初始化项目，由ModelHub XC社区提供模型

Model: Finnish-NLP/gpt2-medium-finnish
Source: Original Platform

.gitattributes

@@ -0,0 +1,27 @@
+*.7z filter=lfs diff=lfs merge=lfs -text
+*.arrow filter=lfs diff=lfs merge=lfs -text
+*.bin filter=lfs diff=lfs merge=lfs -text
+*.bin.* filter=lfs diff=lfs merge=lfs -text
+*.bz2 filter=lfs diff=lfs merge=lfs -text
+*.ftz filter=lfs diff=lfs merge=lfs -text
+*.gz filter=lfs diff=lfs merge=lfs -text
+*.h5 filter=lfs diff=lfs merge=lfs -text
+*.joblib filter=lfs diff=lfs merge=lfs -text
+*.lfs.* filter=lfs diff=lfs merge=lfs -text
+*.model filter=lfs diff=lfs merge=lfs -text
+*.msgpack filter=lfs diff=lfs merge=lfs -text
+*.onnx filter=lfs diff=lfs merge=lfs -text
+*.ot filter=lfs diff=lfs merge=lfs -text
+*.parquet filter=lfs diff=lfs merge=lfs -text
+*.pb filter=lfs diff=lfs merge=lfs -text
+*.pt filter=lfs diff=lfs merge=lfs -text
+*.pth filter=lfs diff=lfs merge=lfs -text
+*.rar filter=lfs diff=lfs merge=lfs -text
+saved_model/**/* filter=lfs diff=lfs merge=lfs -text 
+*.tar.* filter=lfs diff=lfs merge=lfs -text
+*.tflite filter=lfs diff=lfs merge=lfs -text
+*.tgz filter=lfs diff=lfs merge=lfs -text
+*.xz filter=lfs diff=lfs merge=lfs -text
+*.zip filter=lfs diff=lfs merge=lfs -text
+*.zstandard filter=lfs diff=lfs merge=lfs -text
+*tfevents* filter=lfs diff=lfs merge=lfs -text

README.md

@@ -0,0 +1,132 @@
+---
+language:
+- fi
+license: apache-2.0
+tags:
+- finnish
+- gpt2
+datasets:
+- Finnish-NLP/mc4_fi_cleaned
+- wikipedia
+widget:
+- text: "Tekstiä tuottava tekoäly on"
+
+---
+
+# GPT-2 medium for Finnish
+
+Pretrained GPT-2 medium model on Finnish language using a causal language modeling (CLM) objective. GPT-2 was introduced in
+[this paper](https://d4mucfpksywv.cloudfront.net/better-language-models/language_models_are_unsupervised_multitask_learners.pdf)
+and first released at [this page](https://openai.com/blog/better-language-models/).
+
+**Note**: this model is 345M parameter variant as in Huggingface's [GPT-2-medium config](https://huggingface.co/gpt2-medium), so not the famous big 1.5B parameter variant by OpenAI. We also have bigger 774M parameter variant [gpt2-large-finnish](https://huggingface.co/Finnish-NLP/gpt2-large-finnish) available which performs better compared to this model.
+
+## Model description
+
+Finnish GPT-2 is a transformers model pretrained on a very large corpus of Finnish data in a self-supervised fashion. This
+means it was pretrained on the raw texts only, with no humans labelling them in any way (which is why it can use lots
+of publicly available data) with an automatic process to generate inputs and labels from those texts. More precisely,
+it was trained to guess the next word in sentences.
+
+More precisely, inputs are sequences of continuous text of a certain length and the targets are the same sequence,
+shifted one token (word or piece of word) to the right. The model uses internally a mask-mechanism to make sure the
+predictions for the token `i` only uses the inputs from `1` to `i` but not the future tokens.
+
+This way, the model learns an inner representation of the Finnish language that can then be used to extract features
+useful for downstream tasks. The model is best at what it was pretrained for however, which is generating texts from a
+prompt.
+
+## Intended uses & limitations
+
+You can use the raw model for text generation or fine-tune it to a downstream task. See the
+[model hub](https://huggingface.co/models?filter=gpt2) to look for fine-tuned versions on a task that interests you.
+
+### How to use
+
+You can use this model directly with a pipeline for text generation:
+
+```python
+>>> from transformers import pipeline
+>>> generator = pipeline('text-generation', model='Finnish-NLP/gpt2-medium-finnish')
+>>> generator("Tekstiä tuottava tekoäly on", max_length=30, num_return_sequences=5)
+
+[{'generated_text': 'Tekstiä tuottava tekoäly on tullut ihmisten arkeen viime vuosina. Se auttaa hahmottamaan ja tulkitsemaan monimutkaisia kokonaisuuksia ja ilmiöitä, joita ihmiset tekevät esimerkiksi ruokakaupassa'},
+ {'generated_text': 'Tekstiä tuottava tekoäly on jo ottanut haltuunsa myös ihmisten käyttämiä sovelluksia ja esimerkiksi pankkipalveluita. Sen vuoksi tekoäly on tärkeä kumppani etenkin yritysten liiketoiminnan kehittämisessä.-'},
+ {'generated_text': 'Tekstiä tuottava tekoäly on tekoälylle luonnollinen valinta, sillä sen avulla voi kommunikoida ihmisten kanssa hyvin pitkälle samalla tavalla kuin tietokoneiden kanssa. Se on kehittynyt muun'},
+ {'generated_text': 'Tekstiä tuottava tekoäly on ihmisen kehittämä tekoäly, jota ei vielä ole pystytty rakentamaan. Tekoäly kykenee toimimaan esimerkiksi matemaattisissa, tilastollisissa ja sosiaalisissa'},
+ {'generated_text': 'Tekstiä tuottava tekoäly on jo niin iso juttu ettei sitä kannata rajoittaakaan. Ja jos se saadaan käyttöön, niin se voi jo pian syrjäyttää perinteisen'}]
+```
+
+Here is how to use this model to get the features of a given text in PyTorch:
+
+```python
+from transformers import GPT2Tokenizer, GPT2Model
+tokenizer = GPT2Tokenizer.from_pretrained('Finnish-NLP/gpt2-medium-finnish')
+model = GPT2Model.from_pretrained('Finnish-NLP/gpt2-medium-finnish')
+text = "Replace me by any text you'd like."
+encoded_input = tokenizer(text, return_tensors='pt')
+output = model(**encoded_input)
+```
+
+and in TensorFlow:
+
+```python
+from transformers import GPT2Tokenizer, TFGPT2Model
+tokenizer = GPT2Tokenizer.from_pretrained('Finnish-NLP/gpt2-medium-finnish')
+model = TFGPT2Model.from_pretrained('Finnish-NLP/gpt2-medium-finnish', from_pt=True)
+text = "Replace me by any text you'd like."
+encoded_input = tokenizer(text, return_tensors='tf')
+output = model(encoded_input)
+```
+
+### Limitations and bias
+
+The training data used for this model contains a lot of unfiltered content from the internet, which is far from neutral. Therefore, the model can have biased predictions. This bias will also affect all fine-tuned versions of this model.
+
+As with all language models, it is hard to predict in advance how the Finnish GPT-2 will respond to particular prompts and offensive content may occur without warning. We recommend having a human curate or filter the outputs before releasing them, both to censor undesirable content and to improve the quality of the results.
+
+## Training data
+
+This Finnish GPT-2 model was pretrained on the combination of six datasets:
+- [mc4_fi_cleaned](https://huggingface.co/datasets/Finnish-NLP/mc4_fi_cleaned), the dataset mC4 is a multilingual colossal, cleaned version of Common Crawl's web crawl corpus. We used the Finnish subset of the mC4 dataset and further cleaned it with our own text data cleaning codes (check the dataset repo).
+- [wikipedia](https://huggingface.co/datasets/wikipedia) We used the Finnish subset of the wikipedia (August 2021) dataset
+- [Yle Finnish News Archive 2011-2018](http://urn.fi/urn:nbn:fi:lb-2017070501)
+- [Yle Finnish News Archive 2019-2020](http://urn.fi/urn:nbn:fi:lb-2021050401)
+- [Finnish News Agency Archive (STT)](http://urn.fi/urn:nbn:fi:lb-2018121001)
+- [The Suomi24 Sentences Corpus](http://urn.fi/urn:nbn:fi:lb-2020021803)
+
+Raw datasets were cleaned to filter out bad quality and non-Finnish examples. Together these cleaned datasets were around 84GB of text.
+
+## Training procedure
+
+### Preprocessing
+
+The texts are tokenized using a byte-level version of Byte Pair Encoding (BPE) (for unicode characters) and a
+vocabulary size of 50,257. The inputs are sequences of 512 consecutive tokens.
+
+### Pretraining
+
+The model was trained on TPUv3-8 VM, sponsored by the [Google TPU Research Cloud](https://sites.research.google/trc/about/), for 360k steps (a bit over 1 epoch, 128 batch size). The optimizer used was a AdamW with learning rate 1e-4, learning rate warmup for 4000 steps and cosine decay of the learning rate after.
+
+
+## Evaluation results
+
+Evaluation was done using the *validation* split of the [mc4_fi_cleaned](https://huggingface.co/datasets/Finnish-NLP/mc4_fi_cleaned) dataset with [Perplexity](https://huggingface.co/course/chapter7/3#perplexity-for-language-models) (smaller score the better) as the evaluation metric. As seen from the table below, this model (the first row of the table) performs better than our smaller [gpt2-finnish](https://huggingface.co/Finnish-NLP/gpt2-finnish) model variant but loses to our bigger [gpt2-large-finnish](https://huggingface.co/Finnish-NLP/gpt2-large-finnish) model.
+
+|                                          | Perplexity |
+|------------------------------------------|------------|
+|Finnish-NLP/gpt2-medium-finnish           |34.08       |
+|Finnish-NLP/gpt2-finnish                  |44.19       |
+|Finnish-NLP/gpt2-large-finnish            |**30.74**   |
+
+## Acknowledgements
+
+This project would not have been possible without compute generously provided by Google through the
+[TPU Research Cloud](https://sites.research.google/trc/).
+
+## Team Members
+
+- Aapo Tanskanen, [Hugging Face profile](https://huggingface.co/aapot), [LinkedIn profile](https://www.linkedin.com/in/aapotanskanen/)
+- Rasmus Toivanen, [Hugging Face profile](https://huggingface.co/RASMUS), [LinkedIn profile](https://www.linkedin.com/in/rasmustoivanen/)
+
+Feel free to contact us for more details 🤗

