MovieGPT

An English-language, document-oriented dialogue corpus prepared for causal language-model training. All sources use one shared schema and one set of train/validation/test splits. The source field records provenance without separating the corpora into different dataset configurations.

Current data

Source	Documents	Train	Validation	Test
Friends	236	—	—	—
Cornell Movie-Dialogs	617	—	—	—
OpenSubtitles v2024 English	2092508	—	—	—
Unified total	2093361	1883987	105007	104367

Every row is one complete source document with a stable document_id and a text field. Friends documents correspond to episodes, Cornell documents to movies, and OpenSubtitles documents to artificial 512-line chunks because the OPUS plain-text archive does not preserve movie boundaries. Utterances are separated by newlines and every document ends with GPT-2's <|endoftext|> token. Train, validation, and test assignment is a stable SHA-256-based 90/5/5 split, so adding sources later will not reshuffle existing documents.

Load for GPT-2

from datasets import load_dataset
from transformers import AutoTokenizer

dataset = load_dataset("relentlessml/moviegpt")
tokenizer = AutoTokenizer.from_pretrained("gpt2")
block_size = min(tokenizer.model_max_length, 1024)

tokenized = dataset.map(
    lambda batch: tokenizer(batch["text"], add_special_tokens=False),
    batched=True,
    remove_columns=dataset["train"].column_names,
)

def group_texts(batch):
    concatenated = {key: sum(batch[key], []) for key in batch}
    usable = len(concatenated["input_ids"]) // block_size * block_size
    result = {
        key: [
            values[index:index + block_size]
            for index in range(0, usable, block_size)
        ]
        for key, values in concatenated.items()
    }
    result["labels"] = [ids.copy() for ids in result["input_ids"]]
    return result


language_model_data = tokenized.map(group_texts, batched=True)

Use dataset.filter(lambda row: row["source"] == "friends") when a single source is needed. The Hub repository intentionally exposes one unified configuration.

Fields

• document_id: globally unique, stable ID prefixed by source.
• source and source_id: corpus provenance.
• title, language, and license: document metadata.
• text: GPT-2-ready document text ending in <|endoftext|>.
• text_sha256, character_count, byte_count, and line_count: integrity and auditing fields.

Cleaning and provenance

Friends excludes transcript notes and nonverbal-only rows, removes stage directions and transcript-parser artifacts, preserves scene boundaries, and does not serialize speaker or annotation metadata into the training text. Cornell Movie-Dialogs reconstructs each reply chain, removes empty utterances and known screenplay-formatting tags, and preserves conversation boundaries. OpenSubtitles removes subtitle markup, nonverbal cues, music-marked lines, caption credits, and immediate duplicates before fixed-size chunking.

The Friends source is ConvoKit's Friends Corpus, whose documentation identifies the original Emory NLP data as Apache-2.0. Cornell data comes from ConvoKit's Cornell Movie-Dialogs Corpus, and subtitle text comes from OPUS OpenSubtitles v2024. Licensing is recorded per document because the sources use different terms. Users remain responsible for reviewing source-specific rights before redistribution or model release.

Build metadata and counts are recorded in dataset_summary.json.