DataTrove

DataTrove is a library to process, filter and deduplicate text data at a very large scale. It provides a set of prebuilt commonly used processing blocks with a framework to easily add custom functionality.

DataTrove processing pipelines are platform-agnostic, running out of the box locally or on a slurm cluster. Its (relatively) low memory usage and multiple step design makes it ideal for large workloads, such as to process an LLM's training data.

Local, remote and other file systems are supported through fsspec.

Table of contents

• Installation
• Quickstart examples
• Terminology
• Pipeline
  • DataTrove Document
  • Types of pipeline blocks
  • Full pipeline
• Executors
  • LocalPipelineExecutor
  • SlurmPipelineExecutor
  • RayPipelineExecutor
• Logging
• DataFolder / paths
• Practical guides
  • Reading data
  • Synthetic data generatioon
  • Extracting text
  • Filtering data
  • Saving data
  • Deduplicating data
  • Summary Statistics
  • Custom blocks
  • Simple data
  • Custom function
  • Custom block
• Contributing
• Citation

Installation

bash pip install datatrove[FLAVOUR] Available flavours (combine them with , i.e. [processing,s3]): - all installs everything: pip install datatrove[all] - io dependencies to read warc/arc/wet files and arrow/parquet formats: pip install datatrove[io] - processing dependencies for text extraction, filtering and tokenization: pip install datatrove[processing] - s3 s3 support: pip install datatrove[s3] - cli for command line tools: pip install datatrove[cli] - ray for distributed compute engine: pip install datatrove[ray]

Quickstart examples

You can check the following examples: - fineweb.py full reproduction of the FineWeb dataset - processcommoncrawl_dump.py full pipeline to read commoncrawl warc files, extract their text content, filters and save the resulting data to s3. Runs on slurm - tokenize_c4.py reads data directly from huggingface's hub to tokenize the english portion of the C4 dataset using the gpt2 tokenizer - minhash_deduplication.py full pipeline to run minhash deduplication of text data - sentence_deduplication.py example to run sentence level exact deduplication - exact_substrings.py example to run ExactSubstr (requires this repo)

Terminology

• pipeline: a list of processing steps to execute (read data, filter, write to disk, etc)
• executor: runs a specific pipeline on a given execution environment (slurm, multi cpu machine, etc)
• job: the execution of a pipeline on a given executor
• task: a job is comprised of multiple tasks, and these are used to parallelize execution, usually by having each task process a shard of data. Datatrove keeps track of which tasks have completed and when you relaunch only incomplete tasks will run.
• file: an individual input file (.json, .csv, etc). > [!TIP] > Note that each file will be processed by a single task. Datatrove does not automatically split a file into multiple parts, so to fully parallelize you should have multiple medium sized files rather than a single large file)
• shard: a group of input data (usually a group of files), which will be assigned to a specific task. Each task will process a different non overlapping shard of data, from the full list of input files
• worker: compute resource that will execute a single task at a time, e.g., if you have 50 cpu cores you can run a LocalPipelineExecutor with workers=50, to execute 50 tasks simultaneously (one per cpu). Once a worker is done with a task, it will start processing another waiting task

[!TIP] Your number of tasks controls how much you can parallelize and also how much time each individual processing unit will take. If you have a small number of tasks (and they each therefore have to process a large number of files) and they fail, you will have to restart from scratch, whereas if you have a larger number of small tasks each failed task will take way less time to rerun.

[!CAUTION] If your tasks > files, some tasks will not process any data, so there usually isn't a point in setting tasks to a number larger than files.

Example

Running a job to process 10000 files, on a machine with 100 cpu cores (workers). If we choose to use 1000 tasks, each one will process a shard of 10 files. workers=100 means that we can process 100 tasks at a time.

Pipeline

DataTrove Document

Each pipeline block processes data in the datatrove Document format: - text the actual text content for each sample - id a unique id (string) for this sample - metadata a dictionary where any additional info may be stored

Types of pipeline blocks

Each pipeline block takes a generator of Document as input and returns another generator of Document. - readers read data from different formats and yield Document - writers save Document to disk/cloud in different formats - extractors extract text content from raw formats (such as webpage html) - filters filter out (remove) some Documents based on specific rules/criteria - stats blocks to collect statistics on the dataset - tokens blocks to tokenize data or count tokens - dedup blocks for deduplication

Full pipeline

A pipeline is defined as a list of pipeline blocks. As an example, the following pipeline would read data from disk, randomly filter (remove) some documents and write them back to disk: ```python from datatrove.pipeline.readers import CSVReader from datatrove.pipeline.filters import SamplerFilter from datatrove.pipeline.writers import JsonlWriter

pipeline = [ CSVReader( datafolder="/my/input/path" ), SamplerFilter(rate=0.5), JsonlWriter( outputfolder="/my/output/path" ) ] ```

Executors

Pipelines are platform-agnostic, which means that the same pipeline can smoothly run on different execution environments without any changes to its steps. Each environment has its own PipelineExecutor.