added_tokens.json

@@ -0,0 +1 @@
+{"<|endoftext|>": 50257}

config.json

@@ -0,0 +1,41 @@
+{
+  "_name_or_path": "./",
+  "activation_function": "gelu_new",
+  "architectures": [
+    "GPT2LMHeadModel"
+  ],
+  "attn_pdrop": 0.1,
+  "bos_token_id": 50256,
+  "embd_pdrop": 0.1,
+  "eos_token_id": 50256,
+  "initializer_range": 0.02,
+  "layer_norm_epsilon": 1e-05,
+  "model_type": "gpt2",
+  "n_ctx": 1024,
+  "n_embd": 1024,
+  "n_head": 16,
+  "n_inner": null,
+  "n_layer": 24,
+  "n_positions": 1024,
+  "n_special": 0,
+  "predict_special_tokens": true,
+  "reorder_and_upcast_attn": false,
+  "resid_pdrop": 0.1,
+  "scale_attn_by_inverse_layer_idx": false,
+  "scale_attn_weights": true,
+  "summary_activation": null,
+  "summary_first_dropout": 0.1,
+  "summary_proj_to_labels": true,
+  "summary_type": "cls_index",
+  "summary_use_proj": true,
+  "task_specific_params": {
+    "text-generation": {
+      "do_sample": true,
+      "max_length": 100
+    }
+  },
+  "torch_dtype": "float32",
+  "transformers_version": "4.16.1",
+  "use_cache": true,
+  "vocab_size": 50257
+}

flax_model.msgpack

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:cc280704c27ddf0e5800394ac06001b34cf42b137e8173c48b10b731e9d68b45
+size 1419302302

flax_model_to_pytorch.py

@@ -0,0 +1,27 @@
+from transformers import AutoModelForCausalLM, FlaxAutoModelForCausalLM, AutoTokenizer
+import torch
+import numpy as np
+import jax
+import jax.numpy as jnp
+
+def to_f32(t):
+    return jax.tree_map(lambda x: x.astype(jnp.float32) if x.dtype == jnp.bfloat16 else x, t)
+
+jax.config.update('jax_platform_name', 'cpu')
+MODEL_PATH = "./"
+model = FlaxAutoModelForCausalLM.from_pretrained(MODEL_PATH)
+model.params = to_f32(model.params)
+model.save_pretrained(MODEL_PATH)
+
+pt_model = AutoModelForCausalLM.from_pretrained(
+    MODEL_PATH, from_flax=True).to('cpu')
+
+input_ids = np.asarray(2 * [128 * [0]], dtype=np.int32)
+input_ids_pt = torch.tensor(input_ids)
+
+logits_pt = pt_model(input_ids_pt).logits
+print(logits_pt)
+logits_fx = model(input_ids).logits
+print(logits_fx)
+
+pt_model.save_pretrained(MODEL_PATH)

pytorch_model.bin

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:5070fde5ed7d34c5230f1c11aab6352b27c8e33a31ae77170bf4b44486d038d7
+size 1444576537

