- Notifications
You must be signed in to change notification settings - Fork19
DSIR large-scale data selection framework for language model training
License
p-lambda/dsir
Folders and files
Name | Name | Last commit message | Last commit date | |
---|---|---|---|---|
Repository files navigation
This repository contains theDSIR data selection tool for selecting relevant language model training data from any raw data source given a target dataset, as well as pre-filtered datasets and some pretrained models.
DSIR is built for:
- fast, large-scale (trillion-token scale) data selection from large raw text datasets (Pile, RefinedWeb, RedPajama, ...). There is almost no overhead to selecting more examples (unlike retrieval), other than the time it takes to write the extra examples to disk.
- selecting data that is distributed like a given target dataset (domain-specific data, Wikipedia, ...). Relevance and diversity are balanced automatically by matching the distribution of the target dataset on a feature space (e.g., n-gram frequencies).
Compute needed:
- 1 CPU node
- a decent amount of RAM (at least 64GB for most large datasets - need to hold a few floats per example in memory)
- a high number of cores. The data selection speed scales linearly with the CPU cores.
Code related to the DSIR paper's experiments are in theexperimental/
directory.
Install with pip:
pip install data-selection
Install from source by cloning this repo and installing via pip:
git clone git@github.com:/p-lambda/dsirpip install ./dsir
To select data, simply initialize aHashedNgramDSIR
object and call the following functions:
fromdata_selectionimportHashedNgramDSIRraw_datasets= [<listofpaths>]target_datasets= [<listofpaths>]dsir=HashedNgramDSIR(raw_datasets,target_datasets,cache_dir='/path/to/dsir_cache')dsir.fit_importance_estimator(num_tokens_to_fit='auto')dsir.compute_importance_weights()dsir.resample(out_dir='resampled',num_to_sample=10000000,cache_dir='/path/to/resampled_cache')
Running this would write 10M documents injsonl
files inside an output directory namedresampled
. The files will first be written tocache_dir
and moved toout_dir
upon completion (setcache_dir
toNone
to skip this step). For best performance, use uncompressedjsonl
files stored on local file storage for all data paths and use as many CPU cores as possible, which allows each file to be virtually sharded across multiple cores. Custom functions for reading the data paths and extracting the text field from each example can be provided via the{raw,target}_load_dataset_fn
and{raw,target}_parse_example_fn
arguments to the constructor. The number of tokens to use for fitting the importance weight estimator can be tuned with thenum_tokens_to_fit
argument (set toall
to fit on full dataset). Top-k retrieval instead of sampling without replacement (the default) can be done by specifyingtop_k=True
to theresample
method.
(Note: for results similar to the paper, first preprocess the documents by breaking them into equal-word-length chunks, and usetokenizer="word_tokenize"
in theHashedNgramDSIR
constructor.)
Thedsir
intermediate results (afterfit_importance_estimator
andcompute_importance_weights
) can be saved and loaded for later use, for example to resample 100M documents instead:
dsir.save('/path/to/dsir_params.pkl')# later ondsir.load('/path/to/dsir_params.pkl')dsir.resample(out_dir='/path/to/out_dir',num_to_sample=100000000,cache_dir='/path/to/resampled_cache')
Thesave
method can be called at any time to save partial results.
SeeUsage documentation for full details.
Using 1 CPU node with 96GB RAM and 96 cores, we can select data from the full (decompressed) Pile dataset in less than4.5 hours.The Pile dataset was first decompressed and placed onto the node's local file storage. The breakdown of timings for each step are:
- Fit importance estimator (with
num_tokens_to_fit="auto"
): 59.28 seconds - Compute importance weights: 4.36 hours
- Resample 10M documents (with
cache_dir=None
andout_dir
is a local storage location): 353.68 seconds - Total: 4.47 hours
Subsequent resampling with the same target data is very cheap, and the runtime does not scale with the number of documents to select (unlike retrieval). Resampling 100M documents takes the same amount of time (less than6 minutes) as resampling 10M documents:
- Resample 10M documents: 353.68 seconds
- Resample 100M documents: 352.69 seconds
To select data from the Pile:
fromdata_selectionimportHashedNgramDSIR# 2-digit integers up to 29subsets= [str(i).zfill(2)foriinrange(0,30)]raw_datasets= [f'/path/to/pile/{subset}.jsonl'forsubsetinsubsets]target_datasets= ['/path/to/target.jsonl']dsir=HashedNgramDSIR(raw_datasets=raw_datasets,target_datasets=target_datasets,cache_dir='/path/to/dsir_cache')dsir.fit_importance_estimator(num_tokens_to_fit='auto')dsir.compute_importance_weights()dsir.resample(out_dir='/path/to/out_dir',num_to_sample=10000000,cache_dir='/path/to/resample_cache')
HuggingFace datasets can also be used in eitherraw_datasets
ortarget_datasets
(note: streaming a large raw dataset directly will be very slow - we recommend this more for target datasets):
fromdata_selectionimportHashedNgramDSIRfromdatasetsimportload_datasetsubsets= [str(i).zfill(2)foriinrange(0,30)]raw_datasets= [f'/path/to/pile/{subset}.jsonl'forsubsetinsubsets]target_datasets= ['codeparrot/self-instruct-starcoder','SetFit/mnli']deftarget_load_dataset_fn(dataset):ifdataset=='codeparrot/self-instruct-starcoder':ds=load_dataset(dataset,streaming=True,split='raw')else:ds=load_dataset(dataset,streaming=True,split='train').take(10000)returndsdeftarget_parse_example_fn(ex):if'output'inex:returnex['output']else:returnex['text1']+' '+ex['text2']dsir=HashedNgramDSIR(raw_datasets=raw_datasets,target_datasets=target_datasets,cache_dir='/path/to/dsir_cache',target_parse_example_fn=target_parse_example_fn,target_load_dataset_fn=target_load_dataset_fn,separate_targets=True)dsir.fit_importance_estimator(num_tokens_to_fit='auto')dsir.compute_importance_weights()dsir.resample(out_dir='/path/to/out_dir',num_to_sample=10000000,cache_dir='/path/to/resample_cache')
For use-cases where the target datasets are quite different (here, a mix of code and natural language), we recommend passing inseparate_targets
into the constructor.separate_targets
controls whether to select data separately for each target and then join them. For example, when including two target datasets, one natural language dataset and one code, the most heavily upweighted data whenseparate_targets=False
may skew towards documents with a mix of natural language and code, such as StackExchange. Whenseparate_targets=True
, two separate DSIR runs will occur in parallel, selecting a mixture of documents from each target according totarget_proportions
. Whentarget_proportions
is unspecified, the number of documents to select for each target is weighted according to the token sizes of each target dataset.
Paper:https://arxiv.org/abs/2302.03169
@article{xie2023data, author = {Sang Michael Xie and Shibani Santurkar and Tengyu Ma and Percy Liang}, journal = {Advances in Neural Information Processing Systems (NeurIPS)}, title = {Data Selection for Language Models via Importance Resampling}, year = {2023},}
About
DSIR large-scale data selection framework for language model training