RichardErkhov commited on
Commit
7be0536
1 Parent(s): 4aebe97

uploaded readme

Browse files
Files changed (1) hide show
  1. README.md +299 -0
README.md ADDED
@@ -0,0 +1,299 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ Quantization made by Richard Erkhov.
2
+
3
+ [Github](https://github.com/RichardErkhov)
4
+
5
+ [Discord](https://discord.gg/pvy7H8DZMG)
6
+
7
+ [Request more models](https://github.com/RichardErkhov/quant_request)
8
+
9
+
10
+ pythia-410m-deduped - bnb 8bits
11
+ - Model creator: https://huggingface.co/EleutherAI/
12
+ - Original model: https://huggingface.co/EleutherAI/pythia-410m-deduped/
13
+
14
+
15
+
16
+
17
+ Original model description:
18
+ ---
19
+ language:
20
+ - en
21
+ tags:
22
+ - pytorch
23
+ - causal-lm
24
+ - pythia
25
+ license: apache-2.0
26
+ datasets:
27
+ - EleutherAI/the_pile_deduplicated
28
+ ---
29
+
30
+ The *Pythia Scaling Suite* is a collection of models developed to facilitate
31
+ interpretability research [(see paper)](https://arxiv.org/pdf/2304.01373.pdf).
32
+ It contains two sets of eight models of sizes
33
+ 70M, 160M, 410M, 1B, 1.4B, 2.8B, 6.9B, and 12B. For each size, there are two
34
+ models: one trained on the Pile, and one trained on the Pile after the dataset
35
+ has been globally deduplicated. All 8 model sizes are trained on the exact
36
+ same data, in the exact same order. We also provide 154 intermediate
37
+ checkpoints per model, hosted on Hugging Face as branches.
38
+
39
+ The Pythia model suite was designed to promote scientific
40
+ research on large language models, especially interpretability research.
41
+ Despite not centering downstream performance as a design goal, we find the
42
+ models <a href="#evaluations">match or exceed</a> the performance of
43
+ similar and same-sized models, such as those in the OPT and GPT-Neo suites.
44
+
45
+ <details>
46
+ <summary style="font-weight:600">Details on previous early release and naming convention.</summary>
47
+
48
+ Previously, we released an early version of the Pythia suite to the public.
49
+ However, we decided to retrain the model suite to address a few hyperparameter
50
+ discrepancies. This model card <a href="#changelog">lists the changes</a>;
51
+ see appendix B in the Pythia paper for further discussion. We found no
52
+ difference in benchmark performance between the two Pythia versions.
53
+ The old models are
54
+ [still available](https://huggingface.co/models?other=pythia_v0), but we
55
+ suggest the retrained suite if you are just starting to use Pythia.<br>
56
+ **This is the current release.**
57
+
58
+ Please note that all models in the *Pythia* suite were renamed in January
59
+ 2023. For clarity, a <a href="#naming-convention-and-parameter-count">table
60
+ comparing the old and new names</a> is provided in this model card, together
61
+ with exact parameter counts.
62
+ </details>
63
+ <br>
64
+
65
+ # Pythia-410M-deduped
66
+
67
+ ## Model Details
68
+
69
+ - Developed by: [EleutherAI](http://eleuther.ai)
70
+ - Model type: Transformer-based Language Model
71
+ - Language: English
72
+ - Learn more: [Pythia's GitHub repository](https://github.com/EleutherAI/pythia)
73
+ for training procedure, config files, and details on how to use.
74
+ [See paper](https://arxiv.org/pdf/2304.01373.pdf) for more evals and implementation
75
+ details.
76
+ - Library: [GPT-NeoX](https://github.com/EleutherAI/gpt-neox)
77
+ - License: Apache 2.0
78
+ - Contact: to ask questions about this model, join the [EleutherAI
79
+ Discord](https://discord.gg/zBGx3azzUn), and post them in `#release-discussion`.
80
+ Please read the existing *Pythia* documentation before asking about it in the
81
+ EleutherAI Discord. For general correspondence: [contact@eleuther.
82
+ ai](mailto:contact@eleuther.ai).
