Datasetpublic

AdoCleanCode/llasa_stage1_turkish_tinystories_phonemes

LLASA Stage 1 Training Dataset (Text-Phoneme Conversion) This dataset contains filtered text-phoneme pairs for Stage 1 curriculum learning in LLASA. Purpose Stage 1 training teaches the model text ↔ phoneme conversion using LLAMA's existing text understanding capabilities. Columns phoneme_text_sequence: Phoneme tokens → Text conversion sequence ("<|ph_0068|>... Convert into text: Hello world") text_phoneme_sequence: Text → Phoneme tokens conversion… See the full description on the dataset page: https://huggingface.co/datasets/AdoCleanCode/llasa_stage1_turkish_tinystories_phonemes.

sourceHugging Faceupdated 8mo agoView on Hugging Face

0likes37downloads

Dataset Card

LLASA Stage 1 Training Dataset (Text-Phoneme Conversion)

This dataset contains filtered text-phoneme pairs for Stage 1 curriculum learning in LLASA.

Purpose

Stage 1 training teaches the model text ↔ phoneme conversion using LLAMA's existing text understanding capabilities.

Columns

—phoneme_text_sequence: Phoneme tokens → Text conversion sequence ("<|ph_0068|>... Convert into text: Hello world")
—text_phoneme_sequence: Text → Phoneme tokens conversion sequence ("Hello world Convert into phonemes: <|ph_0068|>...")
—text: Original text for reference
—phonemes_ipa: Original IPA phonemes for verification
—phonemes_hex: Hex-encoded phoneme tokens for verification

Filtering

The dataset has been filtered to exclude:

—Invalid MFA phonemizations where phonemes equal the normalized text (fallback cases like "spn")

Usage

Use this dataset for Stage 1 training with --training_stage words_phonemes in train_llasa_with_phonemes.py.

Total samples: 199656