Some options common to all executors: - pipeline a list consisting of the pipeline steps that should be run - logging_dir a datafolder where log files, statistics and more should be saved. Do not reuse folders for different pipelines/jobs as this will overwrite your stats, logs and completions. - skip_completed (bool, True by default) datatrove keeps track of completed tasks so that when you relaunch a job they can be skipped. Set this to False to disable this behaviour - randomize_start_duration (int, 0 by default) the maximum number of seconds to delay the start of each task to prevent all tasks from starting simultaneously and potentially overloading the system.

Call an executor's run method to execute its pipeline.

[!TIP] Datatrove keeps track of which tasks successfully completed by creating a marker (an empty file) in the ${logging_dir}/completions folder. Once the job finishes, if some of its tasks have failed, you can simply relaunch the exact same executor and datatrove will check and only run the tasks that were not previously completed.

[!CAUTION] If you relaunch a pipeline because some tasks failed, do not change the total number of tasks as this will affect the distribution of input files/sharding.

LocalPipelineExecutor

This executor will launch a pipeline on a local machine. Options: - tasks total number of tasks to run - workers how many tasks to run simultaneously. If -1, no limit. Anything > 1 will use multiprocessing to execute the tasks. - start_method method to use to spawn a multiprocessing Pool. Ignored if workers is 1

Example executor

```python from datatrove.executor import LocalPipelineExecutor executor = LocalPipelineExecutor( pipeline=[ ... ], logging_dir="logs/", tasks=10, workers=5 ) executor.run() ```

Multi-node parallelism

You can have different nodes/machines process different parts of the total tasks by using the `local_tasks` and `local_rank_offset`. For each node/instance/machine, launch with the following options: - `tasks` the total tasks to be executed (across all machines). **This value must be the same on each machine or the input file distribution may overlap!** Example: 500 - `local_tasks` how many tasks of the total will be executed on this particular machine. Note that you can use different values for each machine. Example: 100 - `local_rank_offset` the rank of the first task to be executed on this machine. If this is the 3rd machine where you are launching a job, and the 2 previous machines each ran 250 and 150 jobs, this would be `400` for the current machine. To get final merged stats you will have to invoke the `merge_stats` script manually on a path containing the stats from all machines.

SlurmPipelineExecutor

This executor will launch a pipeline on a slurm cluster, using slurm job arrays to group and manage tasks. Options: - tasks total number of tasks to run. required - time slurm time limit string. required - partition slurm partition. required - workers how many tasks to run simultaneously. If -1, no limit. Slurm will run workers tasks at a time. (default: -1) - job_name slurm job name (default: "dataprocessing) - depends another SlurmPipelineExecutor instance, which will be a dependency of this pipeline (current pipeline will only start executing after the depended on pipeline successfully completes) - `sbatchargsdictionary with any other arguments you would like to pass to sbatch -slurmlogsfolderwhere to save the slurm log files. If using a local path forloggingdir, they will be saved onloggingdir/slurm_logs`. If not, they will be saved as a subdir of the current directory.

Other options

• cpus_per_task how many cpus to give each task (default: 1)
• qos slurm qos (default: "normal")
• mem_per_cpu_gb memory per cpu, in GB (default: 2)
• env_command custom command to activate a python environment, if needed
• condaenv conda environment to activate
• venv_path path to a python environment to activate
• max_array_size the MaxArraySize value in $ scontrol show config. If number of tasks exceeds this number, it will split into multiple array jobs (default: 1001)
• max_array_launch_parallel if we need multiple jobs due to maxarraysize, whether to launch them all in one go (parallel) or sequentially (default: False)
• stagger_max_array_jobs when maxarraylaunch_parallel is True, this determines how many seconds to wait between launching each of the parallel jobs (default: 0)
• run_on_dependency_fail start executing when a job we depend on finishes even if it has failed (default: False)
• randomize_start randomize the start of each task in a job in a ~3 min window. Useful when heavily hitting an s3 bucket for example. (default: False)

Example executor

