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.

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.

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:

Fraud outcome — what the attack is: fraud_type_key (the 7 types above).
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.

from datasets import load_dataset
ds = load_dataset("<repo>", "hi")          # Hindi only
ds = load_dataset("<repo>")                # all languages

Roadmap / how this repo grows

This layout is designed so each phase is added without rewriting earlier data:

Phase 1 — text (this release). data/<lang>/train.jsonl.
Phase 2 — audio. TTS mp3s land in audio/<lang>/<id>.mp3; rows already carry the matching audio_file path. An <lang> audio config will be added.
Phase 3 — preference pairs. Chosen/rejected pairs for ALM alignment go under preferences/<lang>/ 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.

Downloads last month: 72

Total file size:

16.1 GB

Request access to FraudAlign-MCS