83
+
84
+ <figure>
85
+
86
+ | Pythia model | Non-Embedding Params | Layers | Model Dim | Heads | Batch Size | Learning Rate | Equivalent Models |
87
+ | -----------: | -------------------: | :----: | :-------: | :---: | :--------: | :-------------------: | :--------------------: |
88
+ | 70M | 18,915,328 | 6 | 512 | 8 | 2M | 1.0 x 10<sup>-3</sup> | — |
89
+ | 160M | 85,056,000 | 12 | 768 | 12 | 2M | 6.0 x 10<sup>-4</sup> | GPT-Neo 125M, OPT-125M |
90
+ | 410M | 302,311,424 | 24 | 1024 | 16 | 2M | 3.0 x 10<sup>-4</sup> | OPT-350M |
91
+ | 1.0B | 805,736,448 | 16 | 2048 | 8 | 2M | 3.0 x 10<sup>-4</sup> | — |
92
+ | 1.4B | 1,208,602,624 | 24 | 2048 | 16 | 2M | 2.0 x 10<sup>-4</sup> | GPT-Neo 1.3B, OPT-1.3B |
93
+ | 2.8B | 2,517,652,480 | 32 | 2560 | 32 | 2M | 1.6 x 10<sup>-4</sup> | GPT-Neo 2.7B, OPT-2.7B |
94
+ | 6.9B | 6,444,163,072 | 32 | 4096 | 32 | 2M | 1.2 x 10<sup>-4</sup> | OPT-6.7B |
95
+ | 12B | 11,327,027,200 | 36 | 5120 | 40 | 2M | 1.2 x 10<sup>-4</sup> | — |
96
+ <figcaption>Engineering details for the <i>Pythia Suite</i>. Deduped and
97
+ non-deduped models of a given size have the same hyperparameters. “Equivalent”
98
+ models have <b>exactly</b> the same architecture, and the same number of
99
+ non-embedding parameters.</figcaption>
100
+ </figure>
101
+
102
+ ## Uses and Limitations
103
+
104
+ ### Intended Use
105
+
106
+ The primary intended use of Pythia is research on the behavior, functionality,
107
+ and limitations of large language models. This suite is intended to provide
108
+ a controlled setting for performing scientific experiments. We also provide
109
+ 154 checkpoints per model: initial `step0`, 10 log-spaced checkpoints
110
+ `step{1,2,4...512}`, and 143 evenly-spaced checkpoints from `step1000` to
111
+ `step143000`. These checkpoints are hosted on Hugging Face as branches. Note
112
+ that branch `143000` corresponds exactly to the model checkpoint on the `main`
113
+ branch of each model.
114
+
115
+ You may also further fine-tune and adapt Pythia-410M-deduped for deployment,
116
+ as long as your use is in accordance with the Apache 2.0 license. Pythia
117
+ models work with the Hugging Face [Transformers
118
+ Library](https://huggingface.co/docs/transformers/index). If you decide to use
119
+ pre-trained Pythia-410M-deduped as a basis for your fine-tuned model, please
120
+ conduct your own risk and bias assessment.
121
+
122
+ ### Out-of-scope use
123
+
124
+ The Pythia Suite is **not** intended for deployment. It is not a in itself
125
+ a product and cannot be used for human-facing interactions. For example,
126
+ the model may generate harmful or offensive text. Please evaluate the risks
127
+ associated with your particular use case.
128
+
129
+ Pythia models are English-language only, and are not suitable for translation
130
+ or generating text in other languages.
131
+
132
+ Pythia-410M-deduped has not been fine-tuned for downstream contexts in which
133
+ language models are commonly deployed, such as writing genre prose,
134
+ or commercial chatbots. This means XNPythia-410M-dedupedAME will **not**
135
+ respond to a given prompt the way a product like ChatGPT does. This is because,
136
+ unlike this model, ChatGPT was fine-tuned using methods such as Reinforcement
137
+ Learning from Human Feedback (RLHF) to better “follow” human instructions.
138
+
139
+ ### Limitations and biases
140
+
141
+ The core functionality of a large language model is to take a string of text
142
+ and predict the next token. The token used by the model need not produce the
143
+ most “accurate” text. Never rely on Pythia-410M-deduped to produce factually accurate
144
+ output.
145
+
146
+ This model was trained on [the Pile](https://pile.eleuther.ai/), a dataset
147
+ known to contain profanity and texts that are lewd or otherwise offensive.
