Move texts into app.py

EVALUATION.md

@@ -1,59 +0,0 @@
-## Evaluation
-
-### Running evaluation runs
-
-Each pre-trained model was evaluated by fine-tuning on summarization and translation. The learning-rate was set to
-a constant schedule after a small warmup of 32 steps.
-Fine-tuning for evaluation was done on a limited set of 50K examples from the fine-tuning datasets.
-
-|                  | Summarization    | Translation       |
-|-----------------:|------------------|-------------------|
-|          Dataset | CNN Dailymail NL | CCMatrix en -> nl |
-|   #train samples | 50K              | 50K               |
-|        Optimizer | AdamW            | AdamW             |
-|    learning rate | 0.001            | 0.0005            |
-|    source length | 1024             | 128               |
-|    target length | 142              | 128               |
-|    #eval samples | 1000             | 1000              |
- | wandb link      | [eval_summ](https://wandb.ai/yepster/eval_dutch_cnndaily_202302_flax)|[eval_transl](https://wandb.ai/yepster/eval_dutch_ccmatrix_202302_flax) |
-
-The graph below shows the Rouge1 score for the summarization runs, evaluated
-after 25K and 50K examples on the [CNN Dailymail Dutch](https://huggingface.co/datasets/yhavinga/cnn_dailymail_dutch) dataset:
-
-![Rouge1 summarization](eval_summ_rouge1_202302.png)
-
-* Flan models perform almost instantly well on the summarization task, with `flan-t5-small`
-  showing performance comparable to Dutch T5 base models.
-* After 50K examples, the `ul2` models exhibit similar performance to the `flan` models.
-* I am surprised by the consistent bad scores for the `long-t5` runs. I've retried the fine-tuning of these models with
-  `float32` instead of `bfloat16`, but the results were the same. Maybe this is normal behaviour for these models
-  targeted at dealing with longer sequence lengths.
-
-The graph below shows the Bleu score for the translation runs, evaluated at step 25K and
-50K on the [CCMatrix](https://huggingface.co/datasets/yhavinga/ccmatrix_en_nl) dataset, from
-English to Dutch:
-
-![Bleu score translation](eval_transl_bleu_202302.png)
-
-* For the translation task from English to Dutch, the Dutch+English pre-trained models perform well. Also
-  `ul2` pre-trained models are consistently better than their `Flan`, `T5 Dutch` and
-  `mT5` counterparts.
-* Like with the summarization task, the `long-t5` models show bad performance, even after 50K examples. I do not understand 
-  cannot explain this at all for this translation task. With a sequence length of 128 input and output
-  tokens, the sliding attention window with radius length 127 of the `long-t5` models should be able to handle this.
-
-The figure below shows the evaluation scores for most models, with summarization Rouge1 on the x-axis (higher is better),
-and translation English to Dutch Bleu score on the y-axis (higher is better).
-The point size is proportional to the model size. UL2 models are blue, Flan models
-red, mT5 green and the other models black.
-
-![Evaluation T5 Dutch English](eval_t5_dutch_english.png)
-
-* For clarity not all models are shown. `t5-base-36L-dutch-english-cased` is model with
-  scores comparable to `ul2-large-dutch-english`, but with slower inference. All long-t5
-  runs are left out, as well as the `t5-v1.1-large-dutch-cased` model whose translation fine-tuning
-  diverged.
-* Across the board, for translation the models pre-trained with Dutch+English or Dutch converge faster than other models.
-  I was surprised to see `t5-xl-4l` among the best models on translation, as it has only 4 layers, and previous tests
-  showed that it had a very bad performance (In those tests I had forgot to force set the dropout rate to 0.0, and
-  apparently this model was very sensitive to dropout). 

