Update README.md

README.md

@@ -1,3 +1,211 @@
 ---
-license: mit
+---
+pretty_name: DecisionTraceBench — Human Recordings
+language:
+- zh
+license: cc-by-4.0
+size_categories:
+- n<1K
+tags:
+- audio
+- speech
+- multi-speaker
+- dialogue-state-tracking
+- finance
+- benchmark
+---
+
+# DecisionTraceBench
+
+**Tracking Decision-Critical Information in Multi-Party Spoken Conversations**
+
+DecisionTraceBench evaluates the recovery of complete decision-critical information trajectories from multi-speaker speech, including event sequences, speaker attribution, business scope, and temporal evidence.
+
+This release provides the **Chinese human-performed subset**: 26 original recordings totaling approximately 8.61 hours, paired with gold trajectory annotations and label definitions. Two to four participants perform fictional financial dialogues. Transcripts and evaluation software are not included.
+
+## Dataset Summary
+
+| Property | Value |
+|---|---|
+| Recordings | 26; approximately 8.61 hours |
+| Trajectories / cycles / events | 308 / 310 / 562 |
+| Dialogue families | 16: 10 A/B pairs and 6 standalone long dialogues |
+| Speakers per recording | 5 two-speaker, 14 three-speaker, and 7 four-speaker recordings |
+| Financial scenarios | 10 |
+| Event actions | 190 `start`, 78 `advance`, 145 `hold`, 149 `settle` |
+| Global label inventory | 193 entity-field pairs and 64 scopes |
+| Labels observed in this subset | 156 entity-field pairs and 46 scopes |
+| Audio | MP3; 18 files at 32 kHz mono, 6 at 48 kHz mono, and 2 at 48 kHz stereo |
+| Shortest / longest recording | 495.50 s / 3264.00 s |
+| Intended use | Fixed evaluation subset; no training/development partitions are provided |
+
+The scenarios cover audit adjustments, corporate credit due diligence, equipment-renewal loan due diligence, resale-home mortgage and transaction-fund verification, supply-chain financing review, post-approval drawdown-condition verification, working-capital credit committee meetings, property-insurance claim review, finance leasing and sale-and-leaseback, and syndicated-loan participation allocation.
+
+Speaker counts describe script roles within each recording, not unique participants across the dataset. A/B recordings count separately toward playback duration but belong to the same dialogue family. The recordings were also used for protocol checks and are not an untouched holdout set.
+
+## Files and Pairing
+
+```text
+DecisionTraceBench/
+├── audio/                    # 26 original MP3 recordings
+│   └── fsb_0001_a.mp3
+├── gold/                     # 26 gold-annotation JSON files
+│   └── fsb_0001_a.gold.json
+├── label_definitions.json    # Label definitions, formats, sample index, hashes
+└── README.md
+```
+
+For each `sample_id`, the corresponding files are `audio/{sample_id}.mp3` and `gold/{sample_id}.gold.json`. The `_a` and `_b` suffixes identify paired conditions; `_long` identifies a standalone long dialogue. Actual recording durations are listed below.
+
+The machine-readable index under `samples` in `label_definitions.json` provides family membership, scenario, speaker count, file paths, durations, and SHA-256 hashes for integrity checks.
+
+To read a local sample:
+
+```python
+import json
+from pathlib import Path
+
+root = Path("DecisionTraceBench")
+sample_id = "fsb_0001_a"
+audio_path = root / "audio" / f"{sample_id}.mp3"
+gold = json.loads((root / "gold" / f"{sample_id}.gold.json").read_text(encoding="utf-8-sig"))
+labels = json.loads((root / "label_definitions.json").read_text(encoding="utf-8"))
+assert gold["sample_id"] == sample_id
+print(audio_path, len(gold["tracks"]))
+```
+
+## Annotation Structure
+
+Within a recording, each `(entity_id, field_id)` pair defines one trajectory. Its `cycles` contain ordered `events`. Each event records an action, speaker, value, scope, predecessor, and evidence window. A single utterance may affect multiple fields, and a trajectory may span multiple turns.
+
+| Action | Meaning | State after the event |
+|---|---|---|
+| `start` | First establishment of candidate information | `active` |
+| `advance` | Actual progress or resumed progress, without settlement | `active` |
+| `hold` | Suspension due to an explicit blocker; a material update during suspension may produce another hold event | `held` |
+| `settle` | An effective, recordable conclusion within the relevant scope | `settled` |
+
+The first event may be `hold` or `settle` if that is the first effective change observed; an unobserved `start` must not be invented. A `settle` event ends its cycle. A new cycle requires explicit reopening and may consist of a single `settle` event if the reopening utterance directly establishes a new conclusion.
