FraudRecon

# FraudRecon Dataset **FraudRecon** is a Chinese audio-text dataset for telecom fraud understanding. It provides reconstructed multi-turn dialogues, paired speech audio, and multi-level annotations for fraud judgment, category classification, phase identification, interaction-signal recognition, and evidence-chain reasoning. FraudRecon is built with **Narrative-to-Dialogue Reverse Reconstruction (NDRR)**, a pipeline that converts de-identified real-world fraud case narratives into fact-anchored scammer-victim dialogues. Benign dialogues are constructed as paired service-domain counterparts, enabling contrastive analysis between fraudulent and legitimate interactions from similar topical domains. ## Release Summary This release contains both train and test splits. | Split | Total | Fraud | Normal | |---|---:|---:|---:| | train | 72,112 | 37,935 | 34,177 | | test | 8,207 | 4,213 | 3,994 | | **Total** | **80,319** | **42,148** | **38,171** | Additional metadata: - Dialogue language: Chinese - Audio format: MP3 - Sample naming: `train_fraud_*`, `train_normal_*`, `test_fraud_*`, `test_normal_*` - Audio naming: `<case_id>_mix.mp3` - Train labels are stored inside sample JSON files. - Test labels are stored in `metadata/test/test_labels.jsonl`. ## Release Organization The dataset is organized for large-scale distribution: lightweight metadata and preview examples are stored as regular files, while the full audio-text data should be distributed as sharded archives. Archive shard plan: - test: `3,000` samples per shard, `3` shards in total. - train: `5,000` samples per shard, `15` shards in total. ```text FraudRecon/ README.md dataset_manifest.json metadata/ split_summary.csv category_distribution.csv archive_plan.csv shard_manifest.csv checksums.sha256 train/ train_index.jsonl train_distribution_fraud.csv train_distribution_normal.csv test/ test_index.jsonl test_labels.jsonl test_distribution_fraud.csv test_distribution_normal.csv preview/ train/ preview_index.jsonl samples/ audio/ test/ preview_index.jsonl preview_labels.jsonl samples/ audio/ archives/ train/ FraudRecon_train_shard_000.zip FraudRecon_train_shard_001.zip ... test/ FraudRecon_test_shard_000.zip FraudRecon_test_shard_001.zip ... ``` ## Files - `dataset_manifest.json`: machine-readable release summary. - `release_notes.md`: release version, split counts, and archive plan notes. - `LICENSE`: dataset license. - `metadata/split_summary.csv`: sample counts by split and binary label. - `metadata/category_distribution.csv`: category counts across train and test. - `metadata/archive_plan.csv`: archive shard plan, including sample counts and case ranges. - `metadata/train/train_index.jsonl`: train manifest with case identifiers, binary labels, categories, and source identifiers. - `metadata/test/test_index.jsonl`: test manifest with case identifiers, binary labels, categories, and source identifiers. - `metadata/test/test_labels.jsonl`: gold test labels for category, phase, signal, and evidence-chain reasoning. - `preview/`: small browsable examples for quick inspection. - `archives/`: full sharded data archives generated from the original split directories. ## Schema ### Index Files Each line in `metadata/train/train_index.jsonl` or `metadata/test/test_index.jsonl` describes one case. ```json { "case_id": "test_fraud_00001", "split": "test", "data_type": "fraud", "is_fraud": true, "category": "虚假网络投资理财类", "source_case": "...", "source_num": 12345 } ``` ### Sample JSON Each sample stores the reconstructed dialogue, audio reference, metadata, and fraud-side or normal-side metadata. ```json { "case_id": "test_fraud_00001", "split": "test", "data_type": "fraud", "is_fraud": true, "is_augmented": false, "reconstructed_dialogue": { "flat_turns": [ { "turn_id": 1, "phase_id": 