datasets
Training and evaluation data, with the modality, task and licence stated up front. Listed live from the Hugging Face Hub.
t2a-mommy
t2a-mommy
Female-voice ASMR corpus for the text2asmr project.
Previously published as aoxo/audios2.
Companion repos: aoxo/t2a-daddy (male voice),
aoxo/t2a-audios-v1 (the original v1 corpus).
Layout
path
what
<creator>/<title>.m4a
source audio, 48 kHz AAC, one folder per creator
<creator>/<title>.json
word-level Whisper large-v3 alignment ([] = skipped: near-silent or undecodable)
labels/qwen3omni.jsonl
non-speech ontology labels for gap clips… See the full description on the dataset page: https://huggingface.co/datasets/aoxo/t2a-mommy.atlas-32-turn-level-actor-critic
32. A turn-level actor-critic derived from the value of computation
1. Question and links
Read this first. The reading copy of this directory is t2ance/atlas-experiments under 32-turn-level-actor-critic/; the saved training steps and the per-token training arrays are only in the Hugging Face repository t2ance/atlas-32-turn-level-actor-critic.
Does a critic that predicts the return at the start of each turn, and is supervised there alone, learn on the… See the full description on the dataset page: https://huggingface.co/datasets/t2ance/atlas-32-turn-level-actor-critic.atlas-30-openmathreasoning-genselect-training
30. Does Qwen3.5-9B learn to use candidates on OpenMathReasoning?
1. Question and links
Trained by reinforcement learning on rows of eight OpenMathReasoning GenSelect candidates with one to seven of them correct, does Qwen3.5-9B's accuracy with eight candidates on held-out rows of the same distribution rise above its untrained accuracy and above the majority vote of the eight?
Report source:… See the full description on the dataset page: https://huggingface.co/datasets/t2ance/atlas-30-openmathreasoning-genselect-training.t2a-daddy
t2a-daddy
Male-voice ASMR corpus for the text2asmr project.
Previously published as aoxo/audios3.
Companion repos: aoxo/t2a-mommy (female voice),
aoxo/t2a-audios-v1 (the original v1 corpus).
Layout
path
what
<creator>/<title>.m4a
source audio, 48 kHz AAC, one folder per creator
<creator>/<title>.json
word-level Whisper large-v3 alignment ([] = skipped: near-silent or undecodable)
labels/qwen3omni.jsonl
non-speech ontology labels for gap clips… See the full description on the dataset page: https://huggingface.co/datasets/aoxo/t2a-daddy.atlas-31-strengthening-candidate-verification-under-rl
31. Strengthening candidate verification under reinforcement learning
1. Question and links
Read this first. The reading copy of this directory is t2ance/atlas-experiments under 31-strengthening-candidate-verification-under-rl/; the saved training steps and the per-token training arrays are on the Hugging Face repository t2ance/atlas-31-strengthening-candidate-verification-under-rl only.
How can reinforcement learning make the orchestrator's comparing and… See the full description on the dataset page: https://huggingface.co/datasets/t2ance/atlas-31-strengthening-candidate-verification-under-rl.t2a-audios-v1
t2a-audios-v1
The original text2asmr corpus (previously aoxo/audios): 48 kHz stereo ASMR audio with word-level
alignments, used for the v1 generator (Chatterbox speech LoRA, Stable Audio Open trigger LoRA) and as
the source for the reconstructed trigger ontology.
Superseded for ontology work by aoxo/t2a-mommy and
aoxo/t2a-daddy, which are larger, creator-attributed
and split by voice.
path
what
<id>.m4a
source audio, 48 kHz
<id>.json
word-level alignment + silence… See the full description on the dataset page: https://huggingface.co/datasets/aoxo/t2a-audios-v1.atlas-25-sequential-tool-runtime-upgrade
ATLAS report 25: the sequential tool runtime on verl V1
1. Question and links
Read this first. Every stage of the bring-up ran to its evidence; the report is complete for the correctness acceptance of issue 59 and for its performance stack (a second pass: the call parser fixed after an independent judgement, a boundary rollout at a 1024-token cap, one stacked performance ladder whose first tier, a48k, is now the campaign's default) and for its first research use:… See the full description on the dataset page: https://huggingface.co/datasets/t2ance/atlas-25-sequential-tool-runtime-upgrade.Urbansound8K_t2a
Dataset Card for "Urbansound8K_t2a"
More Information needed
MACS_t2a
Dataset Card for "MACS_t2a"
More Information needed
gigaspeech_t2aspoken-squad-t2aatlas-24-frozen-prefix-potential-shaping
ATLAS report 24: frozen-prefix potential shaping
1. Question and links
Read this first. This data root holds the first attempt of report 24 on the campaign's old harness (verl 0.7.1): the shaped training is complete and the unshaped training stopped at step 20 with a known problem (the subsection at the end of this section). The question was rerun on the runtime of report 25 with both trainings at 40 steps; that rerun's trajectories, exports, checkpoints and… See the full description on the dataset page: https://huggingface.co/datasets/t2ance/atlas-24-frozen-prefix-potential-shaping.T2AV_TemporalConstraintTalker-T2AV-Data
Talker-T2AV-Data
Joint Talking Audio-Video Generation with Autoregressive Diffusion Modeling
Paper (arXiv 2604.23586) ·
Code (GitHub) ·
Model ·
Samples
Clean training data package for Talker-T2AV. Paths in metadata/train.csv are relative to the dataset root after extracting the shard archives.
