TASTE-IF-SFT-48K

A 48K-example English speech-oriented Instruction-Following (IF) SFT dataset, built for fine-tuning Spoken Language Models (SLMs) such as TASTE on basic content-level instruction following (as a precursor to paralinguistic / voice-style instruction following, e.g. VStyle-style tasks).

Motivation

VStyle-like prompts (e.g. "With a very slow pace, please count from one to five.") combine two abilities:

1. Content-level instruction following — understanding what to say (e.g. "count from one to five" -> "One, two, three, four, five.")
2. Paralinguistic / acoustic control — controlling how it is said (e.g. speaking slowly)

This dataset isolates ability (1): every instruction is a short, speech-friendly English command without any paralinguistic/acoustic conditioning words (slow, fast, loud, whisper, angry, sad, happy, tone, pitch, volume, etc.). Each example also includes synthesized speech audio for both the instruction (user turn) and the expected spoken response (assistant turn), so the dataset can be used directly for SFT of speech-in/speech-out models.

Configs

This repo provides two configs of the same 48K examples:

• default (under data/): includes top-level instruction_audio / response_audio columns stored as the standard HF Audio() struct ({"bytes", "path"}), so the Dataset Viewer can play the audio directly in the browser. Recommended for browsing and for general-purpose use with datasets.load_dataset.
• training (under data_training/): the exact format expected by the TASTE SFT data loader (sft_processor.py) — each message entry carries its own audio (raw bytes, not wrapped in an Audio() feature) and there are no top-level instruction_audio/response_audio columns. Use this config if you want to feed the data directly into TASTE SFT training:

  ds = load_dataset("Jaylin0418/TASTE-IF-SFT-48K", "training")
  

Dataset Structure

• Splits: 11 train shards (shuffled_train_part_0000.parquet ... shuffled_train_part_0010.parquet, 4,000 examples each, 44,000 total) + 1 dev shard (shuffled_dev.parquet, 4,000 examples). Total = 48,000 examples.
• Format: Parquet, one row per dialogue example.

Columns

• idx (str): unique example id, prefixed by the instruction-following ability category (e.g. read_aloud_047388).
• system_prompt (str): a generic system prompt describing the IF task (same for every example). Provided as a convenience starting point; replace it as needed for your own training setup.
• meta (dict):
  • instruction_type (str): the instruction-following ability category (see below)
• instruction_audio / response_audio (Audio): the synthesized speech audio for the instruction (user turn) and the expected spoken response (assistant turn).
• message (list[dict]): a 2-turn conversation, one user turn and one assistant turn, each with:
  • role (str): "user" or "assistant"
  • text (str): the instruction (user) or expected target response (assistant)
  • timestamp_range (list[int]): start/end timestep range for this turn within the full dialogue audio
  • para_tags (str or null, JSON-encoded): paralinguistic tags applied during TTS synthesis of this turn's audio (e.g. {"emotion": "surprised"}). Always null for the user turn, since the instruction audio is always synthesized in a neutral/calm voice; only the assistant (response) audio is rendered with the requested style.

Example

{
  "idx": "counting_001234",
  "system_prompt": "You are a helpful spoken assistant. Listen to the user's spoken instruction and respond with only the requested spoken content, following the instruction exactly.",
  "meta": {"instruction_type": "counting"},
  "instruction_audio": {"bytes": b"...", "path": None},
  "response_audio": {"bytes": b"...", "path": None},
  "message": [
    {"role": "user", "text": "Please count from one to five.", "timestamp_range": [0, 3000], "para_tags": "{}"},
    {"role": "assistant", "text": "One, two, three, four, five.", "timestamp_range": [3000, 5500], "para_tags": "{}"}
  ]
}

Ability Categories (meta.instruction_type)

read_aloud, counting, sequence, reverse_sequence, listing, repetition, spelling,
number_reading, time_date_reading, format_constraint, negative_constraint,
required_word, word_extraction, replacement, filtering, selection, ordering,
comparison, completion, transformation, short_description, short_generation,
simple_arithmetic, conditional, multi_step

Loading

from datasets import load_dataset

ds = load_dataset("Jaylin0418/TASTE-IF-SFT-48K")
print(ds)
print(ds["train"][0]["message"])

Intended Use

This dataset was used to SFT TASTE on basic spoken instruction-following ability before evaluating paralinguistic / voice-style instruction following (VStyle-like tasks). It can be used to fine-tune any speech-in/speech-out SLM that consumes paired (instruction audio, instruction text, response audio, response text) examples.