THE CURIOUS SIDE OF AI
TEXTDATASET PREPARATIONsmol-smoltalk
A conversation mixture designed for small-model instruction tuning. Select a small, relevant subset; the full dataset is not tiny.
Recipe-compatible after subset preparationRead guide →
AIBRAINWORX / DATASET LIBRARY
Small experiments.
Useful training data.
Six source-reviewed preparation guides for text, voice and vision workflows. Read formats, compatibility and rights here before visiting the source.
Data guides, not bundled datasets
Our experimental trainer accepts local text JSONL and small floating-point text checkpoints. Audio and image tasks use separate trainers. A schema match is not a successful training run or blanket rights clearance.
TEXTDATASET PREPARATION
TEXTDATASET PREPARATIONultrachat_200k
Synthetic multi-turn conversations. Keep only the needed messages field and maintain a held-out evaluation set.
Recipe-compatible after split/field selectionRead guide →
VISIONDATASET PREPARATIONthe_cauldron
A collection of visual task datasets. Source-specific image rights vary; no blanket license clearance is claimed.
Schema adaptation + subset rights review requiredRead guide →
VISIONDATASET PREPARATIONbeans
A small bean-leaf image classification dataset. This is a classifier workflow, not direct text LoRA data.
Separate classification trainer requiredRead guide →
ASRDATASET PREPARATIONlibrispeech_asr
Transcribed speech for ASR. The full collection is large; start with an explicitly selected subset.
Separate ASR trainer + prepared subset requiredRead guide →
TTSDATASET PREPARATION
LJ Speech 1.1
The original publisher provides 13,100 clips from one speaker, approximately 24 hours. Use the official source, not an unverified mirror.
Resampling and audio/text/ref_audio JSONL preparation requiredRead guide →
Choose a task before a dataset
Begin with a small relevant subset, inspect representative examples and keep evaluation data separate. Retain original source terms, speaker/image rights and required notices. The launcher never runs remote dataset Python code or downloads a whole collection automatically.
Find a compatible model workflow →