replace_token_script.py

@@ -0,0 +1,80 @@
+''''This script was used to replace the final index of tokenizer.json and vocab.json
+     with "<|endoftext|>" token. Also reassociate the corresponding merges'''
+
+import json
+
+tokenizer_path = 'tokenizer.json'
+model_config_path = 'config.json'
+vocab_path = 'vocab.json'
+
+with open(vocab_path, "r") as f:
+    vocab_data = json.load(f)
+
+with open(tokenizer_path, "r") as f:
+    tokenizer_data = json.load(f)
+
+with open(model_config_path, "r") as f:
+    model_config = json.load(f)
+
+model_vocab_size = model_config['vocab_size']
+tokenizer_vocab = tokenizer_data['model']['vocab']
+
+mergeslength = len(tokenizer_data['model']['merges'])
+
+#readjust added_tokens 'id' to model_vocab_size - 1
+tokenizer_data['added_tokens'][-1]['id'] = model_vocab_size - 1
+
+final_index = model_vocab_size - 1
+eos = '<|endoftext|>'
+
+#retrieve the key of final index
+old_key_final_index_tokenizer = list(tokenizer_data['model']['vocab'].keys())[final_index]
+old_key_final_index_vocab = list(vocab_data.keys())[final_index]
+old_key_final_index_vocab_min2 = list(vocab_data.keys())[final_index - 1]
+old_key_final_index_tokenizer_merges = tokenizer_data['model']['merges'][mergeslength - 1]
+
+print(f"old_key_final_index_tokenizer = {old_key_final_index_tokenizer}")
+print(f"old_key_final_index_vocab = {old_key_final_index_vocab}")
+print(f"old_key_final_index_vocab_min2 = {old_key_final_index_vocab_min2}")
+print(f"old_key_final_index_tokenizer_merges = {old_key_final_index_tokenizer_merges}")
+
+#replace old key with new key
+tokenizer_data['model']['vocab']['<|endoftext|>'] = tokenizer_data['model']['vocab'][old_key_final_index_tokenizer]
+vocab_data[eos] = vocab_data[old_key_final_index_vocab]
+
+#replace the final merges idx with vocab_data - 1
+tokenizer_data['model']['merges'] = tokenizer_data['model']['merges'][: mergeslength - 1]
+
+
+#delete old key
+del tokenizer_data['model']['vocab'][old_key_final_index_tokenizer]
+del vocab_data[old_key_final_index_vocab]
+
+#check updated key
+old_key_final_index_tokenizer = list(tokenizer_data['model']['vocab'].keys())[final_index]
+old_key_final_index_vocab = list(vocab_data.keys())[final_index]
+old_key_final_index_tokenizer_merges = tokenizer_data['model']['merges'][mergeslength - 2]
+
+print(len(tokenizer_data['model']['merges']))
+print()
+print(f"updated old_key_final_index_tokenizer = {old_key_final_index_tokenizer}")
+print(f"updated old_key_final_index_vocab = {old_key_final_index_vocab}")
+print(f"updated old_key_final_index_tokenizer_merges = {old_key_final_index_tokenizer_merges}")
+
+with open(tokenizer_path, "w")as f:
+    json.dump(tokenizer_data, f)
+
+with open(vocab_path, "w")as f:
+    json.dump(vocab_data, f)
+
+with open('merges.txt') as f:
+    lines = f.readlines()
+
+with open("merges.txt", "w") as f:
+    for i in range(len(lines) - 1):
+        f.write(lines[i])
+
+with open('merges.txt') as f:
+    newlines = f.readlines()
+
+print(f"newlines[len(newlines) - 1] = {newlines[len(newlines) - 1]}")