148
+ See [Section 6 of the Pile paper](https://arxiv.org/abs/2101.00027) for a
149
+ discussion of documented biases with regards to gender, religion, and race.
150
+ Pythia-410M-deduped may produce socially unacceptable or undesirable text, *even if*
151
+ the prompt itself does not include anything explicitly offensive.
152
+
153
+ If you plan on using text generated through, for example, the Hosted Inference
154
+ API, we recommend having a human curate the outputs of this language model
155
+ before presenting it to other people. Please inform your audience that the
156
+ text was generated by Pythia-410M-deduped.
157
+
158
+ ### Quickstart
159
+
160
+ Pythia models can be loaded and used via the following code, demonstrated here
161
+ for the third `pythia-70m-deduped` checkpoint:
162
+
163
+ ```python
164
+ from transformers import GPTNeoXForCausalLM, AutoTokenizer
165
+
166
+ model = GPTNeoXForCausalLM.from_pretrained(
167
+ "EleutherAI/pythia-70m-deduped",
168
+ revision="step3000",
169
+ cache_dir="./pythia-70m-deduped/step3000",
170
+ )
171
+
172
+ tokenizer = AutoTokenizer.from_pretrained(
173
+ "EleutherAI/pythia-70m-deduped",
174
+ revision="step3000",
175
+ cache_dir="./pythia-70m-deduped/step3000",
176
+ )
177
+
178
+ inputs = tokenizer("Hello, I am", return_tensors="pt")
179
+ tokens = model.generate(**inputs)
180
+ tokenizer.decode(tokens[0])
181
+ ```
182
+
183
+ Revision/branch `step143000` corresponds exactly to the model checkpoint on
184
+ the `main` branch of each model.<br>
185
+ For more information on how to use all Pythia models, see [documentation on
186
+ GitHub](https://github.com/EleutherAI/pythia).
187
+
188
+ ## Training
189
+
190
+ ### Training data
191
+
192
+ Pythia-410M-deduped was trained on the Pile **after the dataset has been globally
193
+ deduplicated**.<br>
194
+ [The Pile](https://pile.eleuther.ai/) is a 825GiB general-purpose dataset in
195
+ English. It was created by EleutherAI specifically for training large language
196
+ models. It contains texts from 22 diverse sources, roughly broken down into
197
+ five categories: academic writing (e.g. arXiv), internet (e.g. CommonCrawl),
198
+ prose (e.g. Project Gutenberg), dialogue (e.g. YouTube subtitles), and
199
+ miscellaneous (e.g. GitHub, Enron Emails). See [the Pile
200
+ paper](https://arxiv.org/abs/2101.00027) for a breakdown of all data sources,
201
+ methodology, and a discussion of ethical implications. Consult [the
202
+ datasheet](https://arxiv.org/abs/2201.07311) for more detailed documentation
203
+ about the Pile and its component datasets. The Pile can be downloaded from
204
+ the [official website](https://pile.eleuther.ai/), or from a [community
205
+ mirror](https://the-eye.eu/public/AI/pile/).
206
+
207
+ ### Training procedure
208
+
209
+ All models were trained on the exact same data, in the exact same order. Each
210
+ model saw 299,892,736,000 tokens during training, and 143 checkpoints for each
211
+ model are saved every 2,097,152,000 tokens, spaced evenly throughout training,
212
+ from `step1000` to `step143000` (which is the same as `main`). In addition, we
213
+ also provide frequent early checkpoints: `step0` and `step{1,2,4...512}`.
214
+ This corresponds to training for just under 1 epoch on the Pile for
215
+ non-deduplicated models, and about 1.5 epochs on the deduplicated Pile.
216
+
217
+ All *Pythia* models trained for 143000 steps at a batch size
218
+ of 2M (2,097,152 tokens).<br>
219
+ See [GitHub](https://github.com/EleutherAI/pythia) for more details on training
220
+ procedure, including [how to reproduce
221
+ it](https://github.com/EleutherAI/pythia/blob/main/README.md#reproducing-training).<br>
222
+ Pythia uses the same tokenizer as [GPT-NeoX-
223
+ 20B](https://huggingface.co/EleutherAI/gpt-neox-20b).