Contents
metadata/train.csv: training index used by Talker-T2AV.
shards/*.tar: clean archive shards grouped by modality and dataset source.
audio/, motion/, video/: directories… See the full description on the dataset page: https://huggingface.co/datasets/HKUSTAudio/Talker-T2AV-Data.atlas-28-explore-cost-in-dollars
ATLAS report 28: scaling the training up from step 29
1. Question and links
Read this first. This data root is published whole to the Hugging Face repository t2ance/atlas-28-explore-cost-in-dollars and, without the saved steps, the weight files and the per-token arrays, as the directory 28-explore-cost-in-dollars/ of the GitHub reading copy t2ance/atlas-experiments. The saved training steps are on the Hub only.
Question. Can a larger-scale training be brought up… See the full description on the dataset page: https://huggingface.co/datasets/t2ance/atlas-28-explore-cost-in-dollars.mat-01-lr-versus-learning-signal
01 Learning rate versus the learning signal
When the round-best score on a CPU-only Kaggle task stops rising during GRPO training, is the binding constraint the learning rate (too small to move the policy, or too large to keep it stable), or the learning signal itself (what the search samples and how the reward separates it)? This repository is the data root of that question: every training run's tree-search rollout archive it produced between 2026-06-29 and 2026-07-03, minus… See the full description on the dataset page: https://huggingface.co/datasets/t2ance/mat-01-lr-versus-learning-signal.atlas-26-pattern-teacher-sft-on-qwen3-6-27b
Qwen3.6-27B SFT
Question and links
Read this first: The reading copy is t2ance/atlas-experiments, report 26. Saved training weights are on this Hugging Face repository only. The reading copy is updated when publication completes.
The latest three-epoch SFT completed the full 175-question LCB evaluation. It scored 113/175, compared with 118/175 for the original one-epoch SFT and 134/175 for base. The new results are under lcb/sft_first80_3ep/. The comparison below… See the full description on the dataset page: https://huggingface.co/datasets/t2ance/atlas-26-pattern-teacher-sft-on-qwen3-6-27b.atlas-04-update-dose-and-entropy-collapse
ATLAS report 04: update dose and entropy collapse
1. Question and links
Read this first. This repository is the data root of report 04, closed. A reader who cannot fetch files from the Hub finds a reading copy of this whole data root in the private GitHub repository t2ance/atlas-experiments, directory 04-update-dose-and-entropy-collapse/: the same layout and file names, this page included, minus the per-token arrays of trajectory records, the .npz and .npy… See the full description on the dataset page: https://huggingface.co/datasets/t2ance/atlas-04-update-dose-and-entropy-collapse.mat-02-9b-capability-ceiling
02 9B capability ceiling
Is Qwen3.5-9B (19.3 GB of bf16 weights) a viable fast-iteration platform for the project's tree-search GRPO training, in place of the much larger Qwen3.6-27B (54 GB) student, without losing so much task capability that a cheaper training step stops being a useful learning iteration? The report's answer, quoting its abstract: "on these four tasks the 9B is not a viable fast platform" -- per node a 9B training step was 2.7 to 3.0x cheaper than the matching… See the full description on the dataset page: https://huggingface.co/datasets/t2ance/mat-02-9b-capability-ceiling.T2AV-Compass
Dataset Card for T2AV-Compass
Dataset Details
Dataset Description
T2AV-Compass is a unified benchmark for evaluating Text-to-Audio-Video (T2AV) generation, targeting not only unimodal quality (video/audio) but also cross-modal alignment & synchronization, complex instruction following, and perceptual realism grounded in physical/common-sense constraints.
Text-to-Audio-Video (T2AV) generation aims to synthesize temporally coherent video and semantically… See the full description on the dataset page: https://huggingface.co/datasets/NJU-LINK/T2AV-Compass.atlas-16-verifier-permission-prompt-ablation
ATLAS report 16: does the orchestrator's "cannot solve" clause suppress candidate verification?
Complete raw products of the ATLAS rl-training report 16 experiment
(GitHub issue #36). Two system-prompt arms of the same model over the
same 78 fixed states, greedy decoding, one shared vLLM server.
What the experiment did
The ATLAS orchestrator's frozen system prompt contains the clause
You cannot solve the problem yourself; you decide when to explore
further and… See the full description on the dataset page: https://huggingface.co/datasets/t2ance/atlas-16-verifier-permission-prompt-ablation.atlas-22-prefix-accuracy-curves-neutral-vs-submit
ATLAS report 22: prefix accuracy curves k = 1..8 on the neutral selector and on the neutral-text submit surface
Complete raw products of ATLAS rl-training report 22 (GitHub issue #47).
