JimHue/Taste-S-SLM-TrainingData-bidirectional

TASTE-S Bidirectional SLM Training Data

This repository contains two parallel SLM training-data variants and the tokenizer checkpoint used to produce TASTE speech tokens.

Data variants

• slm_training_data/ctc_text_conditioned/: existing SLM training data conditioned on the bidirectional tokenizer's CTC text.
• slm_training_data/parakeet_tdt_0.6b_v2_text_conditioned/: SLM training data using Parakeet TDT 0.6B v2 ASR text with the same tokenizer pipeline.

Each variant contains the following corpus directories: emilia_en_small, gigaspeech, libritts, resource_management, vctk, voxpopuli_en, wsj0, deepdialogue_turn_sequence, and spokenqa_turn_sequence. Dataset files retain their split/shard hierarchy and .npz format.

Tokenizer checkpoint

tokenizers/stage1_novq_author_recipe_fast_ourdata_ctc_bidir_spokenqa_asrtext/ contains the valid_best checkpoint (config.json and model.safetensors). The checkpoint uses bidirectional CTC attention configuration.

Incremental Tok014 variant

slm_training_data/parakeet_tdt_0.6b_v2_text_conditioned_stage2_fsq_tok014/ contains the incremental Tok014 data variant, generated with Parakeet TDT 0.6B v2 ASR text and the TASTE-S Stage-2 FSQ bidirectional tokenizer. Upload begins while extraction is still running. Check slm_training_data/parakeet_tdt_0.6b_v2_text_conditioned_stage2_fsq_tok014/_UPLOAD_STATUS.json before using it; the data are partial until its state is complete. This upload contains NPZ shards only; it does not include raw audio, ASR JSONL files, processing logs, or the tokenizer checkpoint.

Provenance

The Parakeet-conditioned data uses nvidia/parakeet-tdt-0.6b-v2 transcripts. No source audio, intermediate ASR JSONL files, manifests, processing logs, or temporary prepared-job files are included. Refer to the dataset source projects and papers for corpus-specific attribution and usage terms.