run_clm_flax.py

@@ -0,0 +1,892 @@
+#!/usr/bin/env python
+# coding=utf-8
+# Copyright 2021 The HuggingFace Team All rights reserved.
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+"""
+Pre-training/Fine-tuning the library models for causal language modeling (GPT, GPT-2, CTRL, ...) on a text file or a dataset.
+
+Here is the full list of checkpoints on the hub that can be fine-tuned by this script:
+https://huggingface.co/models?filter=text-generation
+"""
+# You can also adapt this script on your own causal language modeling task. Pointers for this are left as comments.
+
+import json
+import logging
+import math
+import os
+import sys
+import time
+import gc
+from dataclasses import asdict, dataclass, field
+from enum import Enum
+from itertools import chain
+from pathlib import Path
+from typing import Callable, Optional
+
+import datasets
+import numpy as np
+from datasets import Dataset, load_dataset, load_from_disk
+from tqdm import tqdm
+
+import jax
+import jax.numpy as jnp
+import optax
+import transformers
+from flax import jax_utils, traverse_util
+from flax.jax_utils import unreplicate
+from flax.training import train_state
+from flax.training.common_utils import get_metrics, onehot, shard, shard_prng_key
+from huggingface_hub import Repository
+from transformers import (
+    CONFIG_MAPPING,
+    FLAX_MODEL_FOR_CAUSAL_LM_MAPPING,
+    AutoConfig,
+    AutoTokenizer,
+    FlaxAutoModelForCausalLM,
+    HfArgumentParser,
+    is_tensorboard_available,
+    set_seed,
+)
+from transformers.file_utils import get_full_repo_name
+from transformers.testing_utils import CaptureLogger
+
+from distributed_shampoo import distributed_shampoo, GraftingType
+
+
+logger = logging.getLogger(__name__)
+
+MODEL_CONFIG_CLASSES = list(FLAX_MODEL_FOR_CAUSAL_LM_MAPPING.keys())
+MODEL_TYPES = tuple(conf.model_type for conf in MODEL_CONFIG_CLASSES)
+
+
+@dataclass
+class TrainingArguments:
+    output_dir: str = field(
+        metadata={"help": "The output directory where the model predictions and checkpoints will be written."},
+    )
+    overwrite_output_dir: bool = field(
+        default=False,
+        metadata={
+            "help": (
+                "Overwrite the content of the output directory. "
+                "Use this to continue training if output_dir points to a checkpoint directory."
+            )
+        },
+    )
+    do_train: bool = field(default=False, metadata={"help": "Whether to run training."})
+    do_eval: bool = field(default=False, metadata={"help": "Whether to run eval on the dev set."})
+    per_device_train_batch_size: int = field(
+        default=8, metadata={"help": "Batch size per GPU/TPU core/CPU for training."}
+    )
+    per_device_eval_batch_size: int = field(
+        default=8, metadata={"help": "Batch size per GPU/TPU core/CPU for evaluation."}
+    )
+    gradient_accumulation_steps: int = field(
+        default=1,
+        metadata={
+            "help": "Number of updates steps to accumulate before performing a backward/update pass."
+        },
+    )
+    learning_rate: float = field(default=5e-5, metadata={"help": "The initial learning rate for AdamW."})
+    weight_decay: float = field(default=0.0, metadata={"help": "Weight decay for AdamW if we apply some."})
+    adam_beta1: float = field(default=0.9, metadata={"help": "Beta1 for AdamW optimizer"})
+    adam_beta2: float = field(default=0.999, metadata={"help": "Beta2 for AdamW optimizer"})
+    adam_epsilon: float = field(default=1e-8, metadata={"help": "Epsilon for AdamW optimizer."})
+    adafactor: bool = field(default=False, metadata={"help": "Whether or not to replace AdamW by Adafactor."})
+    distributed_shampoo: bool = field(
+        default=False, metadata={"help": "Use Distributed Shampoo optimizer instead of AdamW."},
+    )
+    quantize_shampoo: bool = field(
+        default=False, metadata={"help": "Quantize Distributed Shampoo optimizer."},
+    )
+    num_train_epochs: float = field(default=3.0, metadata={"help": "Total number of training epochs to perform."})
+    warmup_steps: int = field(default=0, metadata={"help": "Linear warmup over warmup_steps."})
+    warmup_ratio: float = field(default=0.0, metadata={"help": "Linear warmup ratio of total train steps."})
+    cosine_decay: bool = field(
+        default=False, metadata={"help": "Whether or not to use cosine decay instead of the basic linear decay schedule."}
+    )
+    gradient_clipping: bool = field(
+        default=False, metadata={"help": "Whether or not to use gradient clipping."}
+    )
+    logging_steps: int = field(default=500, metadata={"help": "Log every X updates steps."})
+    save_steps: int = field(default=500, metadata={"help": "Save checkpoint every X updates steps."})
+    eval_steps: int = field(default=None, metadata={"help": "Run an evaluation every X steps."})
+    seed: int = field(default=42, metadata={"help": "Random seed that will be set at the beginning of training."})
+    push_to_hub: bool = field(
+        default=False, metadata={"help": "Whether or not to upload the trained model to the model hub after training."}
+    )
+    hub_model_id: str = field(
+        default=None, metadata={"help": "The name of the repository to keep in sync with the local `output_dir`."}
+    )
+    hub_token: str = field(default=None, metadata={"help": "The token to use to push to the Model Hub."})
+
+    def __post_init__(self):
+        if self.output_dir is not None:
+            self.output_dir = os.path.expanduser(self.output_dir)
+
+    def to_dict(self):
+        """
+        Serializes this instance while replace `Enum` by their values (for JSON serialization support). It obfuscates
+        the token values by removing their value.
+        """
+        d = asdict(self)
+        for k, v in d.items():
+            if isinstance(v, Enum):
+                d[k] = v.value
+            if isinstance(v, list) and len(v) > 0 and isinstance(v[0], Enum):
+                d[k] = [x.value for x in v]
+            if k.endswith("_token"):
+                d[k] = f"<{k.upper()}>"
+        return d
+
+
+@dataclass
+class ModelArguments:
+    """
+    Arguments pertaining to which model/config/tokenizer we are going to fine-tune, or train from scratch.
+    """
+
+    model_name_or_path: Optional[str] = field(
+        default=None,
+        metadata={
+            "help": "The model checkpoint for weights initialization."
+            "Don't set if you want to train a model from scratch."
+        },
+    )
+    model_type: Optional[str] = field(
+        default=None,
+        metadata={"help": "If training from scratch, pass a model type from the list: " + ", ".join(MODEL_TYPES)},
+    )
+    config_name: Optional[str] = field(
+        default=None, metadata={"help": "Pretrained config name or path if not the same as model_name"}
+    )
+    tokenizer_name: Optional[str] = field(
+        default=None, metadata={"help": "Pretrained tokenizer name or path if not the same as model_name"}
+    )
+    cache_dir: Optional[str] = field(
+        default=None, metadata={"help": "Where do you want to store the pretrained models downloaded from s3"}
+    )
+    use_fast_tokenizer: bool = field(
+        default=True,
+        metadata={"help": "Whether to use one of the fast tokenizer (backed by the tokenizers library) or not."},
+    )
+    dtype: Optional[str] = field(
+        default="float32",
+        metadata={
+            "help": "Floating-point format in which the model weights should be initialized and trained. Choose one of `[float32, float16, bfloat16]`."
+        },
+    )
+
+
+@dataclass
+class DataTrainingArguments:
+    """
+    Arguments pertaining to what data we are going to input our model for training and eval.
+    """
+
+    dataset_name: Optional[str] = field(
+        default=None, metadata={"help": "The name of the dataset to use (via the datasets library)."}
+    )
+    dataset_config_name: Optional[str] = field(
+        default=None, metadata={"help": "The configuration name of the dataset to use (via the datasets library)."}
+    )
+    dataset_filepath: Optional[str] = field(
+        default=None, metadata={"help": "Filepath to locally saved HF Dataset (with 'dataset.save_to_disk' method) to use for training"}
+    )
+    train_file: Optional[str] = field(default=None, metadata={"help": "The input training data file (a text file)."})
+    validation_file: Optional[str] = field(
+        default=None,
+        metadata={"help": "An optional input evaluation data file to evaluate the perplexity on (a text file)."},
+    )
+    max_train_samples: Optional[int] = field(
+        default=None,
+        metadata={
+            "help": "For debugging purposes or quicker training, truncate the number of training examples to this "
+            "value if set."
+        },
+    )
+    max_eval_samples: Optional[int] = field(
+        default=None,
+        metadata={
+            "help": "For debugging purposes or quicker training, truncate the number of evaluation examples to this "
+            "value if set."
+        },
+    )
+    overwrite_cache: bool = field(
+        default=False, metadata={"help": "Overwrite the cached training and evaluation sets"}
+    )
+    validation_split_percentage: Optional[int] = field(
+        default=5,
+        metadata={
+            "help": "The percentage of the train set used as validation set in case there's no validation split"
+        },
+    )
+    block_size: Optional[int] = field(
+        default=None,
+        metadata={
+            "help": "Optional input sequence length after tokenization. "
+            "The training dataset will be truncated in block of this size for training. "
+            "Default to the model max input length for single sentence inputs (take into account special tokens)."
+        },
+    )
+    overwrite_cache: bool = field(
+        default=False, metadata={"help": "Overwrite the cached training and evaluation sets"}