INTRO.md

@@ -1,27 +0,0 @@
-# Dutch T5 models : UL2, T5, ByT5 and Long-T5 🇳🇱🇧🇪
-
-TL;DR: Dutch NLP gets a boost with state-of-the-art T5 models trained on the largest Dutch corpus, mC4, and additional datasets.
-See below for model lists and comparison.
-
-During the [HuggingFace Flax/Jax community week](https://discuss.huggingface.co/t/open-to-the-community-community-week-using-jax-flax-for-nlp-cv/7104) in the summer of 2021,
-I was granted access to Google's TPU Research Cloud (TRC),
-a cloud-based platform for machine learning research and development that provides access to Google's
-Tensor Processing Units (TPUs). My goal was to address the (then) shortage of T5 models for the Dutch language.
--- T5 is a state-of-the-art AI model architecture that can handle text as input and output,
-making it an ideal tool for NLP tasks such as summarization, translation, and question-answering --
-Since then, with extended access to the TRC, I have been able to train a variety of T5 models for Dutch.
-
-Relevant papers are:
- 
-* **[Exploring the Limits of Transfer Learning with a Unified Text-to-Text Transformer](https://arxiv.org/abs/1910.10683)** by *Colin Raffel, Noam Shazeer, Adam Roberts, Katherine Lee, Sharan Narang, Michael Matena, Yanqi Zhou, Wei Li, Peter J. Liu*.
-* **[ExT5: Towards Extreme Multi-Task Scaling for Transfer Learning](https://arxiv.org/abs/2111.10952)** by *Vamsi Aribandi, Yi Tay, Tal Schuster, Jinfeng Rao, Huaixiu Steven Zheng, Sanket Vaibhav Mehta, Honglei Zhuang, Vinh Q. Tran, Dara Bahri, Jianmo Ni, Jai Gupta, Kai Hui, Sebastian Ruder, Donald Metzler*.
-* **[Scale Efficiently: Insights from Pre-training and Fine-tuning Transformers](https://arxiv.org/abs/2109.10686)** by *Yi Tay, Mostafa Dehghani, Jinfeng Rao, William Fedus, Samira Abnar, Hyung Won Chung, Sharan Narang, Dani Yogatama, Ashish Vaswani, Donald Metzler*.
-* **[ByT5: Towards a token-free future with pre-trained byte-to-byte models](https://arxiv.org/abs/2105.13626)** by *Linting Xue, Aditya Barua, Noah Constant, Rami Al-Rfou, Sharan Narang, Mihir Kale, Adam Roberts, Colin Raffel*
-* **[LongT5: Efficient Text-To-Text Transformer for Long Sequences](https://arxiv.org/abs/2112.07916)** by *Mandy Guo, Joshua Ainslie, David Uthus, Santiago Ontanon, Jianmo Ni, Yun-Hsuan Sung, Yinfei Yang*
-* **[Scaling Up Models and Data with t5x and seqio](https://arxiv.org/abs/2203.17189)** by *Adam Roberts, Hyung Won Chung, Anselm Levskaya, Gaurav Mishra, James Bradbury, Daniel Andor, Sharan Narang, Brian Lester, Colin Gaffney, Afroz Mohiuddin, Curtis Hawthorne, Aitor Lewkowycz, Alex Salcianu, Marc van Zee, Jacob Austin, Sebastian Goodman, Livio Baldini Soares, Haitang Hu, Sasha Tsvyashchenko, Aakanksha Chowdhery, Jasmijn Bastings, Jannis Bulian, Xavier Garcia, Jianmo Ni, Andrew Chen, Kathleen Kenealy, Jonathan H. Clark, Stephan Lee, Dan Garrette, James Lee-Thorp, Colin Raffel, Noam Shazeer, Marvin Ritter, Maarten Bosma, Alexandre Passos, Jeremy Maitin-Shepard, Noah Fiedel, Mark Omernick, Brennan Saeta, Ryan Sepassi, Alexander Spiridonov, Joshua Newlan, Andrea Gesmundo*
-* **[UL2: Unifying Language Learning Paradigms](https://arxiv.org/abs/2205.05131)** by *Yi Tay, Mostafa Dehghani, Vinh Q. Tran, Xavier Garcia, Jason Wei, Xuezhi Wang, Hyung Won Chung, Dara Bahri, Tal Schuster, Huaixiu Steven Zheng, Denny Zhou, Neil Houlsby, Donald Metzler*
-
-Background on Google's TPU VM's and how to use the Huggingface transformers library to pre-train models can be found
-at the following links
-* https://discuss.huggingface.co/t/open-to-the-community-community-week-using-jax-flax-for-nlp-cv/7104
-* https://github.com/huggingface/transformers/tree/main/examples/research_projects/jax-projects#talks

