OASST1 Huggingface compatible dataset generation scripts

load_hf_ds.py

from datasets import load_dataset

ds = load_dataset("/path/oasst1", name='ready')
train = ds['train']
val = ds['validation']
print(f'{len(train)=}')
print(f'{len(val)=}')

for i in range(5):
   print(train[i]["message_tree_id"])

ds.push_to_hub("OpenAssistant/[name]", private=True, token="...")

#test loading
#ds = load_dataset("OpenAssistant/oasst1")
#ds = load_dataset("OpenAssistant/oasst1")
#print(ds.keys())

oasst1.py

# Copyright 2023 Open-Assistant Contributors.
#
# 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.
"""OpenAssistant dataset loading script"""

import json
import os

import datasets


_CITATION = """\
@InProceedings{huggingface:dataset,
title = {OpenAssistant Conversations - Democratizing Large Language Model Alignment},
author={OpenAssistant Authors},
year={2023}
}
"""

_DESCRIPTION = """\
In an effort to democratize research on large-scale alignment,
we release OpenAssistant Conversations, a human-generated, human-annotated
assistant-style conversation corpus consisting of over 100,000 messages distributed
across 10,940 conversations in 32 different languages, annotated with over 200,000
quality ratings. The corpus is a product of a worldwide crowd-sourcing effort
involving over 13,500 volunteers. To demonstrate the OpenAssistant Conversations
dataset’s effectiveness, we present the first fully open-source large-scale instruction-
tuned model.
"""

_HOMEPAGE = "https://open-assistant.io/"

_LICENSE = "Apache v2.0"

_URLS = {
    #"ready": { "train": "https://huggingface.co/datasets/OpenAssistant/oasst1/resolve/main/2023-04-12_oasst_ready.messages.jsonl.gz" },
    "ready": { 
        "train": "path/DATE_oasst_ready_train.messages.jsonl",
        "valid": "path/DATE_oasst_ready_val.messages.jsonl"
     },
    "all": { "train":"https://huggingface.co/datasets/OpenAssistant/oasst1/resolve/main/2023-04-12_oasst_all.messages.jsonl.gz" },
    "prompts": { "train":"https://huggingface.co/datasets/OpenAssistant/oasst1/resolve/main/2023-04-12_oasst_prompts.messages.jsonl.gz" },
    "spam": { "train":"https://huggingface.co/datasets/OpenAssistant/oasst1/resolve/main/2023-04-12_oasst_spam.messages.jsonl.gz" },
}


class Oasst1(datasets.GeneratorBasedBuilder):
    """OpenAssistant Conversations"""

    VERSION = datasets.Version("1.0.0")

    # If you need to make complex sub-parts in the datasets with configurable options
    # You can create your own builder configuration class to store attribute, inheriting from datasets.BuilderConfig
    # BUILDER_CONFIG_CLASS = MyBuilderConfig

    # You will be able to load one or the other configurations in the following list with
    BUILDER_CONFIGS = [
        datasets.BuilderConfig(
            name="ready",
            version=VERSION,
            description="Messages of message trees in ready_for_export state",
        ),
        datasets.BuilderConfig(
            name="all",
            version=VERSION,
            description="All messages including spam, deleted and synthetic",
        ),
        datasets.BuilderConfig(
            name="prompts",
            version=VERSION,
            description="Only prompts",
        ),
        datasets.BuilderConfig(
            name="spam",
            version=VERSION,
            description="Messages which were deleted or rejected during spam review",
        ),
    ]

    DEFAULT_CONFIG_NAME = "ready"

    def _info(self):

        if self.config.name in _URLS:
            features = datasets.Features(
                {
                    "message_id": datasets.Value("string"),
                    "parent_id": datasets.Value("string"),
                    "user_id": datasets.Value("string"),
                    "created_date": datasets.Value("string"),
                    "text": datasets.Value("string"),
                    "role": datasets.Value("string"),
                    "lang": datasets.Value("string"),
                    "review_count": datasets.Value("int32"),
                    "review_result": datasets.Value("bool"),
                    "deleted": datasets.Value("bool"),
                    "rank": datasets.Value("int32"),
                    "synthetic": datasets.Value("bool"),
                    "model_name": datasets.Value("string"),
                    "detoxify": {
                        "toxicity": datasets.Value("float64"),
                        "severe_toxicity": datasets.Value("float64"),
                        "obscene": datasets.Value("float64"),
                        "identity_attack": datasets.Value("float64"),
                        "insult": datasets.Value("float64"),
                        "threat": datasets.Value("float64"),
                        "sexual_explicit": datasets.Value("float64"),
                    },
                    "message_tree_id": datasets.Value("string"),
                    "tree_state": datasets.Value("string"),
                    "emojis": datasets.Sequence(
                        feature={"name": datasets.Value("string"), "count": datasets.Value("int32")}
                    ),
                    "labels": datasets.Sequence(
                        feature={
                            "name": datasets.Value("string"),
                            "value": datasets.Value("float64"),
                            "count": datasets.Value("int32"),
                        }
                    ),
                }
            )
        else:
            raise RuntimeError("Unsupported configuration")

        return datasets.DatasetInfo(
            description=_DESCRIPTION,
            features=features,
            homepage=_HOMEPAGE,
            license=_LICENSE,
            citation=_CITATION,
        )

    def _split_generators(self, dl_manager):
        # TODO: This method is tasked with downloading/extracting the data and defining the splits depending on the configuration
        # If several configurations are possible (listed in BUILDER_CONFIGS), the configuration selected by the user is in self.config.name

        # dl_manager is a datasets.download.DownloadManager that can be used to download and extract URLS
        # It can accept any type or nested list/dict and will give back the same structure with the url replaced with path to local files.
        # By default the archives will be extracted and a path to a cached folder where they are extracted is returned instead of the archive

        print("_split_generators", self.config.name)
        urls = _URLS[self.config.name]
        data_dir = dl_manager.download_and_extract(urls)
        print("data_dir", data_dir)
        return [
            datasets.SplitGenerator(
                name=datasets.Split.TRAIN,
                # These kwargs will be passed to _generate_examples
                gen_kwargs={
                    "filepath": data_dir["train"],
                    "split": "train",
                },
            ),
            datasets.SplitGenerator(
                name=datasets.Split.VALIDATION,
                # These kwargs will be passed to _generate_examples
                gen_kwargs={
                    "filepath": data_dir["valid"],
                    "split": "dev",
                },
            ),
        ]

    # method parameters are unpacked from `gen_kwargs` as given in `_split_generators`
    def _generate_examples(self, filepath, split):
        print("_generate_examples", filepath, split)
        with open(filepath, encoding="utf-8") as f:
            for line in f:
                msg = json.loads(line)
                if "events" in msg:
                    msg.pop("events")
                if "emojis" in msg:
                    if msg["emojis"]:
                        msg["emojis"] = [{"name": k, "count": v} for k, v in msg["emojis"].items()]
                    else:
                        msg["emojis"] = None
                if "labels" in msg:
                    if msg["labels"]:
                        msg["labels"] = [
                            {"name": k, "value": v["value"], "count": v["count"]} for k, v in msg["labels"].items()
                        ]
                    else:
                        msg["labels"] = None

                for optional_field in ('parent_id','rank', 'model_name', 'emojis', 'labels', 'detoxify', 'review_result'):
                    if optional_field not in msg:
                        msg[optional_field] = None

                yield msg["message_id"], msg