+    )
+    preprocessing_num_workers: Optional[int] = field(
+        default=None,
+        metadata={"help": "The number of processes to use for the preprocessing."},
+    )
+    keep_linebreaks: bool = field(
+        default=True, metadata={"help": "Whether to keep line breaks when using TXT files or not."}
+    )
+
+    def __post_init__(self):
+        if self.dataset_name is None and self.train_file is None and self.dataset_filepath is None and self.validation_file is None:
+            raise ValueError("Need either a dataset name or a training/validation file.")
+        else:
+            if self.train_file is not None:
+                extension = self.train_file.split(".")[-1]
+                assert extension in ["csv", "json", "txt"], "`train_file` should be a csv, a json or a txt file."
+            if self.validation_file is not None:
+                extension = self.validation_file.split(".")[-1]
+                assert extension in ["csv", "json", "txt"], "`validation_file` should be a csv, a json or a txt file."
+
+
+class TrainState(train_state.TrainState):
+    dropout_rng: jnp.ndarray
+
+    def replicate(self):
+        return jax_utils.replicate(self).replace(dropout_rng=shard_prng_key(self.dropout_rng))
+
+
+def data_loader(rng: jax.random.PRNGKey, dataset: Dataset, batch_size: int, shuffle: bool = False):
+    """
+    Returns batches of size `batch_size` from truncated `dataset`, sharded over all local devices.
+    Shuffle batches if `shuffle` is `True`.
+    """
+    steps_per_epoch = len(dataset) // batch_size
+
+    if shuffle:
+        batch_idx = jax.random.permutation(rng, len(dataset))
+    else:
+        batch_idx = jnp.arange(len(dataset))
+
+    batch_idx = batch_idx[: steps_per_epoch * batch_size]  # Skip incomplete batch.
+    batch_idx = batch_idx.reshape((steps_per_epoch, batch_size))
+
+    for idx in batch_idx:
+        batch = dataset[idx]
+        batch = {k: np.array(v) for k, v in batch.items()}
+
+        yield batch
+
+
+def write_train_metric(summary_writer, train_metrics, train_time, step):
+    summary_writer.scalar("train_time", train_time, step)
+
+    train_metrics = get_metrics(train_metrics)
+    for key, vals in train_metrics.items():
+        tag = f"train_{key}"
+        for i, val in enumerate(vals):
+            summary_writer.scalar(tag, val, step - len(vals) + i + 1)
+
+
+def write_eval_metric(summary_writer, eval_metrics, step):
+    for metric_name, value in eval_metrics.items():
+        summary_writer.scalar(f"eval_{metric_name}", value, step)
+
+
+def create_linear_learning_rate_fn(
+    train_ds_size: int, train_batch_size: int, num_train_epochs: int, num_warmup_steps: int, learning_rate: float
+) -> Callable[[int], jnp.array]:
+    """Returns a linear warmup, linear decay learning rate function."""
+    steps_per_epoch = train_ds_size // train_batch_size
+    num_train_steps = steps_per_epoch * num_train_epochs
+    warmup_fn = optax.linear_schedule(init_value=0.0, end_value=learning_rate, transition_steps=num_warmup_steps)
+    decay_fn = optax.linear_schedule(
+        init_value=learning_rate, end_value=0, transition_steps=num_train_steps - num_warmup_steps
+    )
+    schedule_fn = optax.join_schedules(schedules=[warmup_fn, decay_fn], boundaries=[num_warmup_steps])
+    return schedule_fn
+
+def create_cosine_learning_rate_fn(
+    train_ds_size: int, train_batch_size: int, num_train_epochs: int, num_warmup_steps: int, learning_rate: float
+) -> Callable[[int], jnp.array]:
+    """Returns a linear warmup, cosine decay learning rate function."""
+    steps_per_epoch = train_ds_size // train_batch_size
+    num_train_steps = steps_per_epoch * num_train_epochs
+    warmup_fn = optax.linear_schedule(init_value=0.0, end_value=learning_rate, transition_steps=num_warmup_steps)
+    decay_fn = optax.cosine_decay_schedule(
+        init_value=learning_rate, decay_steps=num_train_steps - num_warmup_steps, alpha=0.1
+    )
+    schedule_fn = optax.join_schedules(schedules=[warmup_fn, decay_fn], boundaries=[num_warmup_steps])
+    return schedule_fn
+
+
+def main():
+    # See all possible arguments in src/transformers/training_args.py
+    # or by passing the --help flag to this script.
+    # We now keep distinct sets of args, for a cleaner separation of concerns.
+
+    parser = HfArgumentParser((ModelArguments, DataTrainingArguments, TrainingArguments))
+    if len(sys.argv) == 2 and sys.argv[1].endswith(".json"):
+        # If we pass only one argument to the script and it's the path to a json file,
+        # let's parse it to get our arguments.
+        model_args, data_args, training_args = parser.parse_json_file(json_file=os.path.abspath(sys.argv[1]))
+    else:
+        model_args, data_args, training_args = parser.parse_args_into_dataclasses()
+
+    if (
+        os.path.exists(training_args.output_dir)
+        and os.listdir(training_args.output_dir)
+        and training_args.do_train
+        and not training_args.overwrite_output_dir
+    ):
+        raise ValueError(
+            f"Output directory ({training_args.output_dir}) already exists and is not empty."
+            "Use --overwrite_output_dir to overcome."
+        )
+
+    # Make one log on every process with the configuration for debugging.
+    logging.basicConfig(
+        format="%(asctime)s - %(levelname)s - %(name)s - %(message)s",
+        datefmt="%m/%d/%Y %H:%M:%S",
+        level=logging.INFO,
+    )
+    # Setup logging, we only want one process per machine to log things on the screen.
+    logger.setLevel(logging.INFO if jax.process_index() == 0 else logging.ERROR)
+    if jax.process_index() == 0:
+        datasets.utils.logging.set_verbosity_warning()
+        transformers.utils.logging.set_verbosity_info()
+    else:
+        datasets.utils.logging.set_verbosity_error()
+        transformers.utils.logging.set_verbosity_error()
+
+    # Set the verbosity to info of the Transformers logger (on main process only):
+    logger.info(f"Training/evaluation parameters {training_args}")
+
+    # Set seed before initializing model.
+    set_seed(training_args.seed)
+
+    # Handle the repository creation
+    if training_args.push_to_hub:
+        if training_args.hub_model_id is None:
+            repo_name = get_full_repo_name(
+                Path(training_args.output_dir).absolute().name, token=training_args.hub_token
+            )
+        else:
+            repo_name = training_args.hub_model_id
+        repo = Repository(training_args.output_dir, clone_from=repo_name)
+
+    #  Get the datasets: you can either provide your own CSV/JSON/TXT training and evaluation files (see below)
+    # or just provide the name of one of the public datasets available on the hub at https://huggingface.co/datasets/
+    # (the dataset will be downloaded automatically from the datasets Hub).
+    #
+    # For CSV/JSON files, this script will use the column called 'text' or the first column if no column called
+    # 'text' is found. You can easily tweak this behavior (see below).
+    #
+    # In distributed training, the load_dataset function guarantees that only one local process can concurrently
+    # download the dataset.
+    if data_args.dataset_name is not None:
+        # Downloading and loading a dataset from the hub.
+        dataset = load_dataset(
+            data_args.dataset_name, data_args.dataset_config_name, cache_dir=model_args.cache_dir, keep_in_memory=False
+        )
+
+        if "validation" not in dataset.keys():
+            dataset["validation"] = load_dataset(
+                data_args.dataset_name,
+                data_args.dataset_config_name,
+                split=f"train[:{data_args.validation_split_percentage}%]",
+                cache_dir=model_args.cache_dir,
+            )
+            dataset["train"] = load_dataset(
+                data_args.dataset_name,
+                data_args.dataset_config_name,
+                split=f"train[{data_args.validation_split_percentage}%:]",
+                cache_dir=model_args.cache_dir,
+            )
+
+    elif data_args.dataset_filepath is not None:
+        # Loading a dataset from local file.
+        dataset = load_from_disk(data_args.dataset_filepath)
+        if "validation" not in dataset.keys():
+            dataset = datasets.train_test_split(test_size=data_args.validation_split_percentage/100)
+            dataset["validation"] = dataset["test"]
+            del dataset["test"]
+
+    else:
+        data_files = {}
+        dataset_args = {}
+        if data_args.train_file is not None:
+            data_files["train"] = data_args.train_file
+        if data_args.validation_file is not None:
+            data_files["validation"] = data_args.validation_file
+        extension = data_args.train_file.split(".")[-1]
+        if extension == "txt":
+            extension = "text"
+            dataset_args["keep_linebreaks"] = data_args.keep_linebreaks
+        dataset = load_dataset(extension, data_files=data_files, cache_dir=model_args.cache_dir, **dataset_args)
+
+        if "validation" not in dataset.keys():
+            dataset["validation"] = load_dataset(
+                extension,
+                data_files=data_files,
+                split=f"train[:{data_args.validation_split_percentage}%]",
+                cache_dir=model_args.cache_dir,
+                **dataset_args,
+            )
+            dataset["train"] = load_dataset(
+                extension,
+                data_files=data_files,
+                split=f"train[{data_args.validation_split_percentage}%:]",
+                cache_dir=model_args.cache_dir,
+                **dataset_args,
+            )
+    # See more about loading any type of standard or custom dataset (from files, python dict, pandas DataFrame, etc) at
+    # https://huggingface.co/docs/datasets/loading_datasets.html.
+
+    # Load pretrained model and tokenizer
+
+    # Distributed training:
+    # The .from_pretrained methods guarantee that only one local process can concurrently
+    # download model & vocab.
+    if model_args.config_name:
+        config = AutoConfig.from_pretrained(model_args.config_name, cache_dir=model_args.cache_dir)
+    elif model_args.model_name_or_path:
+        config = AutoConfig.from_pretrained(model_args.model_name_or_path, cache_dir=model_args.cache_dir)
+    else:
+        config = CONFIG_MAPPING[model_args.model_type]()
+        logger.warning("You are instantiating a new config instance from scratch.")
+
+    if model_args.tokenizer_name:
+        tokenizer = AutoTokenizer.from_pretrained(
+            model_args.tokenizer_name, cache_dir=model_args.cache_dir, use_fast=model_args.use_fast_tokenizer
+        )
+    elif model_args.model_name_or_path:
+        tokenizer = AutoTokenizer.from_pretrained(
+            model_args.model_name_or_path, cache_dir=model_args.cache_dir, use_fast=model_args.use_fast_tokenizer
+        )
+    else:
+        raise ValueError(
+            "You are instantiating a new tokenizer from scratch. This is not supported by this script."
+            "You can do it from another script, save it, and load it from here, using --tokenizer_name."
+        )
+
+    if tokenizer.pad_token is None:
+        tokenizer.pad_token = tokenizer.eos_token
+
+    if model_args.model_name_or_path:
+        model = FlaxAutoModelForCausalLM.from_pretrained(
+            model_args.model_name_or_path, config=config, seed=training_args.seed, dtype=getattr(jnp, model_args.dtype)
+        )
+    else:
+        model = FlaxAutoModelForCausalLM.from_config(
+            config, seed=training_args.seed, dtype=getattr(jnp, model_args.dtype)
+        )
+
+    # Preprocessing the datasets.
+    # First we tokenize all the texts.
+    if training_args.do_train:
+        column_names = dataset["train"].column_names
+    else:
+        column_names = dataset["validation"].column_names
+    text_column_name = "text" if "text" in column_names else column_names[0]
+
+    # since this will be pickled to avoid _LazyModule error in Hasher force logger loading before tokenize_function
+    tok_logger = transformers.utils.logging.get_logger("transformers.tokenization_utils_base")
+
+    def tokenize_function(examples):
+        with CaptureLogger(tok_logger) as cl:
+            output = tokenizer(examples[text_column_name])
+        # clm input could be much much longer than block_size
+        if "Token indices sequence length is longer than the" in cl.out:
+            tok_logger.warning(
+                "^^^^^^^^^^^^^^^^ Please ignore the warning above - this long input will be chunked into smaller bits before being passed to the model."
+            )
+        return output
+
+    tokenized_datasets = dataset.map(
+        tokenize_function,
+        batched=True,
+        num_proc=data_args.preprocessing_num_workers,
+        remove_columns=column_names,
+        load_from_cache_file=not data_args.overwrite_cache,
+    )
+
+    if data_args.block_size is None:
+        block_size = tokenizer.model_max_length
+        if block_size > config.max_position_embeddings:
+            logger.warning(
+                f"The tokenizer picked seems to have a very large `model_max_length` ({tokenizer.model_max_length}). "
+                "Picking 1024 instead. You can change that default value by passing --block_size xxx."
+            )
+            block_size = 1024
+    else:
+        if data_args.block_size > tokenizer.model_max_length:
+            logger.warning(
+                f"The block_size passed ({data_args.block_size}) is larger than the maximum length for the model"
+                f"({tokenizer.model_max_length}). Using block_size={tokenizer.model_max_length}."
+            )
+        block_size = min(data_args.block_size, tokenizer.model_max_length)
+
+    # Main data processing function that will concatenate all texts from our dataset and generate chunks of block_size.
+    def group_texts(examples):
+        # Concatenate all texts.
+        concatenated_examples = {k: list(chain(*examples[k])) for k in examples.keys()}
+        total_length = len(concatenated_examples[list(examples.keys())[0]])
+        # We drop the small remainder, we could add padding if the model supported it instead of this drop, you can
+        # customize this part to your needs.
+        if total_length >= block_size:
+            total_length = (total_length // block_size) * block_size
+        # Split by chunks of max_len.
+        result = {
+            k: [t[i : i + block_size] for i in range(0, total_length, block_size)]
+            for k, t in concatenated_examples.items()
+        }
+        result["labels"] = result["input_ids"].copy()
+        return result
+
+    # Note that with `batched=True`, this map processes 1,000 texts together, so group_texts throws away a remainder
+    # for each of those groups of 1,000 texts. You can adjust that batch_size here but a higher value might be slower
+    # to preprocess.
+    #
+    # To speed up this part, we use multiprocessing. See the documentation of the map method for more information:
+    # https://huggingface.co/docs/datasets/package_reference/main_classes.html#datasets.Dataset.map
+
+    lm_datasets = tokenized_datasets.map(
+        group_texts,
+        batched=True,
+        num_proc=data_args.preprocessing_num_workers,
+        load_from_cache_file=not data_args.overwrite_cache,
+    )
+
+    if training_args.do_train:
+        if "train" not in tokenized_datasets:
+            raise ValueError("--do_train requires a train dataset")
+        train_dataset = lm_datasets["train"]
+        if data_args.max_train_samples is not None:
+            train_dataset = train_dataset.select(range(data_args.max_train_samples))
+        
+        # test to see that tokenization worked
+        detokenized_example = tokenizer.decode(train_dataset[0]["input_ids"])
+        logger.info(f"Detokenized example: {detokenized_example}")
+        detokenized_example = tokenizer.decode(train_dataset[-1]["input_ids"])
+        logger.info(f"Detokenized example 2: {detokenized_example}")
+
+    if training_args.do_eval:
+        if "validation" not in tokenized_datasets:
+            raise ValueError("--do_eval requires a validation dataset")
+        eval_dataset = lm_datasets["validation"]
+        if data_args.max_eval_samples is not None:
+            eval_dataset = eval_dataset.select(range(data_args.max_eval_samples))
+
+    # Enable tensorboard only on the master node
+    has_tensorboard = is_tensorboard_available()
+    if has_tensorboard and jax.process_index() == 0:
+        try:
+            from flax.metrics.tensorboard import SummaryWriter
+
+            summary_writer = SummaryWriter(log_dir=Path(os.path.join(training_args.output_dir, "runs")))
+        except ImportError as ie:
+            has_tensorboard = False
+            logger.warning(
+                f"Unable to display metrics through TensorBoard because some package are not installed: {ie}"
+            )
+    else:
+        logger.warning(
+            "Unable to display metrics through TensorBoard because the package is not installed: "
+            "Please run pip install tensorboard to enable."
+        )
+
+    # Initialize our training
+    rng = jax.random.PRNGKey(training_args.seed)
+    rng, dropout_rng = jax.random.split(rng)
+
+    # Store some constant
+    num_epochs = int(training_args.num_train_epochs)
+    train_batch_size = int(training_args.per_device_train_batch_size) * jax.device_count() * training_args.gradient_accumulation_steps
+    eval_batch_size = int(training_args.per_device_eval_batch_size) * jax.device_count()
+    steps_per_epoch = len(train_dataset) // train_batch_size
+    total_train_steps = steps_per_epoch * num_epochs
+
+    if training_args.warmup_ratio > 0:
+        warmup_steps = int(total_train_steps * training_args.warmup_ratio)
+    else:
+        warmup_steps = training_args.warmup_steps
+
+    # Create learning rate schedule
+    if training_args.cosine_decay:
+        lr_schedule_fn = create_cosine_learning_rate_fn(
+            len(train_dataset),
+            train_batch_size,
+            training_args.num_train_epochs,
+            warmup_steps,
+            training_args.learning_rate,
+        )
+    else:
+        lr_schedule_fn = create_linear_learning_rate_fn(
+            len(train_dataset),
+            train_batch_size,
+            training_args.num_train_epochs,
+            warmup_steps,
+            training_args.learning_rate,
+        )
+
+    # We use Optax's "masking" functionality to not apply weight decay
+    # to bias and LayerNorm scale parameters. decay_mask_fn returns a
+    # mask boolean with the same structure as the parameters.
+    # The mask is True for parameters that should be decayed.
+    # Note that this mask is specifically adapted for FlaxGPT2.
+    # For other models, one should correct the layer norm parameter naming
+    # accordingly.
+    def decay_mask_fn(params):
+        flat_params = traverse_util.flatten_dict(params)
+        flat_mask = {
+            path: (path[-1] != "bias" and path[-2:] not in [("ln_1", "scale"), ("ln_2", "scale"), ("ln_f", "scale")])
+            for path in flat_params
+        }
+        return traverse_util.unflatten_dict(flat_mask)
+
+    # create adam optimizer
+    if training_args.adafactor:
+        # We use the default parameters here to initialize adafactor,
+        # For more details about the parameters please check https://github.com/deepmind/optax/blob/ed02befef9bf81cbbf236be3d2b0e032e9ed4a40/optax/_src/alias.py#L74
+        optimizer = optax.adafactor(
+            learning_rate=lr_schedule_fn,
+        )
+
+    elif training_args.distributed_shampoo:
+        # parameters from https://github.com/tensorflow/lingvo/blob/03ee9d7cd50764b0424c7c863733c91fc0b053ec/lingvo/jax/optimizers.py#L729
+        # Notes:
+        # - mask for weight decay is not implemented but we don't use it anyway
+        optimizer = distributed_shampoo(
+            lr_schedule_fn,
+            block_size=1536,  # recommended default for large LM is 1536
+            beta1=0.9,
+            beta2=0.999,
+            diagonal_epsilon=1e-10,
+            matrix_epsilon=1e-8,
+            weight_decay=0.0,
+            start_preconditioning_step=1001,
+            preconditioning_compute_steps=10,
+            statistics_compute_steps=1,
+            best_effort_shape_interpretation=True,
+            graft_type=GraftingType.RMSPROP_NORMALIZED,
+            nesterov=False,
+            exponent_override=0,
+            batch_axis_name="batch",
+            inverse_failure_threshold=0.1,
+            moving_average_for_momentum=True,
+            skip_preconditioning_dim_size_gt=4096,
+            clip_by_scaled_gradient_norm=None,
+            precision=jax.lax.Precision.HIGHEST,
+            best_effort_memory_usage_reduction=training_args.quantize_shampoo,
+        )
+    else:
+        optimizer = optax.adamw(
+            learning_rate=lr_schedule_fn,
+            b1=training_args.adam_beta1,
+            b2=training_args.adam_beta2,
+            eps=training_args.adam_epsilon,
+            weight_decay=training_args.weight_decay,
+            mask=decay_mask_fn,
+        )
+        if training_args.gradient_clipping:
+            optimizer = optax.chain(
+                optax.clip_by_global_norm(1.),
+                optimizer
+            )
+
+    # add gradient accumulation
+    if training_args.gradient_accumulation_steps > 1:
+        optimizer = optax.chain(
+            optax.apply_every(training_args.gradient_accumulation_steps), optimizer
+        )
+
+    # Setup train state
+    state = TrainState.create(apply_fn=model.__call__, params=model.params, tx=optimizer, dropout_rng=dropout_rng)
+
+    def loss_fn(logits, labels):
+        shift_logits = logits[..., :-1, :]
+        shift_labels = labels[..., 1:]
+        loss = optax.softmax_cross_entropy(shift_logits, onehot(shift_labels, shift_logits.shape[-1]))
+        return loss.mean()
+
+    # Define gradient update step fn
+    def train_step(state, batch):
+        dropout_rng, new_dropout_rng = jax.random.split(state.dropout_rng)
+
+        def compute_loss(params):
+            labels = batch.pop("labels")
+            logits = state.apply_fn(**batch, params=params, dropout_rng=dropout_rng, train=True)[0]
+            loss = loss_fn(logits, labels)
+            return loss
+
+        grad_fn = jax.value_and_grad(compute_loss)
+        loss, grad = grad_fn(state.params)
+        grad = jax.lax.pmean(grad, "batch")
+
+        new_state = state.apply_gradients(grads=grad, dropout_rng=new_dropout_rng)
+
+        metrics = {"loss": loss, "learning_rate": lr_schedule_fn(state.step // training_args.gradient_accumulation_steps)}
+        metrics = jax.lax.pmean(metrics, axis_name="batch")
+
+        return new_state, metrics
+
+    # Define eval fn
+    def eval_step(params, batch):
+        labels = batch.pop("labels")
+        logits = model(**batch, params=params, train=False)[0]
+        loss = loss_fn(logits, labels)
+
+        # summarize metrics
+        metrics = {"loss": loss}
+        metrics = jax.lax.pmean(metrics, axis_name="batch")
+        return metrics
+
+    # Create parallel version of the train and eval step
+    p_train_step = jax.pmap(train_step, "batch", donate_argnums=(0,))
+    p_eval_step = jax.pmap(eval_step, "batch")
+
+    # Replicate the train state on each device
+    state = state.replicate()
+
+    logger.info("***** Running training *****")
+    logger.info(f"  Num examples = {len(train_dataset)}")
+    logger.info(f"  Num Epochs = {num_epochs}")
+    logger.info(f"  Instantaneous batch size per device = {training_args.per_device_train_batch_size}")
+    logger.info(f"  Total train batch size (w. parallel & distributed) = {train_batch_size}")
+    logger.info(f"  Total optimization steps = {total_train_steps}")
+
+    train_time = 0
+    train_metrics = []
+    epochs = tqdm(range(num_epochs), desc="Epoch ... ", position=0)
+    for epoch in epochs:
+        # ======================== Training ================================
+        train_start = time.time()
+
+        # Create sampling rng
+        rng, input_rng = jax.random.split(rng)
+
+        # Generate an epoch by shuffling sampling indices from the train dataset
+        train_loader = data_loader(input_rng, train_dataset, train_batch_size // training_args.gradient_accumulation_steps, shuffle=True)
+        steps_per_epoch = len(train_dataset) // train_batch_size
+        # train
+        steps_trained_progress_bar = tqdm(range(steps_per_epoch), desc="Training...", position=1,
+                                          leave=False)
+        for step in range(steps_per_epoch * training_args.gradient_accumulation_steps):
+            batch = next(train_loader)
+            batch = shard(batch)
+            state, train_metric = p_train_step(state, batch)
+            train_metrics.append(train_metric)
+
+            cur_step = epoch * (steps_per_epoch*training_args.gradient_accumulation_steps) + step
+
+            if step % training_args.gradient_accumulation_steps == 0:
+                steps_trained_progress_bar.update(1)
+
+            if cur_step % (training_args.logging_steps * training_args.gradient_accumulation_steps) == 0 and cur_step > 0:
+                # Save metrics
+                train_metric = unreplicate(train_metric)
+                train_time += time.time() - train_start
+                if has_tensorboard and jax.process_index() == 0:
+                    write_train_metric(summary_writer, train_metrics, train_time, cur_step)
+
+                epochs.write(
+                    f"Step... ({cur_step} | Loss: {train_metric['loss'].mean()}, Learning Rate: {train_metric['learning_rate'].mean()})"
+                )
+
+                train_metrics = []
+
+            if cur_step % (training_args.eval_steps * training_args.gradient_accumulation_steps) == 0 and cur_step > 0:
+                # ======================== Evaluating ==============================
+                eval_metrics = []
+                eval_loader = data_loader(input_rng, eval_dataset, eval_batch_size)
+                eval_steps = len(eval_dataset) // eval_batch_size
+                for _ in tqdm(range(eval_steps), desc="Evaluating...", position=2, leave=False):
+                    # Model forward
+                    batch = next(eval_loader)
+                    batch = shard(batch)
+                    metrics = p_eval_step(state.params, batch)
+                    eval_metrics.append(metrics)
+
+                # normalize eval metrics
+                eval_metrics = get_metrics(eval_metrics)
+                eval_metrics = jax.tree_map(jnp.mean, eval_metrics)
+
+                try:
+                    eval_metrics["perplexity"] = math.exp(eval_metrics["loss"])
+                except OverflowError:
+                    eval_metrics["perplexity"] = float("inf")
+
+                # Print metrics and update progress bar
+                desc = f"Step... ({cur_step} | Eval Loss: {eval_metrics['loss']} | Eval Perplexity: {eval_metrics['perplexity']})"
+                epochs.write(desc)
+                epochs.desc = desc
+
+                # Save metrics
+                if has_tensorboard and jax.process_index() == 0:
+                    write_eval_metric(summary_writer, eval_metrics, cur_step)
+
+            if cur_step % (training_args.save_steps * training_args.gradient_accumulation_steps) == 0 and cur_step > 0:
+                # save checkpoint after each epoch and push checkpoint to the hub
+                if jax.process_index() == 0:
+                    params = jax.device_get(unreplicate(state.params))
+                    model.save_pretrained(training_args.output_dir, params=params)
+                    tokenizer.save_pretrained(training_args.output_dir)
+                    if training_args.push_to_hub:
+                        repo.push_to_hub(commit_message=f"Saving weights and logs of step {cur_step}", blocking=False)
+
+        # save also at the end of epoch
+        try:
+            if jax.process_index() == 0:
+                params = jax.device_get(unreplicate(state.params))
+                model.save_pretrained(training_args.output_dir, params=params)
+                tokenizer.save_pretrained(training_args.output_dir)
+                if training_args.push_to_hub:
+                    repo.push_to_hub(commit_message=f"Saving weights and logs of epoch {epoch}", blocking=False)
+        except:
+            # push to hub fails the whole script if nothing new to commit
+            pass
+
+    # Eval after training
+    if training_args.do_eval:
+        eval_metrics = []
+        eval_loader = data_loader(input_rng, eval_dataset, eval_batch_size)
+        eval_steps = len(eval_dataset) // eval_batch_size
+        for _ in tqdm(range(eval_steps), desc="Evaluating...", position=2, leave=False):
+            # Model forward
+            batch = shard(next(eval_loader))
+            metrics = p_eval_step(state.params, batch)
+            eval_metrics.append(metrics)
+
+        # normalize eval metrics
+        eval_metrics = get_metrics(eval_metrics)
+        eval_metrics = jax.tree_map(lambda x: jnp.mean(x).item(), eval_metrics)
+
+        try:
+            eval_metrics["perplexity"] = math.exp(eval_metrics["loss"])
+        except OverflowError:
+            eval_metrics["perplexity"] = float("inf")
+
+        if jax.process_index() == 0:
+            eval_metrics = {f"eval_{metric_name}": value for metric_name, value in eval_metrics.items()}
+            path = os.path.join(training_args.output_dir, "eval_results.json")
+            with open(path, "w") as f:
+                json.dump(eval_metrics, f, indent=4, sort_keys=True)
+
+
+if __name__ == "__main__":
+    main()

