FlexiSLM/FlexiSLM-Data-2M-s2s-compact

FlexiSLM-Data — Speech-to-Speech Part (2.43M filtered samples, 385G in size)

• Paper: https://arxiv.org/abs/2606.31247
• Demo page: https://flexislm.github.io/
• Code: https://github.com/AmphionTeam/FlexiSLM

FlexiSLM-Data is a large-scale, single-turn English speech-to-speech dialogue dataset for training FlexiSLM, a spoken language model. This repository contains the paired prompt-and-response audio portion of the release in WebDataset format.

Related data releases

• FlexiSLM/FlexiSLM-Data-4M-s2s is the larger speech-to-speech release with prompt deduplication and WER-based filtering. Its question audios are stored in 44k wav, while response audios are stored in mp3. Size is around 2.8T.
• FlexiSLM/FlexiSLM-Data-2M-s2s-compact (this repo) is the filtered, mp3-compressed subset of FlexiSLM/FlexiSLM-Data-4M-s2s that takes only 385G storage while maintaining high data quality.
• FlexiSLM/FlexiSLM-Data-5M-t2t provides the raw text input-output pairs without audio.

Data construction pipeline

1. Prompt collection and response generation (released in FlexiSLM/FlexiSLM-Data-5M-t2t) Text prompts are collected from public QA, instruction-following, and dialogue datasets (see the table below). For multi-turn datasets, the user's first-turn utterance is used as the prompt; samples whose first turn is a generic greeting (e.g. "Hello") are skipped. Then, all text responses are generated with Qwen3-Omni-30B-A3B; the responses shipped with the source datasets are discarded. We use a spoken language model rather than a text-only LLM because Qwen3-Omni produces speech-friendly responses: short, conversational, and free of formatting that does not transfer to speech (bullet points, code blocks, long enumerations). The text version of the prompt is fed to the model rather than its synthesized speech, since text-input responses are typically more accurate.
2. Speech synthesis. (released in FlexiSLM/FlexiSLM-Data-4M-s2s) Responses are synthesized with Qwen3-TTS using the fixed speaker "Ryan". Prompts are synthesized with Fish-Audio TTS, with speaker prompts randomly sampled from English Emilia utterances longer than 5 seconds. This yields the 4,222,459 samples and 26,735 hours of audio shipped here.
3. Quality filtering and mp3-format compression. (released in FlexiSLM/FlexiSLM-Data-2M-s2s-compact) Before any shards are written, this release applies prompt deduplication and response-text filtering. The kept subset uses normalized prompt deduplication first, then drops responses that contain !, contain ?, look non-English, or look like code. From 4,222,459 raw rows, 1,790,681 are filtered out and 2,431,778 remain. The punctuation-based filtering is empirical: we find that questions with more knowledge density are responded more formally, while casual questions tend to be responded with ! or ?.

Prompt sources

Only the user prompts come from these datasets; their original answers are not used.

Dataset	Type	# Prompts
TriviaQA	QA	138K
WebQuestions	QA	3.8K
TyDiQA	Multilingual QA	167K
Alpaca	Instruction	52K
SmolTalk2	Instruction / Dialogue	385K
SODA	Dialogue	1.48M
Magpie-Pro	Instruction / Dialogue	1M
UltraChat	Multi-turn Dialogue	949K
HH-RLHF	Dialogue / Preference	167K
WildChat	Real-user Dialogue	159K

Statistics of this data

	value
Samples (train)	2,425,778 across 243 shards
Samples (validation)	6,000 across 1 shard
Samples (total)	2,431,778
Filtered out	1,790,681 of 4,222,459 raw rows
Prompt dedup	91,869 duplicate prompts removed
Punctuation filter	1,375,207 removed (! / ?)
Non-English filter	116,412 removed
Code filter	207,193 removed
Prompt audio	4,088.9 hours (14,720,133.735 s)
Response audio	10,728.2 hours (38,621,498.771 s)
Total audio	14,817.1 hours (53,341,632.505 s)
Mean prompt duration	6.05 s
Mean response duration	15.88 s

Layout

Each sample is three tar members sharing an eight-digit key, unique across every shard and both splits:

00000001.question.mp3   # user prompt, transcoded from WAV during packing
00000001.response.mp3   # assistant reply, Qwen3-TTS speaker "Ryan"
00000001.json           # transcripts and per-sample metadata

• train: 2425778 samples across 243 shards (data/train-{00000..00242}-of-00243.tar)
• validation: 6000 samples across 1 shard (data/validation-{00000..00000}-of-00001.tar)

Exact per-shard sample counts and byte offsets live in shards.json; per-sample records live in manifest.jsonl.

Each metadata record carries:

field	description
key	Eight-digit WebDataset key
uuid	Stable identifier, shared with the t2t release
split	train or validation
shard	Shard the sample lives in
question_text / response_text	Transcripts of the two audio members
question_duration / response_duration	Seconds
total_audio_duration	Seconds
num_tokens_est	Estimated audio tokens at 12 tokens/s

Loading

from datasets import load_dataset
dataset = load_dataset("FlexiSLM/FlexiSLM-Data-2M-s2s-compact", split="train", streaming=True)

Or straight from WebDataset, using the brace pattern recorded in shards.json:

import webdataset as wds
url = "https://huggingface.co/datasets/FlexiSLM/FlexiSLM-Data-2M-s2s-compact/resolve/main/data/train-{00000..00099}-of-01406.tar"
dataset = wds.WebDataset(f"pipe:curl -sL {url}")

Filtering

Filtering is applied before any tar shards are written. The packer first performs prompt deduplication, then keeps only responses that pass all of the following checks:

• response does not contain !
• response does not contain ?
• response does not look non-English
• response does not look like code

The punctuation-based filtering is empirical: we find that questions with more knowledge density are responded more formally, while casual questions tend to be responded with ! or ?. Higher knowledge density data aligns with knwoledge benchmarks like llama-questions better.

The filtered release was built from 4,222,459 raw rows. The exact drop counts are:

reason	dropped
duplicate prompts	91,869
contains !	966,181
contains ?	409,026
looks non-English	116,412
looks like code	207,193

Total removed: 1,790,681 rows.

Limitations

• Single-turn only. Multi-turn sources are truncated to their first user turn.
• Filtered, but not perfect. The release removes the most obvious punctuation, non-English, and code-like responses, but some imperfect samples may still remain.
• Synthetic speech. All audio is TTS-generated, with a single fixed response voice; it does not reflect real conversational acoustics, noise, or speaker diversity on the response side.
• Model-generated text. Every response comes from Qwen3-Omni-30B-A3B and inherits its biases and factual errors; none are human-verified.
• Derived data. Prompts inherit the licenses and terms of their source datasets; please check each source before redistribution.

References

• Qwen3-Omni — Xu et al., 2025
• Qwen3-TTS — Hu et al., 2026
• Fish-Audio TTS — Liao et al., 2024
• Emilia — He et al., 2024
• MLS — Pratap et al., 2020
• LibriSpeech — Panayotov et al., 2015
• LLaSO-Instruct — Sun et al., 2025
• TriviaQA — Joshi et al., 2017
• WebQuestions — Berant et al., 2013
• TyDiQA — Clark et al., 2020
• Alpaca — Taori et al., 2023
• SODA — Kim et al., 2023
• Magpie — Xu et al., 2025
• UltraChat — Ding et al., 2023
• HH-RLHF — Bai et al., 2022
• WildChat — Zhao et al., 2024