DiCoTA

DiCoTA is a text-only derived annotation dataset for difficulty-adaptive compressed chain-of-thought audio reasoning. It is derived from the annotations of Audio-Reasoner-CoTA. This repository does not redistribute audio. Each row retains the stable source coordinates needed to retrieve the corresponding CoTA audio row.

The S2TT subset is excluded from this release.

Splits and counts

Split	Rows
sft	257,440
rl	110,351
Total	367,791

Subset	SFT	RL	Total
audiocap	68,220	29,241	97,461
clotho	5,967	2,558	8,525
complex_audio	13,410	5,747	19,157
meld	20,423	8,752	29,175
multispeaker	59,694	25,586	85,280
musicbench	89,726	38,467	128,193

From the 368,047 non-S2TT rows in the starting full split, this release excludes:

• 87 rows in the reviewed malformed compressed-answer denylist; see exclusions/quality_exclusions.jsonl.
• 169 non-S2TT rows whose compressed answers contain repeated opening tags; see exclusions/repeated_structure_exclusions.jsonl.

Loading

from datasets import load_dataset

ds = load_dataset("assasinatee/DiCoTA", split="sft")
print(ds[0])

Joining audio back from CoTA

This dataset stores only derived text annotations. Audio is not uploaded here. Use:

source_config
source_split
source_row

to index the corresponding row from zhifeixie/Audio-Reasoner-CoTA.

Conceptually:

from datasets import load_dataset

row = ds[0]
cota = load_dataset(
    "zhifeixie/Audio-Reasoner-CoTA",
    row["source_config"],
    split=row["source_split"],
)
cota_row = cota[row["source_row"]]
audio = cota_row["audio"]

For efficient batch use, group DiCoTA rows by (source_config, source_split) and reuse each loaded CoTA split.

Schema

Field	Description
id	Stable derived-row ID: <source_file>#<source_row>.
split	sft or rl.
subset	Source subset name.
task_type	choice or open, derived from the source file.
source_config	CoTA dataset config name.
source_split	CoTA split name.
source_file	Source annotation JSON path used to build this row.
source_row	Zero-based row index in the corresponding CoTA split.
user	Original question, including its <audio> marker.
assistant	Compressed chain of thought with rating tokens.
Reasoning_Demand	Integer teacher score in [1, 5].
Integration_Score	Integer teacher score in [1, 5].
Question_Quality	Integer teacher score in [1, 5].
PQ, PC	AudioBox perceptual-quality and perceptual-complexity scores.
target_total, orig_len	Compression metadata retained from generation.

Reproducibility

manifest/manifest.json records:

• final split and subset counts,
• exclusion policy and counts,
• output Parquet sizes, row counts, and SHA-256 hashes,
• hashes for the 75 source annotation JSON files.

Source

This derived annotation dataset was built from Audio-Reasoner-CoTA.