PRETRAINING.md

@@ -1,66 +0,0 @@
-## Pre-training
-
-### mC4 dataset
-
-Together with the T5 model architecture and SeqIO, the T5 authors also created and released
-the multilingual [mC4 dataset](https://huggingface.co/datasets/allenai/c4).
-It was made available by AllenNLP on the HuggingFace Dataset hub.
-Our team confirmed that the Dutch portion of the mC4 dataset was deduplicated,
-and we cleaned the Dutch portion of the mC4 dataset using [code adapted](https://gitlab.com/yhavinga/c4nlpreproc) from the TensorFlow C4 dataset.
-The resulting [mc4_nl_cleaned](https://huggingface.co/datasets/yhavinga/mc4_nl_cleaned) dataset on the HuggingFace hub
-has configs for several sizes, and also configs for interleaved mixed Dutch and English
-texts, e.g. [micro_en_nl](https://huggingface.co/datasets/yhavinga/mc4_nl_cleaned/viewer/micro_en_nl/train).
-The `_en_nl` configs were added to accommodate multi-language pre-training
-with the Huggingface pre-training script, that accepts only a single dataset as input.
-The full, cleaned Dutch mC4 dataset is 151GB and remains (as of June 2022) the largest available Dutch
-corpus on the HuggingFace Dataset hub.
-
-### Additional books, Wikipedia and Dutch news articles datasets
-
-The `t5_1_1` and `ul2` models have also been trained on Dutch books, the Dutch subset of Wikipedia (2022-03-20),
-the English subset of Wikipedia (2022-03-01), and a subset of "mc4_nl_cleaned" containing only texts
-from Dutch and Belgian newspapers. Mixing in the these datasets was done to bias the model towards
-descriptions of events in the Netherlands and Belgium.
-
-### Pre-Training Objectives
-
-The T5 models are pre-trained using the [span corruption](https://arxiv.org/abs/1910.10683) denoising objective.
-15% of the tokens in the text are masked, and each span
-of masked tokens is replaced with a special token known as a sentinel token, where each span is assigned
-its own sentinel token. The model is then trained to predict for each sentinel token the original text
-that was replaced by the sentinel tokens.
-
-The UL2 models are pre-trained with the [Mixture-of-Denoisers (MoD)](https://arxiv.org/abs/2205.05131) objective, that combines diverse pre-training
-paradigms together. UL2 frames different objective functions for training language models as denoising tasks, where
-the model has to recover missing sub-sequences of a given input. During pre-training it uses a novel mixture-of-denoisers
-that samples from a varied set of such objectives, each with different configurations. UL2 is trained using a mixture of
-three denoising tasks:
-
-1. R-denoising (or regular span corruption), which emulates the standard T5 span corruption objective;
-2. X-denoising (or extreme span corruption); and
-3. S-denoising (or sequential PrefixLM).
-
-### Pre-training software
-
-#### Huggingface [run_t5_mlm_flax.py](https://github.com/huggingface/transformers/blob/main/examples/flax/language-modeling/run_t5_mlm_flax.py)
-
-All models except `t5_1_1` and `ul2` were pre-trained using the Huggingface `run_t5_mlm_flax.py` script.
-This script is a good fit if you want to get a grasp what's needed to pre-train a language model
-with Flax and Jax, since all data preparation, model instantiation, loss function, and training loop are
-contained in a single file.
-
-#### Google's [T5X](https://github.com/google-research/t5x)
-
-The Dutch `t5_1_1` and `ul2` models were pre-trained using T5X. This is a modular framework that can be used for
-pre-training, fine-tuning, and evaluation of T5 models. Because of its modular and pluggable design,
-by only supplying a few configuration and code files, it is possible to pre-train with your own definitions.
-It is even possible to define custom neural network layers and architectures, though I did not do this and only
-pre-trained the default T5 encoder-decoder architecture, and varied only the pre-training objective, and the
-datasets used and mixed with SeqIO.
-
-#### Conversion script from T5X to HF
-
-The T5X models were converted to Huggingface Flax T5 format using a script that was adapted from the
-[T5X checkpoint to HuggingFace Flax conversion script](https://github.com/huggingface/transformers/blob/main/src/transformers/models/t5/convert_t5x_checkpoint_to_flax.py).
-This script was modified to cast weights to bf16, and to also convert to pytorch format.
-For this conversion to be successful, the T5X model had to be saved with `use_gda=False` set in the GIN file.