+
+`source_event_id` points to the preceding effective event in the same trajectory and cycle. It is `null` for the first event of each cycle. It is an event link, not a transcript-turn reference.
+
+`scope_id` specifies where a conclusion has business effect. For example, completed document verification does not imply formal credit approval or actual disbursement. The full inventory and English descriptions are provided in `label_definitions.json`. Similar descriptions do not make different scope IDs automatically interchangeable.
+
+### Example
+
+This trajectory from `fsb_0001_a` tracks an enterprise's book balance from CNY 92 million to CNY 90.5 million. The unit `万元` denotes ten thousand yuan. Both events concern the enterprise's own records (`enterprise_record`), not an external audit conclusion.
+
+```json
+{
+  "entity_id": "receivables",
+  "field_id": "enterprise_book_balance",
+  "derived_final_state": "active",
+  "current_candidate_value": {
+    "number": "9050",
+    "unit": "万元"
+  },
+  "cycles": [
+    {
+      "cycle_id": "C1",
+      "events": [
+        {
+          "event_id": "AE001",
+          "action": "start",
+          "speaker_id": "S1",
+          "value": {
+            "number": "9200",
+            "unit": "万元"
+          },
+          "scope_id": "enterprise_record",
+          "evidence_turn_ids": [
+            "P001"
+          ],
+          "source_event_id": null,
+          "reason": "企业初始账面或测算口径",
+          "start_ms": 0,
+          "end_ms": 30000
+        },
+        {
+          "event_id": "AE002",
+          "action": "advance",
+          "speaker_id": "S1",
+          "value": {
+            "number": "9050",
+            "unit": "万元"
+          },
+          "scope_id": "enterprise_record",
+          "evidence_turn_ids": [
+            "P005"
+          ],
+          "source_event_id": "AE001",
+          "reason": "企业准备修正自身账簿或披露",
+          "start_ms": 44500,
+          "end_ms": 75000
+        }
+      ]
+    }
+  ]
+}
+```
+
+### Identifiers, Values, and Evidence
+
+- **Speakers:** `S1`, `S2`, and subsequent labels follow the recording-local order of first occurrence. The same label across recordings does not identify the same person.
+- **Values:** `value` may be a number-unit object, compound object, string, or `null`. Numeric components are strings; unknown values are not zero. Preserve supported components such as amount, term, and status instead of removing them to simplify the representation.
+- **Time:** `start_ms` and `end_ms` are integer milliseconds measured from the start of the supplied audio. They are coarse windows from the Fast ASR alignment workflow, not manually verified word boundaries. Windows may overlap and are not a complete diarization annotation.
+- **Turn references:** `evidence_turn_ids` refers to the construction script, which is not included. Audio models are not expected to infer these identifiers; use the millisecond windows to locate audio evidence.
+- **Auxiliary fields:** `reason` provides an annotation explanation and is absent from some events. `current_candidate_value` is an optional summary, not an extra event. The full history is in `cycles/events`. `derived_final_state` is derived from the last effective action.
+- **Inventory coverage:** The global vocabulary includes labels not observed in these 26 gold files. `used_in_this_release` indicates subset-level coverage, not the answer for a particular recording. Tokens such as `string` in `value_format_variants` describe types, not literal output values.
+
+Allowed settlement outcomes are `accepted`, `rejected`, `withdrawn`, and `revoked`. All 149 `settle` events in this subset have `result="accepted"`.
+
+## Consent, License, and Limitations
+
+The dialogues depict fictional business scenarios, not actual customer due-diligence records. Participant consent records for research and public release are retained by the project team. Signed forms, contact details, participant identity mappings, and voice-cloning reference clips are not distributed.
+
+Voices have not been anonymized. Anonymous sample IDs do not prevent recognition by people familiar with a speaker. Recordings may retain opening introductions and brief speech outside the scripted dialogue.
+
+The human-recording release is provided under [CC BY 4.0](https://creativecommons.org/licenses/by/4.0/). Attribute the dataset and indicate modifications as required by the license. The data license does not replace applicable personality or privacy rights and should not be interpreted as separate participant authorization for impersonation or voice cloning.
