MixlyGames/Orpheus_Zero-v1-dataset

Dataset Card for Orpheus Zero Pretraining Corpus

Dataset Details

Dataset Description

This dataset is a pre-tokenized Russian/English text corpus assembled for pretraining Orpheus Zero, a ~278M parameter causal language model trained from scratch. The corpus combines game lore/wiki text, literature, encyclopedic content, and forum data, cleaned and deduplicated, then tokenized with a custom BPE tokenizer.

• Curated by: MixlyGames
• Language(s): Russian, English
• License: Mixed/unverified — see Bias, Risks, and Limitations for a per-source breakdown. Do not treat as a single clean license.
• Tokenizer: Custom BPE, vocab_size=32,000

Dataset Sources

• Trained model: MixlyGames/Orpheus_Zero-base-0.3b

Uses

Direct Use

Intended for causal language model pretraining on Russian/English text, particularly narrative and game-lore-heavy domains. The dataset is provided already tokenized (input_ids only, no raw text field) using a custom BPE tokenizer trained specifically for this project.

⚠️ Important: This dataset is tokenized with a project-specific custom BPE tokenizer (vocab_size=32,000). Token IDs are not compatible with other tokenizers (e.g. GPT-2/LLaMA/BERT vocabularies) — reusing these input_ids with a different tokenizer's decoder will produce garbage output. If you intend to use this data with your own tokenizer, you will need to obtain the source text and re-tokenize it yourself; the source text is not directly included in this release. No compatibility with third-party tokenizers is guaranteed or supported.

Out-of-Scope Use

• Not suitable as raw/plaintext corpus without decoding via the original tokenizer.
• Not intended for direct use as in-game dialogue simulation data — wiki-sourced game content consists of lore summaries, character bios, and plot descriptions rather than verbatim in-game dialogue/scripts.
• Not vetted for PII removal beyond source-level cleaning; not recommended for downstream use requiring strict privacy guarantees without further audit.

Dataset Structure

Stored via Hugging Face datasets (save_to_disk format), single split (full), sharded internally across 43 Arrow files.

• Features: input_ids (token ID sequences only — no raw text field)
• Rows: 1,222,747
• Total tokens: 5,008,371,712 (~5.0B)
• Average sequence length: ~4,096 tokens/row
• Raw pre-tokenization size: ~20 GB
• Split: Single (full) — train/validation split is applied downstream at training time, not baked into the dataset itself.

Dataset Creation

Curation Rationale

Assembled to provide a Chinchilla-appropriate token budget (~5B tokens) for pretraining a 278M-parameter model, with an emphasis on narrative/game-lore-style Russian and English text to suit the target model's intended domain.

Source Data

The corpus draws from several categories of sources:

• Game lore/wiki text: Scraped from game-specific Fandom wikis (character bios, plot summaries, lore pages) via a custom scraping script, followed by cleaning to strip markup/navigation artifacts. Covers a wide range of titles (RPGs, visual novels, horror, indie games — see repository for the full source list).
• Literature: books, gutenberg_clean (public-domain texts)
• Encyclopedic: wikipedia_ru, wikisource_ru
• Forums/social: ruforum_clean, reddit_clean (games + science subcategories)
• Other: Songs, subtitles_ru

Data Collection and Processing

• Game lore/wiki text: Custom scraping script targeting Fandom wikis directly.
• Reddit / ruforum: Downloaded as pre-existing dumps from Hugging Face Hub (not scraped directly by this project). Original upstream source/license for these specific dumps was not recorded at collection time — treat as unverified provenance.
• Language filtering (Cyrillic/Latin character ratio thresholds) applied across text sources to remove cross-language leakage (e.g. Norwegian text in forum data).
• Deduplication via SimHash + LSH.
• Tokenization with a custom BPE tokenizer (vocab_size=32,000).
• Sharded and merged into a unified Hugging Face datasets object.

Who are the source data producers?

Mixed: Fandom wiki contributors (game lore pages), public-domain literary authors (Gutenberg), Wikipedia/Wikisource contributors, and forum/Reddit users (ru-language forums, Reddit games/science communities).

Personal and Sensitive Information

Not explicitly audited for PII. Forum and Reddit-sourced data may contain user-generated content with incidental personal information; no targeted anonymization was performed beyond standard text cleaning.

Bias, Risks, and Limitations

• License heterogeneity: Sources within this corpus carry different licensing statuses — Fandom wiki content is typically CC BY-SA (attribution + share-alike required for redistribution), Gutenberg text is public domain, and Reddit/ruforum dumps have unverified provenance/license (downloaded as third-party HF dumps without recorded upstream source). Treat this dataset's overall license as a wrapper around heterogeneous source material, not a blanket relicensing of that material.
• Tokenizer lock-in: As noted above, this data is only usable with the accompanying custom BPE tokenizer. Attempting to reuse it with a different tokenizer is unsupported and will not produce meaningful results — this is the responsibility of downstream users.
• Domain skew: Heavy weighting toward gaming/nerd-culture narrative content may bias a model trained on this data toward those registers of Russian/English.
• No held-out test set: Only a single full split is provided; any train/val partitioning must be done downstream.

Recommendations

Users should independently verify licensing compatibility of any derived/decoded text before redistribution, and should not assume interoperability with tokenizers other than the one this dataset was built for.

Contact

MixlyGames — see the Orpheus Zero model card for the trained model this dataset powers.