README.md

@@ -4,7 +4,6 @@ emoji: 🚀
 colorFrom: blue
 colorTo: pink
 sdk: streamlit
-sdk_version: 1.10.0
 pinned: false
 app_file: app.py
 license: afl-3.0

REMARKS.md

@@ -1,20 +0,0 @@
-## Miscellaneous remarks
-
-* Use loss regularization when training with `bfloat16` for better results (more info below).
-* Be cautious of the dropout rate in the config.json file and consider training without it.
-  Check in a model's `config.json` what the dropout rate has been set to. Unless you
-  intend to run many epochs on the same data, its worth to try a training run without dropout.
-  If you want to compare losses, be sure to set the dropout rate equal.
-  The smaller models can probably always be trained without.
-* Training with more layers is much slower than you'd expect from the increased model size.
-  It is also more difficult to get batch size and learning rate right. Below is a section
-  about finding the right hyperparameters for the base-36L training.
-* For the translation task, I am not sure that a 'deep-narrow' model (e.g. base-nl36) is better than a normal model
-  of comparable size (e.g. `large`).
-* PyCharm's remote debugging features are useful to inspect variables on either a TPU VM or your deep-learning rig.
-* When increasing the batch size, increase the learning rate. bs * 2 -> lr * sqrt(2) is a good heuristic but mileage may
-  vary.
-* Dataset quality is a key success factor. Do not expect a model to magically turn mediocre data into magic. This holds for 
-  the pre-training data, fine-tuning and also evaluating.
-* Good Bleu score does not necessarily mean fluent text. Evaluation loss on a large translation dataset might be
-  better suited for model comparison, if the models have a tokenizer of comparable size.

app.py

@@ -78,11 +78,103 @@ def main():
     with open("style.css") as f:
         st.markdown(f"<style>{f.read()}</style>", unsafe_allow_html=True)
 
