Language Model Zoo 🦍


Basic attempts with Italian in two weeks time here!

(Hamel Husain) #145

Hello folks does anyone have any advice on where I can host a pre-trained language model for the community without it costing me a bunch of money? I was thinking of hosting it on Google Cloud Bucket but it charges me for every download. Does anyone have any advice?

(Jeremy Howard (Admin)) #146

I’m happy to host any language models people create (with credit, of course).

(Sarada Lee) #147

Jeremy setup a new website for NLP ( There is a place holder for Model Zoo. Keep working on it.

(Sooheon Kim) #148

Try going to

At least that’s what I did for disambiguating ko and kr for Korean.

(Jens Dahl Møllerhøj) #149

I’m going to give Danish ( and Norwegian ( a shot :slight_smile:

(Monique Monteiro) #150

I tried to run @binga data preparation for Telugu, but I’d like to know the exact format used for wikipedia text. The wikipedia data is not on github, and it seems the IMDB’s data format was used, 'cause the code is similar to the one used in the notebook for lesson 10.

Thanks in advance


I’m trying out the LM on BPE (from sentencepiece) input with 30k subword units, and I’m getting ridiculously low perplexity (3, but still not finished training) and high accuracy (81% so far).

My guess is that the LM has an easy time on sentencepiece input, because a lot of the time the “tokens” are just a single character, so maybe the LSTM is just learning to remember the word vocabulary rather than any meaningful semantics.

So my question: anyone else running experiments using sentencepiece input? What kinds of perplexities and accuracies are you seeing? And any results on downstream classification tasks already? I would be very interested in seeing a comparison between sentencepiece-tokenized and word-tokenized input for classification.

(Phani Srikanth) #152

Hi @monilouise,

Thank you for going through my work.

The objective of the language model in Telugu was to take the wikipedia articles in Telugu language and build a LM on top of it.

The file pretty much contains all that you need.

  1. It obtains the all the telugu wikipedia articles of the telugu language from wikimedia dumps (where all the language wiki dumps are hosted). These are html files
  2. The parses the html files and converts them into json articles.
  3. The notebook in the same repo reads the json files, extracts the body of the wikipedia article, converts all the data into one single dataframe and from there you should be good to go.

Jeremy prefers to convert all the source datasets into a common format so that the downstream code doesn’t have to be changed and it fits the fastai framework - way of doing this. We prefer the same approach and hence you see the IMDB way of doing things. The data preparation section in my notebook essentially does that.

I hope this helps. Feel free to ask any questions.

Also, thanks for pointing this out. I fixed a bug in my code while writing this answer. :slight_smile:

(Monique Monteiro) #153

Thank you very much, @binga!

I saw that there are multiple kinds of Wikipedia dumps (at least for Portuguese), and I suspect these different formats lead to different results when we run the notebook, despite using WikiExtractor.

(Kristian Rother) #154

I’ll give German a shot over the next couple of weeks. I was thinking about using Wikipedia or maybe Projekt Gutenberg-DE. I’ll start with a literature review since there have to be language models and classifiers of all sorts already. Hopefully the state of the art can be improved upon :slight_smile:

(Thomas) #155

Hi Kristian
Great! I was surprised that noone was looking at de when I look yesterday. I’ll keep you updated, but I’ll also be very interested to hear what you are doing.
Best regards

(Sarada Lee) #156

I found similar problem. Due to limited documentation from SentencePiece (SP) official website, here are the observations from my experiment (for Chinese Transitional*). It seems that we need to set the “right” vocab_size to get meaningful semantics:

  • When I set vocab_size just right (exact numbers of words plus 4 (automatically added by SP) or rounded up to the nearest thousand, the tokenization is restricted to each word.

  • When I set vocab_size too high, say 103M, I got an error message to adjust to “soft limit”

  • When I used --hard_vocab_size = False (ie soft limit), SP applied vocab_size = 8000 (ie just right limit) for me. You can use sp.GetPieceSize() to check your vocab_size.

  • Then, I played with different vocab_size over and over again and used sp.Encode() to test a sample sentence. I found 200k tokens (around 27 times of alphabet size) work well for me. Here is the example:

Now, I need to do the LM with the new set of tokens before answering the rest of your question. I just found more ways which did not work! :sweat_smile:

Note: *word = character for Chinese Tranditional

My dataset’s basic information (including alphabet size)

(Kristian Rother) #157

@t-v Perfect, I was looking for your name because I saw it in the wiki but didn’t find it in the thread. I’ll keep you updated. Also interested in hearing what you do :slight_smile:

(Monique Monteiro) #158


Does the 100 milion limit refer to the total number of words in all the documents?

(Jeremy Howard (Admin)) #159

You probably don’t want this to get much higher that 30,000 or you’ll have memory and speed problems when training the LM, FYI.

(Sarada Lee) #160

Thanks for the advice. I am planning to try with high vocab_size to capture more interesting phases. Then, I will increase the minimum frequency for each word when training LM. My concern is:

Using the Queen Elizabeth II of United Kingdom (英女王伊利沙伯二世) as an example:
When vocab_size = 30K, sp.PieceId = 0 which means out of range. It was encoded as ‘英’ + ‘女王’ + ‘伊利沙伯二世’.



However, using bigger vocab_size (ie 200k in this case), sp returned it as one token.

I am expecting the same embedding matrix for ‘英女王伊利沙伯二世’ will be updated during forward and backward training. But, I am not sure the embedding for ‘’ (multiple meanings) -> ‘女王’ (Queen) -> ‘伊利沙伯二世’ (name) vs ‘伊利沙伯二世’ (name) -> ‘女王’ (Queen) -> ‘英’ (as surname here?) will generate the same embedding. Is my concern valid?

(Sooheon Kim) #161

Re: the sentencepiece discussion

The official sp repo has two example data files here: botchan.txt and wagahaiwa_nekodearu.txt

In botchan, a “sentence” is just the text line-wrapped at 80 or so characters. There is no significance given to actual semantic sentences, and line breaks occur basically randomly across the text. In wagahaiwa_nekodearu, a “sentence” is more like a paragraph, consisting of multiple semantic sentences.

@Moody, how did you approach sentence tokenization to feed to sentencepiece?

(Sarada Lee) #162

In the SP website, there is a Python wrapper link under the pip install instruction. Then, follow the instructions to tokenise a wiki file (with minimal preprocessing). You can specify the number of output tokens accordingly.

(Sooheon Kim) #163

I am using the python wrapper, have downloaded Korean wiki, and have no problems getting SP to run in general (and output “Pieces”).

The problem is it’s kind of unclear what the right “minimal preprocessing” is, to derive sentences.

I can:

  1. Split at every instance of “.”
  2. Split at every instance of regex “\.\s+”, or something like that
  3. Split every 80 characters (like this)
  4. Split according to natural “\n” in wiki text, corresponding to paragraphs, not sentences. (like this)