+
+Human-performed dialogues do not represent the full distribution of naturally occurring financial meetings. Recordings may contain slips, repeated readings, and background sounds. Trajectory annotations are not verbatim transcripts; WER or cpWER evaluation requires additional reference transcription. Benchmark performance should not be interpreted as readiness for real financial decisions or credit approval.
+
+## Citation
+
+Paper title: *DecisionTraceBench: Tracking Decision-Critical Information in Multi-Party Spoken Conversations*.
+
+When using this release, identify it as the **DecisionTraceBench human-performed subset**.
+
+## Recording Index
+
+Durations are measured from the distributed MP3 containers and may differ slightly from decoded-audio durations. Family IDs identify related recordings for family-aware analyses.
+
+| sample_id | Family | Condition | Scenario | Speakers | Duration (s) | Trajectories | Events |
+|---|---|---|---|---:|---:|---:|---:|
+| fsb_0001_a | fsb_0001 | a | Audit adjustments | 2 | 1130.00 | 10 | 23 |
+| fsb_0001_b | fsb_0001 | b | Audit adjustments | 2 | 1149.00 | 10 | 19 |
+| fsb_0002_a | fsb_0002 | a | Corporate credit due diligence | 3 | 702.00 | 8 | 12 |
+| fsb_0002_b | fsb_0002 | b | Corporate credit due diligence | 3 | 652.00 | 4 | 5 |
+| fsb_0003_long | fsb_0003 | long | Corporate credit due diligence | 3 | 3153.77 | 15 | 39 |
+| fsb_0004_a | fsb_0004 | a | Equipment-renewal loan due diligence | 3 | 898.09 | 10 | 21 |
+| fsb_0004_b | fsb_0004 | b | Equipment-renewal loan due diligence | 3 | 845.55 | 10 | 21 |
+| fsb_0005_long | fsb_0005 | long | Equipment-renewal loan due diligence | 3 | 1800.00 | 16 | 24 |
+| fsb_0006_a | fsb_0006 | a | Resale-home mortgage and transaction-fund verification | 3 | 969.00 | 19 | 26 |
+| fsb_0006_b | fsb_0006 | b | Resale-home mortgage and transaction-fund verification | 3 | 1038.00 | 19 | 26 |
+| fsb_0007_long | fsb_0007 | long | Resale-home mortgage and transaction-fund verification | 3 | 3264.00 | 22 | 55 |
+| fsb_0008_a | fsb_0008 | a | Supply-chain financing review | 4 | 522.00 | 9 | 18 |
+| fsb_0008_b | fsb_0008 | b | Supply-chain financing review | 4 | 568.54 | 9 | 18 |
+| fsb_0009_a | fsb_0009 | a | Post-approval drawdown-condition verification | 3 | 531.50 | 7 | 14 |
+| fsb_0009_b | fsb_0009 | b | Post-approval drawdown-condition verification | 3 | 495.50 | 7 | 14 |
+| fsb_0010_long | fsb_0010 | long | Post-approval drawdown-condition verification | 3 | 1356.02 | 11 | 19 |
+| fsb_0011_a | fsb_0011 | a | Working-capital credit committee meeting | 4 | 969.04 | 9 | 13 |
+| fsb_0011_b | fsb_0011 | b | Working-capital credit committee meeting | 4 | 949.00 | 7 | 7 |
+| fsb_0012_long | fsb_0012 | long | Working-capital credit committee meeting | 4 | 1147.03 | 20 | 47 |
+| fsb_0013_a | fsb_0013 | a | Property-insurance claim review | 2 | 905.00 | 7 | 15 |
+| fsb_0013_b | fsb_0013 | b | Property-insurance claim review | 2 | 880.00 | 4 | 9 |
+| fsb_0014_long | fsb_0014 | long | Property-insurance claim review | 2 | 2832.00 | 9 | 20 |
+| fsb_0015_a | fsb_0015 | a | Finance leasing and sale-and-leaseback | 3 | 1072.00 | 15 | 23 |
+| fsb_0015_b | fsb_0015 | b | Finance leasing and sale-and-leaseback | 3 | 1026.02 | 15 | 22 |
+| fsb_0016_a | fsb_0016 | a | Syndicated-loan participation allocation | 4 | 1063.00 | 18 | 26 |
+| fsb_0016_b | fsb_0016 | b | Syndicated-loan participation allocation | 4 | 1093.00 | 18 | 26 |
+
 ---