-    with open("INTRO.md", "r") as f:
-        st.markdown(f.read())
-
-    with open("PRETRAINING.md", "r") as f:
-        st.markdown(f.read())
+        st.markdown("""# Dutch T5 models : UL2, T5, ByT5 and Long-T5 🇳🇱🇧🇪
+
+TL;DR: Dutch NLP gets a boost with state-of-the-art T5 models trained on the largest Dutch corpus, mC4, and additional datasets.
+See below for model lists and comparison.
+
+During the [HuggingFace Flax/Jax community week](https://discuss.huggingface.co/t/open-to-the-community-community-week-using-jax-flax-for-nlp-cv/7104) in the summer of 2021,
+I was granted access to Google's TPU Research Cloud (TRC),
+a cloud-based platform for machine learning research and development that provides access to Google's
+Tensor Processing Units (TPUs). My goal was to address the (then) shortage of T5 models for the Dutch language.
+-- T5 is a state-of-the-art AI model architecture that can handle text as input and output,
+making it an ideal tool for NLP tasks such as summarization, translation, and question-answering --
+Since then, with extended access to the TRC, I have been able to train a variety of T5 models for Dutch.
+
+Relevant papers are:
+ 
+* **[Exploring the Limits of Transfer Learning with a Unified Text-to-Text Transformer](https://arxiv.org/abs/1910.10683)** by *Colin Raffel, Noam Shazeer, Adam Roberts, Katherine Lee, Sharan Narang, Michael Matena, Yanqi Zhou, Wei Li, Peter J. Liu*.
+* **[ExT5: Towards Extreme Multi-Task Scaling for Transfer Learning](https://arxiv.org/abs/2111.10952)** by *Vamsi Aribandi, Yi Tay, Tal Schuster, Jinfeng Rao, Huaixiu Steven Zheng, Sanket Vaibhav Mehta, Honglei Zhuang, Vinh Q. Tran, Dara Bahri, Jianmo Ni, Jai Gupta, Kai Hui, Sebastian Ruder, Donald Metzler*.
+* **[Scale Efficiently: Insights from Pre-training and Fine-tuning Transformers](https://arxiv.org/abs/2109.10686)** by *Yi Tay, Mostafa Dehghani, Jinfeng Rao, William Fedus, Samira Abnar, Hyung Won Chung, Sharan Narang, Dani Yogatama, Ashish Vaswani, Donald Metzler*.
+* **[ByT5: Towards a token-free future with pre-trained byte-to-byte models](https://arxiv.org/abs/2105.13626)** by *Linting Xue, Aditya Barua, Noah Constant, Rami Al-Rfou, Sharan Narang, Mihir Kale, Adam Roberts, Colin Raffel*
+* **[LongT5: Efficient Text-To-Text Transformer for Long Sequences](https://arxiv.org/abs/2112.07916)** by *Mandy Guo, Joshua Ainslie, David Uthus, Santiago Ontanon, Jianmo Ni, Yun-Hsuan Sung, Yinfei Yang*
+* **[Scaling Up Models and Data with t5x and seqio](https://arxiv.org/abs/2203.17189)** by *Adam Roberts, Hyung Won Chung, Anselm Levskaya, Gaurav Mishra, James Bradbury, Daniel Andor, Sharan Narang, Brian Lester, Colin Gaffney, Afroz Mohiuddin, Curtis Hawthorne, Aitor Lewkowycz, Alex Salcianu, Marc van Zee, Jacob Austin, Sebastian Goodman, Livio Baldini Soares, Haitang Hu, Sasha Tsvyashchenko, Aakanksha Chowdhery, Jasmijn Bastings, Jannis Bulian, Xavier Garcia, Jianmo Ni, Andrew Chen, Kathleen Kenealy, Jonathan H. Clark, Stephan Lee, Dan Garrette, James Lee-Thorp, Colin Raffel, Noam Shazeer, Marvin Ritter, Maarten Bosma, Alexandre Passos, Jeremy Maitin-Shepard, Noah Fiedel, Mark Omernick, Brennan Saeta, Ryan Sepassi, Alexander Spiridonov, Joshua Newlan, Andrea Gesmundo*
+* **[UL2: Unifying Language Learning Paradigms](https://arxiv.org/abs/2205.05131)** by *Yi Tay, Mostafa Dehghani, Vinh Q. Tran, Xavier Garcia, Jason Wei, Xuezhi Wang, Hyung Won Chung, Dara Bahri, Tal Schuster, Huaixiu Steven Zheng, Denny Zhou, Neil Houlsby, Donald Metzler*
+
+Background on Google's TPU VM's and how to use the Huggingface transformers library to pre-train models can be found
+at the following links
+* https://discuss.huggingface.co/t/open-to-the-community-community-week-using-jax-flax-for-nlp-cv/7104
+* https://github.com/huggingface/transformers/tree/main/examples/research_projects/jax-projects#talks
+
+## Pre-training
+
+### mC4 dataset
+
+Together with the T5 model architecture and SeqIO, the T5 authors also created and released
+the multilingual [mC4 dataset](https://huggingface.co/datasets/allenai/c4).
+It was made available by AllenNLP on the HuggingFace Dataset hub.
+Our team confirmed that the Dutch portion of the mC4 dataset was deduplicated,
+and we cleaned the Dutch portion of the mC4 dataset using [code adapted](https://gitlab.com/yhavinga/c4nlpreproc) from the TensorFlow C4 dataset.
+The resulting [mc4_nl_cleaned](https://huggingface.co/datasets/yhavinga/mc4_nl_cleaned) dataset on the HuggingFace hub
+has configs for several sizes, and also configs for interleaved mixed Dutch and English
+texts, e.g. [micro_en_nl](https://huggingface.co/datasets/yhavinga/mc4_nl_cleaned/viewer/micro_en_nl/train).
+The `_en_nl` configs were added to accommodate multi-language pre-training
+with the Huggingface pre-training script, that accepts only a single dataset as input.
+The full, cleaned Dutch mC4 dataset is 151GB and remains (as of June 2022) the largest available Dutch
+corpus on the HuggingFace Dataset hub.
+
+### Additional books, Wikipedia and Dutch news articles datasets
+
+The `t5_1_1` and `ul2` models have also been trained on Dutch books, the Dutch subset of Wikipedia (2022-03-20),
+the English subset of Wikipedia (2022-03-01), and a subset of "mc4_nl_cleaned" containing only texts
+from Dutch and Belgian newspapers. Mixing in the these datasets was done to bias the model towards
+descriptions of events in the Netherlands and Belgium.
+
+### Pre-Training Objectives
+
+The T5 models are pre-trained using the [span corruption](https://arxiv.org/abs/1910.10683) denoising objective.
+15% of the tokens in the text are masked, and each span
+of masked tokens is replaced with a special token known as a sentinel token, where each span is assigned
+its own sentinel token. The model is then trained to predict for each sentinel token the original text
+that was replaced by the sentinel tokens.
+
+The UL2 models are pre-trained with the [Mixture-of-Denoisers (MoD)](https://arxiv.org/abs/2205.05131) objective, that combines diverse pre-training
+paradigms together. UL2 frames different objective functions for training language models as denoising tasks, where
+the model has to recover missing sub-sequences of a given input. During pre-training it uses a novel mixture-of-denoisers
+that samples from a varied set of such objectives, each with different configurations. UL2 is trained using a mixture of
+three denoising tasks:
+
+1. R-denoising (or regular span corruption), which emulates the standard T5 span corruption objective;
+2. X-denoising (or extreme span corruption); and
+3. S-denoising (or sequential PrefixLM).
+
+### Pre-training software
+
+#### Huggingface [run_t5_mlm_flax.py](https://github.com/huggingface/transformers/blob/main/examples/flax/language-modeling/run_t5_mlm_flax.py)
+
+All models except `t5_1_1` and `ul2` were pre-trained using the Huggingface `run_t5_mlm_flax.py` script.
+This script is a good fit if you want to get a grasp what's needed to pre-train a language model
+with Flax and Jax, since all data preparation, model instantiation, loss function, and training loop are
+contained in a single file.
+
+#### Google's [T5X](https://github.com/google-research/t5x)
+
+The Dutch `t5_1_1` and `ul2` models were pre-trained using T5X. This is a modular framework that can be used for
+pre-training, fine-tuning, and evaluation of T5 models. Because of its modular and pluggable design,
+by only supplying a few configuration and code files, it is possible to pre-train with your own definitions.
+It is even possible to define custom neural network layers and architectures, though I did not do this and only
+pre-trained the default T5 encoder-decoder architecture, and varied only the pre-training objective, and the
+datasets used and mixed with SeqIO.
+
+#### Conversion script from T5X to HF
+
+The T5X models were converted to Huggingface Flax T5 format using a script that was adapted from the
+[T5X checkpoint to HuggingFace Flax conversion script](https://github.com/huggingface/transformers/blob/main/src/transformers/models/t5/convert_t5x_checkpoint_to_flax.py).
+This script was modified to cast weights to bf16, and to also convert to pytorch format.
+For this conversion to be successful, the T5X model had to be saved with `use_gda=False` set in the GIN file.
+
+
+""")
 