runs/events.out.tfevents.1642710569.t1v-n-42145f73-w-0.1403347.0.v2

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:b031f0caff3541188f2e55871a0d86173ae697c82d2af899974ae5a60ed6cb65
+size 53662711

special_tokens_map.json

@@ -0,0 +1 @@
+{"bos_token": "<|endoftext|>", "eos_token": "<|endoftext|>", "unk_token": "<|endoftext|>", "pad_token": "<|endoftext|>"}

start_train.sh

@@ -0,0 +1,29 @@
+# set train hyperparams
+unset LD_PRELOAD
+export HF_DATASETS_CACHE="/researchdisk/datasets_cache"
+export USE_TORCH=0
+python3 run_clm_flax.py \
+    --output_dir="./" \
+    --model_type="gpt2" \
+    --config_name="./" \
+    --tokenizer_name="./" \
+    --dataset_filepath="/researchdisk/training_dataset_full_deduplicated" \
+    --do_train --do_eval \
+    --block_size="512" \
+    --per_device_train_batch_size="16" \
+    --per_device_eval_batch_size="16" \
+    --preprocessing_num_workers="1" \
+    --adam_beta1="0.9" \
+    --adam_beta2="0.98" \
+    --learning_rate="1e-4" \
+    --weight_decay="0.01" \
+    --warmup_steps="4000" \
+    --cosine_decay \
+    --overwrite_output_dir \
+    --logging_steps="500" \
+    --eval_steps="10000" \
+    --save_steps="10000" \
+    --num_train_epochs="10" \
+    --dtype="bfloat16" \
+    --push_to_hub \
+    --hub_model_id="Finnish-NLP/gpt2-medium-finnish"

tokenizer_config.json

@@ -0,0 +1 @@
+{"unk_token": "<|endoftext|>", "bos_token": "<|endoftext|>", "eos_token": "<|endoftext|>", "add_prefix_space": false, "special_tokens_map_file": null, "name_or_path": "./", "tokenizer_class": "GPT2Tokenizer"}

train_tokenizer.py

@@ -0,0 +1,30 @@
+from datasets import load_from_disk
+from tokenizers import trainers, Tokenizer, normalizers, ByteLevelBPETokenizer
+from transformers import AutoTokenizer
+
+
+model_dir = "./"
+
+# load dataset
+dataset = load_from_disk("/researchdisk/training_dataset_full_deduplicated")
+dataset = dataset["train"]
+
+# Instantiate tokenizer
+tokenizer = ByteLevelBPETokenizer()
+def batch_iterator(batch_size=1000):
+    for i in range(0, len(dataset), batch_size):
+        yield dataset[i: i + batch_size]["text"]
+
+# Customized training
+tokenizer.train_from_iterator(batch_iterator(), vocab_size=50257, min_frequency=2, special_tokens=[
+    "<s>",
+    "<pad>",
+    "</s>",
+    "<unk>",
+    "<mask>",
+])
+
+# Save files to disk
+tokenizer.save(f"{model_dir}/tokenizer.json")
+tokenizer = AutoTokenizer.from_pretrained(model_dir)
+tokenizer.save_pretrained(model_dir)