/resolve/main/eval_plots/piqa_v1.png" style="width:auto"/>
213
+ </details>
214
+
215
+ <details>
216
+ <summary>WinoGrande</summary>
217
+ <img src="/EleutherAI/pythia-12b/resolve/main/eval_plots/winogrande_v1.png" style="width:auto"/>
218
+ </details>
219
+
220
+ <details>
221
+ <summary>AI2 Reasoning Challenge—Easy Set</summary>
222
+ <img src="/EleutherAI/pythia-12b/resolve/main/eval_plots/arc_easy_v1.png" style="width:auto"/>
223
+ </details>
224
+
225
+ <details>
226
+ <summary>SciQ</summary>
227
+ <img src="/EleutherAI/pythia-12b/resolve/main/eval_plots/sciq_v1.png" style="width:auto"/>
228
+ </details>
229
+
230
+ ## Changelog
231
+
232
+ This section compares differences between previously released
233
+ [Pythia v0](https://huggingface.co/models?other=pythia_v0) and the current
234
+ models. See Appendix B of the Pythia paper for further discussion of these
235
+ changes and the motivation behind them. We found that retraining Pythia had no
236
+ impact on benchmark performance.
237
+
238
+ - All model sizes are now trained with uniform batch size of 2M tokens.
239
+ Previously, the models of size 160M, 410M, and 1.4B parameters were trained
240
+ with batch sizes of 4M tokens.
241
+ - We added checkpoints at initialization (step 0) and steps {1,2,4,8,16,32,64,
242
+ 128,256,512} in addition to every 1000 training steps.
243
+ - Flash Attention was used in the new retrained suite.
244
+ - We remedied a minor inconsistency that existed in the original suite: all
245
+ models of size 2.8B parameters or smaller had a learning rate (LR) schedule
246
+ which decayed to a minimum LR of 10% the starting LR rate, but the 6.9B and
247
+ 12B models all used an LR schedule which decayed to a minimum LR of 0. In
248
+ the redone training runs, we rectified this inconsistency: all models now were
249
+ trained with LR decaying to a minimum of 0.1× their maximum LR.
250
+
251
+ ### Naming convention and parameter count
252
+
253
+ *Pythia* models were renamed in January 2023. It is possible that the old
254
+ naming convention still persists in some documentation by accident. The
255
+ current naming convention (70M, 160M, etc.) is based on total parameter count.
256
+
257
+ <figure style="width:32em">
258
+
259
+ | current Pythia suffix | old suffix | total params | non-embedding params |
260
+ | --------------------: | ---------: | -------------: | -------------------: |
261
+ | 70M | 19M | 70,426,624 | 18,915,328 |
262
+ | 160M | 125M | 162,322,944 | 85,056,000 |
263
+ | 410M | 350M | 405,334,016 | 302,311,424 |
264
+ | 1B | 800M | 1,011,781,632 | 805,736,448 |
265
+ | 1.4B | 1.3B | 1,414,647,808 | 1,208,602,624 |
266
+ | 2.8B | 2.7B | 2,775,208,960 | 2,517,652,480 |
267
+ | 6.9B | 6.7B | 6,857,302,016 | 6,444,163,072 |
268
+ | 12B | 13B | 11,846,072,320 | 11,327,027,200 |
269
+ </figure>