Report 21's two ATLAS-free selector surfaces, run over every identity prefix
k = 1..8 of every question of the canonical LiveCodeBench (175) and GPQA (198)
validation sets: 373 questions x 8 prefixes x 2 surfaces = 5968 states, one
greedy deliberative rollout each, on the same Qwen3-14B checkpoint… See the full description on the dataset page: https://huggingface.co/datasets/t2ance/atlas-22-prefix-accuracy-curves-neutral-vs-submit.t2adata3atlas-15-selector-capacity-vs-training
ATLAS report 15 -- selector capacity vs training (raw products)
Every raw product of ATLAS rl-training report 15, uploaded for the raw-trace
audit asked for in t2ance/ATLAS issue #13.
Nothing is filtered or hand-picked: all 78 states are here for all five model
rows, both surfaces, both modes, both rollouts.
The experiment in one paragraph
39 LiveCodeBench questions whose 8-candidate Sonnet cache holds at least one
correct and one incorrect candidate that carries… See the full description on the dataset page: https://huggingface.co/datasets/t2ance/atlas-15-selector-capacity-vs-training.sounddescs_t2a
SoundDescsT2ARetrieval
An MTEB dataset
Massive Text Embedding Benchmark
Natural language description for different audio sources from the BBC Sound Effects webpage.
Task category
Any2AnyRetrieval (text-to-audio)
Domains
Encyclopaedic, Written
Reference
IEEE Transactions on Multimedia
Source datasets:
mteb/sounddescs_t2a
How to evaluate on this task
You can evaluate an embedding model on this dataset using the following code:
import mteb
task =… See the full description on the dataset page: https://huggingface.co/datasets/mteb/sounddescs_t2a.audiocaps_t2a
AudioCapsT2ARetrieval
An MTEB dataset
Massive Text Embedding Benchmark
Natural language description for any kind of audio in the wild.
Task category
t2a
Domains
Encyclopaedic, Written
Reference
https://audiocaps.github.io/
Source datasets:
mteb/audiocaps_t2a
How to evaluate on this task
You can evaluate an embedding model on this dataset using the following code:
import mteb
task = mteb.get_task("AudioCapsT2ARetrieval")
evaluator = mteb.MTEB([task])… See the full description on the dataset page: https://huggingface.co/datasets/mteb/audiocaps_t2a.atlas-23-prefix-curves-on-a-larger-selector
ATLAS report 23: prefix curves on a larger selector
1. Question and links
Read this first. GPQA ran in full on both surfaces (198 questions, prefix lengths k = 1 to 8, 1584 states each). LiveCodeBench was started and stopped by the user at 455 and 531 of its 1400 states per surface and is not read: every table and the report read GPQA only. A reader who cannot fetch files from the Hub finds this whole data root mirrored in the private GitHub repository… See the full description on the dataset page: https://huggingface.co/datasets/t2ance/atlas-23-prefix-curves-on-a-larger-selector.atlas-17-prompt-interpolation-selector-recovery
ATLAS report 17: which ATLAS prompt component suppresses candidate discrimination?
Complete raw products of ATLAS rl-training report 17 (GitHub issue #38).
Eight prompt arms interpolate from the frozen ATLAS orchestrator prompt
to an empty system message with a neutral closing, plus the neutral A/B
selector rerun on the same server as the interpolation anchor. One
Qwen3-14B, 78 fixed states, greedy, one vLLM process.
Start here
report/report17.pdf, seven pages… See the full description on the dataset page: https://huggingface.co/datasets/t2ance/atlas-17-prompt-interpolation-selector-recovery.atlas-20-information-barrier-and-the-prompt-endpoint
ATLAS report 20: does the information barrier suppress verification, and where does the prompt line end?
Complete raw products of ATLAS rl-training report 20 (GitHub issue #43).
Two new selector surfaces over every question of the canonical LiveCodeBench
(175) and GPQA (198) validation sets, each question with all eight of its
cached candidates revealed. Both carry report 18's comparison closing and
change only the system message:
barrier relaxed — the finalization paragraph's… See the full description on the dataset page: https://huggingface.co/datasets/t2ance/atlas-20-information-barrier-and-the-prompt-endpoint.atlas-18-full-benchmark-closing-selector-transfer
ATLAS report 18: does the comparison closing carry ATLAS to the neutral selector on full GPQA and LiveCodeBench?
Complete raw products of ATLAS rl-training report 18 (GitHub issue #39).
Three selector surfaces over every question of the canonical LiveCodeBench
(175) and GPQA (198) validation sets, each question with all eight of its
cached candidates revealed: the current ATLAS forced-submit surface, the
same surface with only its closing user message replaced by a… See the full description on the dataset page: https://huggingface.co/datasets/t2ance/atlas-18-full-benchmark-closing-selector-transfer.