llm-dataset-converter 0.1.1

Python3 library for converting between various LLM dataset formats.

The llm-dataset-converter allows the conversion between various dataset formats for large language models (LLMs). Filters can be supplied as well, e.g., for cleaning up the data.

Dataset formats:

• pairs: alpaca (r/w), csv (r/w), jsonl (r/w), parquet (r/w), tsv (r/w)

• pretrain: csv (r/w), jsonl (r/w), parquet (r/w), tsv (r/w), txt (r/w)

• translation: csv (r/w), jsonl (r/w), parquet (r/w), tsv (r/w), txt (r/w)

• classification: csv (r/w), jsonl (r/w), parquet (r/w), tsv (r/w)

Compression formats:

• bzip

• gzip

• xz

• zstd

Examples:

Simple conversion with logging info:

llm-convert \
  from-alpaca \
    -l INFO \
    --input ./alpaca_data_cleaned.json \
  to-csv-pr \
    -l INFO \
    --output alpaca_data_cleaned.csv

Automatic decompression/compression (based on file extension):

llm-convert \
  from-alpaca \
    --input ./alpaca_data_cleaned.json.xz \
  to-csv-pr \
    --output alpaca_data_cleaned.csv.gz

Filtering:

llm-convert \
  -l INFO \
  from-alpaca \
    -l INFO \
    --input alpaca_data_cleaned.json \
  keyword \
    -l INFO \
    --keyword function \
    --location any \
    --action keep \
  to-alpaca \
    -l INFO \
    --output alpaca_data_cleaned-filtered.json

Changelog

0.1.1 (2024-02-15)

• added classification domain

• added from-jsonlines-cl reader and to-jsonlines-cl writer for classification data in JSON lines format

• added filter pretrain-sentences-to-classification to turn pretrain data into classification data (with a predefined label)

• added filter classification-label-map that can generate a label string/int map

• the to-llama2-format filter now has the –skip_tokens options to leave out the [INST] [/INST] tokens

• added from-parquet-cl reader and to-parquet-cl writer for classification data in Parquet database format

• added from-csv-cl/from-tsv-cl readers and to-csv-cl/to-tsv-cl writers for classification data in CSV/TSV file format

0.1.0 (2024-02-05)

• fixed output format of to-llama2-format filter

• llama2-to-pairs filter has more robust parsing now

• upgraded seppl to 0.1.0

• switched to seppl classes: Splitter, MetaDataHandler, Reader, Writer, StreamWriter, BatchWriter

0.0.5 (2024-01-24)

• added flag -b/–force_batch to the llm-convert tool which all data to be reader from the reader before filtering it and then passing it to the writer; useful for batch filters.

• added the randomize-records batch filter

• added the –encoding ENC option to file readers

• auto-determined encoding is now being logged (INFO level)

• the LDC_ENCODING_MAX_CHECK_LENGTH environment variable allows overriding the default number of bytes used for determining the file encoding in auto-detect mode

• default max number of bytes inspected for determining file encoding is now 10kb

• method locate_files in base_io no longer includes directories when expanding globs

• added tool llm-file-encoding for determining file encodings of text files

• added method replace_extension to base_io module for changing a files extension (removes any supported compression suffix first)

• stream writers (.jsonl/.txt) now work with –force_batch mode; the output file name gets automatically generated from the input file name when just using a directory for the output

0.0.4 (2023-12-19)

• pairs-to-llama2 filter now has an optional –prefix parameter to use with the instruction

• added the pretrain-sentences-to-pairs filter for generating artificial prompt/response datasets from pretrain data

• requires seppl>=0.0.11 now

• the LDC_MODULES_EXCL environment variable is now used for specifying modules to be excluded from the registration process (e.g., used when generating help screens for derived libraries that shouldn’t output the base plugins as well)

• llm-registry and llm-help now allow specifying excluded modules via -e/–excluded_modules option

• to-alpaca writer now has the -a/–ensure_ascii flag to enforce ASCII compatibility in the output

• added global option -u/–update_interval to convert tool to customize how often progress of # records processed is being output in the console (default: 1000)

• text-length filter now handles None values, i.e., ignores them

• locations (i.e., input/instructions/output/etc) can be specified now multiple times

• the llm-help tool can generate index files for all the plugins now; in case of markdown it will link to the other markdown files

0.0.3 (2023-11-10)

• added the record-window filter

• added the llm-registry tool for querying the registry from the command-line

• added the replace_patterns method to ldc.text_utils module

• added the replace-patterns filter

• added -p/–pretty-print flag to to-alpaca writer

• added pairs-to-llama2 and llama2-to-pairs filter (since llama2 has instruction as part of the string, it is treated as pretrain data)

• added to-llama2-format filter for pretrain records (no [INST]…[/INST] block)

• now requiring seppl>=0.0.8 in order to raise Exceptions when encountering unknown arguments

0.0.2 (2023-10-31)

• added text-stats filter

• stream writers accept iterable of data records now as well to improve throughput

• text_utils.apply_max_length now uses simple whitespace splitting instead of searching for nearest word boundary to break a line, which results in a massive speed improvement

• fix: text_utils.remove_patterns no longer multiplies the generated lines when using more than one pattern

• added remove-patterns filter

• pretrain and translation text writers now buffer records by default (-b, –buffer_size) in order to improve throughput

• jsonlines writers for pair, pretrain and translation data are now stream writers

0.0.1 (2023-10-26)

• initial release