DecisionTraceBench

Tracking Decision-Critical Information in Multi-Party Spoken Conversations

DecisionTraceBench is a Chinese speech benchmark for recovering complete decision-critical information trajectories from multi-party financial conversations. It jointly evaluates event sequences, speaker attribution, business scope, values, predecessor relations and temporal evidence.

🤗 Dataset | 💻 Code and prompts

Dataset composition

Subset	Recordings	Hours	Families	Trajectories	Events	Audio	License
Human-performed	26	8.61	16	308	562	Original MP3	CC BY-NC 4.0
Synthetic	2,246	100.39	1,456	5,319	18,227	Original 24 kHz mono WAV	CC BY-NC 4.0
Total	2,272	109.01	1,472	5,627	18,789		CC BY-NC 4.0

Both subsets are fixed research evaluation cohorts, exposed as test; no training or development split is provided. A/B variants share dialogue-family membership and must be kept together in family-aware analyses. The 26-recording human subset differs from the paper's common 25-recording main comparison cohort.

Files

README.md
LICENSE
assets/
  decisiontracebench-banner.png
  label_vocabulary.json      # Shared prediction vocabulary: 193 field pairs / 64 scopes
human/
  README.md
  LICENSE.txt
  label_definitions.json
  audio/                     # 26 MP3 recordings and metadata.jsonl
  gold/                      # 26 original gold JSON files
synthetic/
  README.md
  LICENSE.txt
  label_definitions.json
  audio/                     # 2,246 WAV recordings and metadata.jsonl
  gold/                      # 2,246 original gold JSON files

Within each subset, the samples array in label_definitions.json pairs audio with gold by sample_id and records family membership, scenario, speaker count, duration and SHA-256 hashes. Paths in this index are relative to the subset directory. Sample identifiers are subset-local.

Load with Hugging Face Datasets

Use Hugging Face Datasets with its audio dependencies (datasets[audio]) and a compatible FFmpeg runtime. Audio(decode=False) avoids decoding when accessing a row; dataset preparation may still require the audio dependencies.

import json
from datasets import Audio, load_dataset

human = load_dataset("QFIN/DecisionTraceBench", "human", split="test")
synthetic = load_dataset("QFIN/DecisionTraceBench", "synthetic", split="test")

# Inspect annotations without decoding the waveform.
row = human.cast_column("audio", Audio(decode=False))[0]
gold = json.loads(row["gold_json"])
assert gold["sample_id"] == row["sample_id"]

Each AudioFolder row includes audio, sample_id, family_id, variant, scene_code, speaker_count, duration_seconds, trajectory_count, event_count, gold_path, gold_json and license. gold_path is relative to the repository root. gold_json stores the complete annotation as a JSON string, preserving mixed value representations without forcing them into one Arrow struct type. Original gold files are also retained under gold/.

metadata.jsonl is an access index generated from the original sample index and annotations; it does not replace the original gold files. Ground truth and sample-specific answer metadata are evaluation material and must be excluded from prediction inputs. The dataset does not include verbatim transcripts, construction scripts, participant identity mappings or voice-cloning reference clips.

For raw audio and gold files, download the repository with huggingface_hub.snapshot_download and follow each subset's sample index. This also retrieves the label definitions, which are not individual AudioFolder rows.

Task and evaluation

An entity–field pair identifies a trajectory. Each trajectory contains ordered decision cycles, and each cycle contains effective events with action, speaker_id, scope_id, value, source_event_id and millisecond evidence windows. Actions are start, advance, hold and settle. A cycle can begin with an observed hold or settle; do not invent an earlier start. Settlement closes a cycle, and reopening requires explicit evidence.

Cycles are identified by their position and order in the cycles array. Predictions do not need cycle_id. Original gold files retain this legacy label for compatibility; it is ignored by the scorer. Cycle count, boundaries and chronological order remain required.

For paper-comparable prediction, both subsets use the same task-wide input vocabulary: 193 entity–field pairs and 64 scopes. Use assets/label_vocabulary.json or the complete GitHub task templates. Do not narrow this vocabulary using the subset, scenario, individual recording or gold annotations.

The subset label_definitions.json files are data dictionaries and coverage indexes. The human subset observes 156 field pairs and 46 scopes; the synthetic subset observes 37 pairs and 25 scopes. These are coverage counts, not alternative inference vocabularies. The human dictionary also retains the complete global inventory, whereas the synthetic dictionary lists its observed subset.

Speakers use recording-local first-occurrence identifiers. Scopes distinguish business authority and effect; verified materials do not imply formal approval or actual disbursement.

DCIST macro-averages complete-trajectory F1 across recordings. Recovering only the terminal value is insufficient. The code repository provides scoring rules, task prompts and evaluation tools. The paper's model-assisted semantic/additional-content adjudication, cross-model review and human resolution of disagreements are distinct from deterministic scoring; a new single-model judge run does not reproduce the frozen paper judgments.

Evidence and limitations

Human timestamps are coarse aligned evidence windows, not word-level boundaries or complete diarization annotations. Synthetic timestamps derive from assembled utterance intervals and may enclose gaps between supporting utterances. Neither should be treated as independently verified verbatim transcription.

The dialogues depict fictional financial scenarios. Human voices are not anonymized, and opening introductions or brief off-script speech may remain. Anonymous speaker IDs do not prevent recognition. Project documentation states that participant research/public-release permissions are retained by the project team; signed forms and identity mappings are not distributed. Dataset licensing does not replace privacy or personality rights or constitute separate participant authorization for impersonation or voice cloning.

Human-performed scripts do not represent the full distribution of spontaneous meetings. The recordings have also been used for protocol checks and are not an untouched holdout. Family dependence and construction templates limit generalization claims. Benchmark performance does not establish suitability for real financial or lending decisions.

License

Both subsets and the accompanying documentation are distributed under CC BY-NC 4.0:

• human/: CC BY-NC 4.0.
• synthetic/: CC BY-NC 4.0.
• Evaluation code and prompts are distributed separately on GitHub under CC BY-NC 4.0.

Third-party reference-voice resources remain subject to their source licenses. This dataset license does not grant additional rights in those resources.

Citation

Please cite DecisionTraceBench: Tracking Decision-Critical Information in Multi-Party Spoken Conversations and identify the subset and version used. The paper link will be added when available.