224
+
225
+ ## Evaluations
226
+
227
+ All 16 *Pythia* models were evaluated using the [LM Evaluation
228
+ Harness](https://github.com/EleutherAI/lm-evaluation-harness). You can access
229
+ the results by model and step at `results/json/*` in the [GitHub
230
+ repository](https://github.com/EleutherAI/pythia/tree/main/results/json/).<br>
231
+ Expand the sections below to see plots of evaluation results for all
232
+ Pythia and Pythia-deduped models compared with OPT and BLOOM.
233
+
234
+ <details>
235
+ <summary>LAMBADA – OpenAI</summary>
236
+ <img src="/EleutherAI/pythia-12b/resolve/main/eval_plots/lambada_openai_v1.png" style="width:auto"/>
237
+ </details>
238
+
239
+ <details>
240
+ <summary>Physical Interaction: Question Answering (PIQA)</summary>
241
+ <img src="/EleutherAI/pythia-12b/resolve/main/eval_plots/piqa_v1.png" style="width:auto"/>
242
+ </details>
243
+
244
+ <details>
245
+ <summary>WinoGrande</summary>
246
+ <img src="/EleutherAI/pythia-12b/resolve/main/eval_plots/winogrande_v1.png" style="width:auto"/>
247
+ </details>
248
+
249
+ <details>
250
+ <summary>AI2 Reasoning Challenge—Easy Set</summary>
251
+ <img src="/EleutherAI/pythia-12b/resolve/main/eval_plots/arc_easy_v1.png" style="width:auto"/>
252
+ </details>
253
+
254
+ <details>
255
+ <summary>SciQ</summary>
256
+ <img src="/EleutherAI/pythia-12b/resolve/main/eval_plots/sciq_v1.png" style="width:auto"/>
257
+ </details>
258
+
259
+ ## Changelog
260
+
261
+ This section compares differences between previously released
262
+ [Pythia v0](https://huggingface.co/models?other=pythia_v0) and the current
263
+ models. See Appendix B of the Pythia paper for further discussion of these
264
+ changes and the motivation behind them. We found that retraining Pythia had no
265
+ impact on benchmark performance.
266
+
267
+ - All model sizes are now trained with uniform batch size of 2M tokens.
268
+ Previously, the models of size 160M, 410M, and 1.4B parameters were trained
269
+ with batch sizes of 4M tokens.
270
+ - We added checkpoints at initialization (step 0) and steps {1,2,4,8,16,32,64,
271
+ 128,256,512} in addition to every 1000 training steps.
272
+ - Flash Attention was used in the new retrained suite.
273
+ - We remedied a minor inconsistency that existed in the original suite: all
274
+ models of size 2.8B parameters or smaller had a learning rate (LR) schedule
275
+ which decayed to a minimum LR of 10% the starting LR rate, but the 6.9B and
276
+ 12B models all used an LR schedule which decayed to a minimum LR of 0. In
277
+ the redone training runs, we rectified this inconsistency: all models now were
278
+ trained with LR decaying to a minimum of 0.1× their maximum LR.
279
+
280
+ ### Naming convention and parameter count
281
+
282
+ *Pythia* models were renamed in January 2023. It is possible that the old
283
+ naming convention still persists in some documentation by accident. The
284
+ current naming convention (70M, 160M, etc.) is based on total parameter count.
285
+
286
+ <figure style="width:32em">
287
+
288
+ | current Pythia suffix | old suffix | total params | non-embedding params |
289
+ | --------------------: | ---------: | -------------: | -------------------: |
290
+ | 70M | 19M | 70,426,624 | 18,915,328 |
291
+ | 160M | 125M | 162,322,944 | 85,056,000 |
292
+ | 410M | 350M | 405,334,016 | 302,311,424 |
293
+ | 1B | 800M | 1,011,781,632 | 805,736,448 |
294
+ | 1.4B | 1.3B | 1,414,647,808 | 1,208,602,624 |
295
+ | 2.8B | 2.7B | 2,775,208,960 | 2,517,652,480 |
296
+ | 6.9B | 6.7B | 6,857,302,016 | 6,444,163,072 |
297
+ | 12B | 13B | 11,846,072,320 | 11,327,027,200 |
298
+ </figure>
299
+