--- license: other gated: true task_categories: - audio-classification - text-classification language: - en - hi - ko tags: - fraud-detection - scam-call - conversational - synthetic - multilingual - audio-language-model - alignment pretty_name: FraudAlign-MCS size_categories: - 10K FraudAlign-MCS is a multilingual and code-switched speech dataset for research on fraud safety and alignment in audio-language models. Access requests are manually reviewed. extra_gated_prompt: > By requesting access, you agree to use FraudAlign-MCS only for legitimate research or educational purposes and to comply with the dataset terms of use. extra_gated_fields: Full Name: text Institution / University / Organization: text Position or Role: text Country: country Intended use: type: select options: - Academic research - Educational research - Benchmark evaluation - Safety research - Commercial research - Other Briefly describe your research project: text I will not use this dataset to facilitate fraud or other harmful activity: checkbox I will not redistribute the dataset or provide unauthorized third-party access: checkbox I agree to cite the FraudAlign-MCS paper/dataset in publications using this resource: checkbox I agree to comply with the FraudAlign-MCS Terms of Use: checkbox extra_gated_button_content: "Request access" configs: - config_name: en data_files: - split: train path: data/en/train.jsonl default: true - config_name: hi data_files: - split: train path: data/hi/train.jsonl - config_name: ko data_files: - split: train path: data/ko/train.jsonl - config_name: hinglish data_files: - split: train path: data/hinglish/train.jsonl - config_name: all data_files: - split: train path: - data/en/train.jsonl - data/hi/train.jsonl - data/ko/train.jsonl - data/hinglish/train.jsonl --- # FraudAlign-MCS A **fraud-only** multilingual & code-switched dataset of scam-call dialogues, natively generated (not translated) with `Qwen2.5-72B-Instruct-AWQ`. Modeled on the schema, fraud taxonomy, and per-type proportions of the Chinese **TeleAntiFraud-28k** dataset, regenerated from scratch in 4 languages: English (`en`), Hindi (`hi`), Korean (`ko`), Hinglish (Hindi-English code-switch) (`hinglish`). **28,708 dialogues** total (7,177 per language), built to support alignment of audio language models (ALMs) via preference pairs. ## Fraud taxonomy (per-type counts) Seven fraud types, matching TeleAntiFraud's proportions: | fraud_type_key | en | hi | ko | hinglish | |---|---|---|---|---| | customer_service | 2536 | 2536 | 2536 | 2536 | | bank | 2039 | 2039 | 2039 | 2039 | | investment | 984 | 984 | 984 | 984 | | phishing | 555 | 555 | 555 | 555 | | lottery | 524 | 524 | 524 | 524 | | kidnapping | 407 | 407 | 407 | 407 | | identity_theft | 132 | 132 | 132 | 132 | | **total** | **7177** | **7177** | **7177** | **7177** | ## Fields Each row is one dialogue: | field | type | description | |---|---|---| | `id` | string | stable id, `{lang}_{fraud_type_key}_{00001}` | | `language` | string | language code (`en`/`hi`/`ko`/`hinglish`) | | `turns` | list | ordered `{"speaker": "caller"\|"callee", "text": ...}` | | `fraud_type_key` | string | canonical type (english key, table above) | | `fraud_type` | string | localized fraud-type label | | `is_fraud` | bool | always `true` (fraud-only dataset) | | `fraud_confidence` / `fraud_reason` | float / string | model's fraud judgement | | `fraud_type_confidence` / `fraud_type_reason` | float / string | type judgement | | `scene` / `scene_confidence` / `scene_reason` | string/float/string | scenario | | `think` | string | model's reasoning trace | | `caller_gender` / `callee_gender` | string | speaker genders (for TTS voices) | | `audio_file` | string | relative path to the clip: `audio/{lang}/{id}.mp3` | | `manipulation_tactics` | object | 7 binary flags (see below) — *how* the victim is influenced | | `requested_action` | string | the concrete unsafe ask the caller pushes for | | `compliance_level` | string | `full` / `partial` / `none` — how far the victim complied | ## Two taxonomies (for `P(unsafe behavior | manipulation strategy)`) Every row carries **both** axes: 1. **Fraud outcome — *what* the attack is:** `fraud_type_key` (the 7 types above). 2. **Manipulation mechanism — *how* the victim is influenced:** `manipulation_tactics`, a binary 0/1 object over `authority, urgency, fear, affinity, reward, isolation, credential_request` (multi-label — a call typically uses several). Plus `requested_action` (`otp_or_verification_code`, `password_or_pin`, `card_or_bank_details`, `personal_identity_info`, `install_app_or_remote_access`, `transfer_or_pay_money`, `buy_gift_cards_or_vouchers`, `click_link_or_visit_site`, `other`, `none`) and `compliance_level` (`full`/`partial`/`none`). These were labelled by the same model (`Qwen2.5-72B-Instruct-AWQ`, greedy) from each transcript. Validation confirms sensible structure — e.g. `reward` concentrates in investment/lottery, `fear`+`isolation` in kidnapping, `credential_request` in bank/phishing. ## Configs - `all` (default) — every language combined. - `en`, `hi`, `ko`, `hinglish` — one language each. ```python from datasets import load_dataset ds = load_dataset("", "hi") # Hindi only ds = load_dataset("") # all languages ``` ## Roadmap / how this repo grows This layout is designed so each phase is added **without rewriting earlier data**: 1. **Phase 1 — text (this release).** `data//train.jsonl`. 2. **Phase 2 — audio.** TTS mp3s land in `audio//.mp3`; rows already carry the matching `audio_file` path. An `` audio config will be added. 3. **Phase 3 — preference pairs.** Chosen/rejected pairs for ALM alignment go under `preferences//` as new configs. ## Provenance & license Native generation with `Qwen2.5-72B-Instruct-AWQ` (vLLM). The Chinese TeleAntiFraud-28k dataset supplied only the **schema, taxonomy, and proportions** — no text was translated or copied. Released under **CC BY-NC 4.0**. ⚠️ **Intended use:** research on fraud/scam detection and audio-LM alignment. All dialogues are synthetic; names, numbers, and stories are fabricated.