OLM GPT-2 October 2022

This is a more up-to-date version of the original GPT-2. In addition to being more up-to-date, it also tends to perform better than the original GPT2 on standard benchmarks. It was trained on a cleaned October 2022 snapshot of Common Crawl and Wikipedia.

This model was created as part of the OLM project, which has the goal of continuously training and releasing models that are up-to-date and comparable in standard language model performance to their static counterparts. This is important because we want our models to know about events like COVID or a presidential election right after they happen.

Intended uses

You can use the raw model for text generation or fine-tune it to a downstream task.

How to use

You can use this model directly with a pipeline for text generation. Since the generation relies on some randomness, we set a seed for reproducibility:

>>> from transformers import pipeline, set_seed
>>> # It is important to include the bad_words_ids=[[0,2]] if you want this model to stay on topic.
>>> # Otherwise, the model may generate start and end tokens followed by text that is not relevant to
>>> # the previous text.
>>> generator = pipeline('text-generation', model='olm/olm-gpt2-oct-2022', bad_words_ids=[[0,2]])
>>> set_seed(42)
>>> # This example also illustrates that sometimes our model generates
>>> # bloggy/spammy/webb-y things, even though it gets higher evaluation results
>>> # than the original GPT-2 accross a variety of benchmarks. See the first output.
>>> generator("Hello, I'm a language model,", max_length=30, num_return_sequences=5)
Setting `pad_token_id` to `eos_token_id`:50256 for open-end generation.
[
{'generated_text': "Hello, I'm a language model, but you can take me if I want.\nReplyDelete\nReplies\nReply\nAnonymous October 17, 2011"},
{'generated_text': "Hello, I'm a language model, and here's some useful news for you all: The release date for the new release of"},
{'generated_text': "Hello, I'm a language model, I'm not a developer or anybody who's working on those. I'm a freelancer... I"},
{'generated_text': "Hello, I'm a language model, a language analyst, and a language system designer. I'm just curious about the"},
{'generated_text': "Hello, I'm a language model, I'm passionate about languages, but I don't understand how my system works, the interaction"}
]

Here is how to use this model to get the features of a given text in PyTorch:

from transformers import AutoTokenizer, AutoModelForCausalLM
tokenizer = AutoTokenizer.from_pretrained('olm/olm-gpt2-oct-2022')
model = AutoModelForCausalLM.from_pretrained('olm/olm-gpt2-oct-2022')
text = "Replace me by any text you'd like."
encoded_input = tokenizer(text, return_tensors='pt')
output = model(**encoded_input)

Dataset

The model and tokenizer were trained with this October 2022 cleaned Common Crawl dataset plus this October 2022 cleaned Wikipedia dataset.
The tokenized version of these concatenated datasets is here.
The datasets were created with this repo.

Training

The model was trained according to the OLM GPT2 instructions at this repo.

Evaluation results

The model achieves the following results without any fine-tuning (zero-shot):

Task	Metric	Original GPT2	OLM GPT2 Oct 2022 (Ours)	Significance of Difference (two-tailed p-value)
rte	acc	0.5307	0.5415	0.7188
piqa	acc/acc_norm	0.6289/0.6251	0.6638/0.6670	0.0020/0.0002
copa	acc	0.6400	0.6900	0.3000
record	f1/em	0.7094/0.7026	0.6874/0.6810	0.0000/0.0000
boolq	acc	0.4872	0.5606	0.0000
cb	acc/f1	0.4101/0.2619	0.3571/0.1754	0.4193/NA
hellaswag	acc/acc_norm	0.2892/0.3114	0.3076/0.3491	0.0000/0.0000
mrpc	acc/f1	0.5662/0.6911	0.6495/0.7741	0.0007/0.0002
multirc	acc	0.0189	0.0115	0.0959
lambada	ppl/acc	40.0554/0.3256	28.6733/0.3625	0.0000/0.0000
wsc	acc	0.4327	0.3654	0.1679
wic	acc	0.4922	0.5	0.6924
mnli	acc	0.3372	0.3471	0.0384
qnli	acc	0.5017	0.4981	0.5884
cola	mcc	0.0126	0.0181	0.8614
triviaqa	acc	0.0151	0.0182	0.0048
winogrande	acc	0.5162	0.5114	0.7360
webqs	acc	0.0030	0.0108	0.0000
arc_easy	acc/acc_norm	0.4381/0.3948	0.4651/0.4247	0.0082/0.0029
arc_challenge	acc/acc_norm	0.1903/0.2270	0.1997/0.2329	0.4132/0.6256

To get these results, we used commit 4f0410a4be0049729078376ce36a42dc308b6e38 of the Eleuther AI evaluation harness here, which can produce results different than those reported in the GPT2 paper. We added a change here to enable evaluation of the OLM GPT2, which has a very slightly different vocab size. The p-values come from the stderr from the evaluation harness, plus a normal distribution assumption.