     st.markdown(
         """## Evaluation
@@ -291,11 +383,30 @@ mT5 green and the other models black.
         )
     )
 
-    with open("REMARKS.md", "r") as f:
-        st.markdown(f.read())
 
     st.markdown(
-        """### Bfloat16 datatype requires loss regularization
+        """## Miscellaneous remarks
+
+* Use loss regularization when training with `bfloat16` for better results (more info below).
+* Be cautious of the dropout rate in the config.json file and consider training without it.
+  Check in a model's `config.json` what the dropout rate has been set to. Unless you
+  intend to run many epochs on the same data, its worth to try a training run without dropout.
+  If you want to compare losses, be sure to set the dropout rate equal.
+  The smaller models can probably always be trained without.
+* Training with more layers is much slower than you'd expect from the increased model size.
+  It is also more difficult to get batch size and learning rate right. Below is a section
+  about finding the right hyperparameters for the base-36L training.
+* For the translation task, I am not sure that a 'deep-narrow' model (e.g. base-nl36) is better than a normal model
+  of comparable size (e.g. `large`).
+* PyCharm's remote debugging features are useful to inspect variables on either a TPU VM or your deep-learning rig.
+* When increasing the batch size, increase the learning rate. bs * 2 -> lr * sqrt(2) is a good heuristic but mileage may
+  vary.
+* Dataset quality is a key success factor. Do not expect a model to magically turn mediocre data into magic. This holds for 
+  the pre-training data, fine-tuning and also evaluating.
+* Good Bleu score does not necessarily mean fluent text. Evaluation loss on a large translation dataset might be
+  better suited for model comparison, if the models have a tokenizer of comparable size.
+        
+### Bfloat16 datatype requires loss regularization
 
 When training models with `bfloat16` and without loss regularization (default in the HuggingFace pre-training script),
 the training losses would plateau or diverge. The graph below displays the results of different attempts