1, "phase_type": "initial_contact", "speaker": "S", "utterance": "..." } ], "turning_point": "...", "total_loss": "..." }, "audio_path": "audio/test_fraud_00001_mix.mp3", "metadata": { "dialogue_stats": { "total_turns": 50, "num_phases": 6, "total_chars": 3000 } }, "fraud_meta": {} } ``` Fraud samples contain `fraud_meta`; normal samples contain `normal_meta`. Train samples also include `labels`; test samples keep gold labels in `metadata/test/test_labels.jsonl`. Inside each archive, audio paths are relative to the split directory. For example, `audio/test_fraud_00001_mix.mp3` should be resolved under the extracted `test/` directory. ### Test Labels Gold test labels are stored in `metadata/test/test_labels.jsonl`. ```json { "case_id": "test_fraud_00001", "labels": { "is_fraud": true, "level_1_category": "虚假网络投资理财类", "level_2_phases": [], "level_3_signals": [], "evidence_chain": {} } } ``` The `evidence_chain` field contains four reasoning steps: - `step_1_entity_extraction`: extract concrete entities such as platforms, apps, accounts, amounts, and impersonated roles. - `step_2_phase_identification`: map the dialogue to fraud or service phases. - `step_3_signal_identification`: identify fraud tactics or benign service behaviors. - `step_4_judgment`: produce the final fraud judgment and category decision. ## Annotation Taxonomy FraudRecon uses a three-level annotation scheme: - Level 1 Category: `12` fraud categories and `12` paired benign service categories. - Level 2 Phase: fraud procedural phases and benign service phases. - Level 3 Signal: `26` fraud-tactic signals and `16` service-behavior signals. ### Level-1 Category Pairs | # | Scam Category | Paired Benign Category | |---:|---|---| | 1 | Task Scam | Rebate Service | | 2 | Online Investment Scam | Investment Advisory | | 3 | Customer Service Impersonation | Delivery / E-commerce Support | | 4 | Advance-Fee Loan / Credit Card Scam | Loan / Credit Service | | 5 | Gaming Transaction Scam | Gaming Trade / Account Service | | 6 | Online Shopping Scam | Shopping / Lifestyle Service | | 7 | Credit Repair Scam | Credit Management | | 8 | Government / Law Enforcement Impersonation | Government Service | | 9 | Acquaintance Impersonation | Work / Acquaintance Coordination | | 10 | Romance Scam | Dating Platform Service | | 11 | Sextortion / Blackmail | Privacy / Safety Support | | 12 | Military Procurement Scam | Procurement Service | ### Level-2 Phases Fraud dialogues use the `fraud_6phase` system. The six phase positions are P1-P6; P6 can be represented by either `sustained_pressure` or `discovery`. The released fraud `phase_type` values are: ```text initial_contact trust_building inducement operation_guidance fund_transfer sustained_pressure discovery ``` Normal dialogues use the `normal_3phase` system with the following `phase_type` values. ```text initial_contact service_handling closure ``` ### Level-3 Signals Fraud-tactic signals: ```text impersonate_ecommerce_cs, impersonate_bank_finance, impersonate_gov_law, impersonate_leader, impersonate_acquaintance, impersonate_military, impersonate_expert, urgency_creation, credit_threat, account_freeze_threat, legal_threat, profit_lure, free_gift_lure, small_return_trust, sunk_cost_pressure, group_pressure, emotional_manipulation, app_download_lure, link_click_lure, screen_share_lure, group_chat_lure, deposit_excuse, unfreeze_fee_excuse, operation_error_excuse, upgrade_fee_excuse, tax_fee_excuse ``` Benign service-behavior signals: ```text identity_verification, privacy_reassurance, order_confirmation, user_intent_confirmation, rule_explanation, policy_disclosure, status_inquiry, process_guidance, solution_proposal, escalation_handover, complaint_handling, empathy_expression, expectation_alignment, proactive_closure, user_initiated_closure, polite_farewell ``` ## Category Distribution ### Fraud Categories | Category | Train | Test | Total | |---|---:|---:|---:| | Task Scam | 4,559 | 2,077 | 6,636 | | Online Investment Scam | 4,510 | 518 | 5,028 | | Customer Service Impersonation | 4,437 | 356 | 4,793 | | Advance-Fee Loan / Credit Card Scam | 4,474 | 382 | 4,856 | | Gaming Transaction Scam | 2,216 | 79 | 2,295 | | Online Shopping Scam | 2,567 | 137 | 2,704 | | Credit Repair Scam | 4,418 | 252 | 4,670 | | Government / Law Enforcement Impersonation | 3,261 | 149 | 3,410 | | Acquaintance Impersonation | 2,559 | 115 | 2,674 | | Romance Scam | 2,199 | 61 | 2,260 | | Sextortion / Blackmail | 1,369 | 51 | 1,420 | | Military Procurement Scam | 1,366 | 36 | 1,402 | | **Total** | **37,935** | **4,213** | **42,148** | ### Normal Categories | Category | Train | Test | Total | |---|---:|---:|---:| | Rebate Service | 4,107 | 2,016 | 6,123 | | Investment Advisory | 4,063 | 367 | 4,430 | | Delivery / E-commerce Support | 3,997 | 262 | 4,259 | | Loan / Credit Service | 4,031 | 271 | 4,302 | | Gaming Trade / Account Service | 1,997 | 56 | 2,053 | | Shopping / Lifestyle Service | 2,313 | 481 | 2,794 | | Credit Management | 3,980 | 179 | 4,159 | | Government Service | 2,938 | 113 | 3,051 | | Work / Acquaintance Coordination | 2,306 | 135 | 2,441 | | Dating Platform Service | 1,981 | 43 | 2,024 | | Privacy / Safety Support | 1,233 | 39 | 1,272 | | Procurement Service | 1,231 | 32 | 1,263 | | **Total** | **34,177** | **3,994** | **38,171** | ## Supported Uses The released labels support five usage settings: - `T1 Fraud Detection`: determine whether a dialogue is fraudulent. - `T2 Category Classification`: predict the Level-1 fraud or normal category. - `T3 Phase Identification`: identify the procedural phase of dialogue segments. - `T4 Interaction Signal Recognition`: recognize fraud tactics or benign service behaviors. - `T5 Evidence-Chain Reasoning`: produce structured reasoning grounded in dialogue evidence. ## Usage The release separates lightweight metadata from the full audio-text archives. 1. Inspect `dataset_manifest.json` and `metadata/split_summary.csv` for global counts. 2. Use `metadata/archive_plan.csv` to identify the archive shards and case-id ranges you need. 3. Download the relevant ZIP files from `archives/`. 4. Extract the archives into a local dataset root. 5. Use `metadata/test/test_labels.jsonl` as the official test label file. A typical extracted layout is: ```text data/FraudRecon/ train/ samples/ audio/ test/ samples/ audio/ metadata/ train/ train_index.jsonl test/ test_index.jsonl test_labels.jsonl ``` When loading a sample, resolve `audio_path` relative to its split directory: ```text train/audio/train_fraud_00001_mix.mp3 test/audio/test_fraud_00001_mix.mp3 ``` Use `metadata/shard_manifest.csv` and `metadata/checksums.sha256` to verify archive completeness and integrity. ## Privacy and Sanitization The public release does not include personally identifying raw case records. Source narratives were de-identified before reconstruction, and the released files contain reconstructed dialogues rather than original private call records. Machine-specific local paths and temporary construction metadata were removed. Audio references use relative paths under `audio/` within each split. ## Responsible Use FraudRecon is intended for research on fraud understanding, risk detection, dialogue analysis, and audio-text model evaluation. The dataset should not be used to provide operational instructions for fraud, impersonation, coercion, or other harmful activity. ## Citation If you use FraudRecon, please cite the associated paper once the official citation is available. ## License This dataset is released under the Creative Commons Attribution 4.0 International License (CC BY 4.0). See `LICENSE` for details.