```python from datatrove.executor import SlurmPipelineExecutor executor1 = SlurmPipelineExecutor( pipeline=[ ... ], jobname="mycooljob1", loggingdir="logs/job1", tasks=500, workers=100, # omit to run all at once time="10:00:00", # 10 hours partition="hopper-cpu" ) executor2 = SlurmPipelineExecutor( pipeline=[ ... ], jobname="mycooljob2", loggingdir="logs/job2", tasks=1, time="5:00:00", # 5 hours partition="hopper-cpu", depends=executor1 # this pipeline will only be launched after executor1 successfully completes )

executor1.run()

executor2.run() # this will actually launch executor1, as it is a dependency, so no need to launch it explicitly ```

RayPipelineExecutor

This executor will launch a pipeline on a ray cluster, using ray tasks for parallel execution. Options: - tasks total number of tasks to run. - workers how many tasks to run simultaneously. If -1, no limit. Ray will run workers tasks at a time. (default: -1) - depends another RayPipelineExecutor instance, which will be a dependency of this pipeline (current pipeline will only start executing after the depended on pipeline successfully completes)

Other options

• cpus_per_task how many cpus to give each task (default: 1)
• mem_per_cpu_gb memory per cpu, in GB (default: 2)
• ray_remote_kwargs Additional kwargs to pass to the ray.remote decorator

Example executor

python import ray from datatrove.executor import RayPipelineExecutor ray.init() executor = RayPipelineExecutor( pipeline=[ ... ], logging_dir="logs/", tasks=500, workers=100, # omit to run all at once ) executor.run()

Logging

For a pipeline with logging_dir mylogspath/exp1, the following folder structure would be created:

See folder structure

``` └── mylogspath/exp1 │── executor.json ⟵ json dump of the executor options and pipeline steps │── launch_script.slurm ⟵ the slurm config created and used to launch this job (if running on slurm) │── executor.pik ⟵ the slurm config created and used to launch this job (if running on slurm) │── ranks_to_run.json ⟵ list of tasks that are being run │── logs/ │ └──[task_00000.log, task_00001.log, task_00002.log, ...] ⟵ individual logging files for each task │── completions/ │ └──[00004, 00007, 00204, ...] ⟵ empty files marking a task as completed. Using when relaunching/resuming a job (only unfinished tasks will be run) │── stats/ │ └──[00000.json, 00001.json, 00002.json, ...] ⟵ individual stats for each task (number of samples processed, filtered, removed, etc) └── stats.json ⟵ global stats from all tasks ```

Colorization

Log messages support colorization. By default, colorization will be auto detected for console messages and disabled for log files (logs/taskXXXXX.log). To explicitly enable or disable colorization, you may set the following environment variables: - `DATATROVECOLORIZELOGS"1" to add ANSI colors to console log messages and "0" to disable colorization. -DATATROVECOLORIZELOGFILES` set to "1" to add ANSI colors to log messages saved to logs/task_XXXXX.log.

DataFolder / paths

Datatrove supports a wide variety of input/output sources through fsspec.

There are a few ways to provide a path to a datatrove block (for input_folder, logging_dir, data_folder and so on arguments): - str: the simplest way is to pass a single string. Example: /home/user/mydir, s3://mybucket/myinputdata, hf://datasets/allenai/c4/en/

• (str, fsspec filesystem instance): a string path and a fully initialized filesystem object. Example: ("s3://mybucket/myinputdata", S3FileSystem(client_kwargs={"endpoint_url": endpoint_uri}))
• (str, dict): a string path and a dictionary with options to initialize a fs. Example (equivalent to the previous line): ("s3://mybucket/myinputdata", {"client_kwargs": {"endpoint_url": endpoint_uri}})
• DataFolder: you can initialize a DataFolder object directly and pass it as an argument

Under the hood these argument combinations are parsed by get_datafolder.

Practical guides

Reading data

Usually, pipelines will start with a Reader block. Most readers take a data_folder argument — a path to a folder containing the data to be read.

These files will be distributed across each task. If you have N tasks, task with rank i (0-based) will process files i, i+N, i+2N, i+3N,....

Internally, each reader reads data and converts it into a dictionary before creating a Document object.

Some options common to most readers: - text_key the dictionary key containing the text content for each sample. Default: text - id_key the dictionary key containing the id for each sample. Default: id - default_metadata a dictionary for any default metadata values you would like to add (such as their source, for example) - recursive whether to look for files recursively in data_folder's subdirectories - glob_pattern use this field to match specific files. For instance, glob_pattern="*/warc/*.warc.gz" will match files with a .warc.gz file extension on the warc/ folder of each of the data_folder's subdirectories - adapter this function takes the raw dictionary obtained from the reader and returns a dictionary with Document's field names. You may overwrite this function (defaultadapter) if you would like. - limit read only a certain number of samples. Useful for testing/debugging

Synthetic data generation

We support vLLM and SGLang for inference using the InferenceRunner block. Each datatrove task will spawn a replica of the target model and full asynchronous continuous batching will guarantee high model throughput.

By setting checkpoints_local_dir and records_per_chunk generations will be written to a local folder until a chunk is complete, allowing for checkpointing in case tasks fail or are preempted.

Tune max_concurrent_requests to tune batching behaviour. If you have slow pre-processing, you can also increase max_concurrent_tasks (to a value higher than max_concurrent_requests).

Refer to the example for more info.

Extracting text

You can use extractors to extract text content from raw html. The most commonly used extractor in datatrove is Trafilatura, which uses the trafilatura library.

Filtering data

Filters are some of the most important blocks of any data processing pipeline. Datatrove's filter blocks take a Document and return a boolean (True to keep a document, False to remove it). Removed samples do not continue to the next pipeline stage. You can also save the removed samples to disk by passing a Writer to the exclusion_writer parameter.

Saving data

Once you are done processing your data you will probably want to save it somewhere. For this you can use a writer. Writers require an output_folder (the path where data should be saved). You can choose the compression to use (default: gzip) and the filename to save each file as. For the output_filename, a template is applied using the following arguments: - ${rank} replaced with the current task's rank. Note that if this tag isn't present, different tasks may try to write to the same location - ${id} replaced with the sample id - metadata: any other ${tag} will be replaced with the corresponding document.metadata['tag'] value

An example to separate samples by language based on their lang metadata field: JsonlWriter( f"{MAIN_OUTPUT_PATH}/non_english/", output_filename="${language}/" + DUMP + "/${rank}.jsonl.gz", # folder structure: language/dump/file )

Deduplicating data

For deduplication check the examples minhash_deduplication.py, sentence_deduplication.py and exact_substrings.py.

Summary Statistics

For summary statistics on your data you can use the Stats blocks. These blocks provide an easy way to collect data-profiles on your dataset in a distributed manner. It's a two step process in which you first: 1) For each shard iterate over documents and collect stats into of the following groupings summary (all docs counted to "summary" key), fqdn (fully qualified domain name grouping), suffix (the last part of the url path grouping) or histogram (value based grouping). 2) Merge the stats from different shards into a single file. See the summary_stats.py for more details.

Each resulting stat is saved in a separate file with following structure: output_folder/{fqdn,suffix,summary,histogram}/{stat_name}/metric.json

Each such file is a MetricStatsDict object, which you can easily load using: ```python from datatrove.pipeline.stats.summarystats import MetricStatsDict import json stats = MetricStatsDict.fromdict(json.load(open("fqdn/length/metric.json")))

E.g for total length of nytimes.com docs

stats["nytimes.com"].total

Or for mean of cnn.com docs

stats["cnn.com"].mean ```

Following stats are available: - contamination_stats.py: word_contamination_{words[0]}: Frequency of words contamination in the document. - doc_stats.py: length: Length of the document, white_space_ratio: Ratio of whitespace characters, non_alpha_digit_ratio: Ratio of non-alphabetic and non-digit characters, digit_ratio: Ratio of digits, uppercase_ratio: Ratio of uppercase letters, elipsis_ratio: Ratio of elipsis characters, punctuation_ratio: Punctuation ratio - lang_stats.py: fasttext_{language}: Score of document being written in language. Score is computed using FastText model. - line_stats.py: n_lines: Number of lines per doc, avg_line_length: Average length of line per doc, long_line_ratio_chars_{chars}: Ratio of lines with more than k chars, short_line_ratio_chars_{chars}: Ratio of lines with less than k chars, bullet_point_lines_ratio: Ratio of line starting with bullet point, line_duplicates: Ratio of lines that are duplicates, line_char_duplicates: Ratio of chars in duplicated lines to all chars. - paragraph_stats.py: n_paragraphs: Number of paragraphs, avg_paragraph_length: Average paragraph length, short_paragraph_ratio_{chars}: Ratio of short paragraphs (<{chars} chars), long_paragraph_ratio_{chars}: Ratio of long paragraphs (>{chars} chars) - perplexity_stats.py: ccnet_perplexity_{model_dataset}_{language}: Perplexity of the document using the CCNet model for {model} on {dataset} in {language} - sentence_stats.py: n_sentences: Number of sentences, avg_sentence_length: Average sentence length, short_sentence_ratio_{chars}: Ratio of short sentences (<{chars} chars), long_sentence_ratio_{chars}: Ratio of long sentences (>{chars} chars) - token_stats.py:token_count: Number of tokens in the document - word_stats.py: n_words: Number of words in the document, avg_word_length: Average length of words in the document, avg_words_per_line: Average number of words per line in the document, short_word_ratio_{chars}: Ratio of words shorter than {chars} characters, stop_word_ratio: Ratio of stop words, long_word_ratio_{chars}: Ratio of words longer than {chars} characters, type_token_ratio: Number of unique words / Number of tokens, capitalized_word_ratio: Ratio of capitalized words, uppercase_word_ratio: Ratio of uppercase words

Custom blocks

Simple data

You can pass an iterable of Document directly as a pipeline block like so: ```python from datatrove.data import Document from datatrove.pipeline.filters import SamplerFilter from datatrove.pipeline.writers import JsonlWriter

pipeline = [ [ Document(text="some data", id="0"), Document(text="some more data", id="1"), Document(text="even more data", id="2"), ], SamplerFilter(rate=0.5), JsonlWriter( output_folder="/my/output/path" ) ] ```

Do note, however, that this iterable will not be sharded (if you launch more than 1 task they will all get the full iterable). This is usually useful for small workloads/testing.

Custom function

For simple processing you can simply pass in a custom function with the following signature: ```python from datatrove.data import DocumentsPipeline

def uppercaseeverything(data: DocumentsPipeline, rank: int = 0, worldsize: int = 1) -> DocumentsPipeline: """ data is a generator of Document. You must also return a generator of Document (yield) You can optionally use rank and world_size for sharding """ for document in data: document.text = document.text.upper() yield document

pipeline = [ ..., uppercase_everything, ... ] ```

[!TIP] You might have some pickling issues due to the imports. If this happens, simply move whatever imports you need inside the function body.

Custom block

You can also define a full block inheriting from PipelineStep or one of its subclasses:

```python from datatrove.pipeline.base import PipelineStep from datatrove.data import DocumentsPipeline from datatrove.io import DataFolderLike, get_datafolder

class UppercaserBlock(PipelineStep): def init(self, somefolder: DataFolderLike, someparam: int = 5): super().init() # you can take whatever parameters you need and save them here self.someparam = someparam # to load datafolders use getdatafolder() self.somefolder = getdatafolder(somefolder)

def run(self, data: DocumentsPipeline, rank: int = 0, world_size: int = 1) -> DocumentsPipeline:
    # you could also load data from the `some_folder`:
    for filepath in self.some_folder.get_shard(rank, world_size): # it also accepts a glob pattern, among other things
        with self.some_folder.open(filepath, "rt") as f:
            # do something
            ...
            yield doc

    #
    # OR process data from previous blocks (`data`)
    #

    for doc in data:
        with self.track_time():
            # you can wrap the main processing code in `track_time` to know how much each document took to process
            nr_uppercase_letters = sum(map(lambda c: c.isupper(), doc.text))
            # you can also keep track of stats per document using stat_update
            self.stat_update("og_upper_letters", value=nr_uppercase_letters)
            doc.text = doc.text.upper()
        # make sure you keep the yield outside the track_time block, or it will affect the time calculation
        yield doc

    #
    # OR save data to disk
    #

    with self.some_folder.open("myoutput", "wt") as f:
        for doc in data:
            f.write(doc...)

```

python pipeline = [ ..., UppercaserBlock("somepath"), ... ]

You could also inherit from BaseExtractor, BaseFilter, BaseReader/BaseDiskReader, or DiskWriter.

Contributing

bash git clone git@github.com:huggingface/datatrove.git && cd datatrove pip install -e ".[dev]"

Install pre-commit code style hooks: bash pre-commit install

Run the tests: bash pytest -sv ./tests/

Citation

bibtex @misc{penedo2024datatrove, author = {Penedo, Guilherme and Kydlíček, Hynek and Cappelli, Alessandro and Sasko, Mario and Wolf, Thomas}, title = {DataTrove: large scale data processing}, year = {2024}, publisher = {GitHub}, journal = {GitHub repository}, url = {https://github.com/huggingface/datatrove} }