Skip to content

project-felys/delta-me13

Folders and files

NameName
Last commit message
Last commit date

Latest commit

 

History

1 Commit
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 

Repository files navigation

δ-me13

δ-me13 generates text and audio datasets for Cyrene, and can be extended to a wider range of applications. Feel free to chat with Cyrene or hear her voice.

Note: this project does not contain any game assets, nor does it reverse engineer any encrypted files.

Setup

The project isn't published to a package index, so install it directly from git.

pip install git+https://github.com/project-felys/delta-me13.git

However, since upstream dependencies can be extremely fragile, I recommend cloning the repository and patching it as needed. You can set it up with pip install -e . or uv sync. Feel free to ask an LLM for help, as some of the data mapping is complicated.

Large-Language-Model

The corpus generation scripts depend on an external game data repository, which I will not name here. If you find that repository, clone it and replace /path/to/game-data-repository with its path. The following commands build the dataset in standard format for all 13 languages.

REPO=/path/to/game-data-repository

# Pre-training
delta-me13 multilingual \
    --turn-based-game-data-dir $REPO \
    --namespace pt \
    --dataset amphoreus \
    --num-proc 13

# Supervised Fine-tuning
delta-me13 multilingual \
    --turn-based-game-data-dir $REPO \
    --namespace sft \
    --dataset cyrene

The vendor data includes the official game wiki and LeetCode problems from COIG.

# Pre-training
delta-me13 vendor --vendor-dir vendor

Text-to-Speech

Generating the audio dataset requires both the external repository (see the previous section) and the game itself. You will need to unpack the audio files first. Replace /path/to/persistent/audio/audio-package/windows with the game's audio directory.

PCK=/path/to/persistent/audio/audio-package/windows

# Unpack *.pck
delta-me13-pck --input-dir $PCK

Once all .pck files have been unpacked, the command-line interface can process the corpus and generate .wav files. Make sure vgmstream-cli is in your PATH, or run export VGMSTREAM=/path/to/vgmstream-cli. Refer to vgmstream for installation guidance.

REPO=/path/to/game-data-repository

# Text-to-Speech
delta-me13 audio \
    --turn-based-game-data-dir $REPO \
    --unpacked-audio-dir audio \
    --dataset cyrene

License

Distributed under the terms of the LICENSE.

Copyright

© All rights reserved by FelysNeko.

Contributors

Languages