saksornr/debt-collection

Debt Collection Conversations and Scripts

A multilingual collection of debt-collection call excerpts, published scripts, a user-supplied transcript, and simulated exchanges. Each row contains a description, an ordered list of messages, and references that trace the example back to its source. The collection includes 1,344 examples and 7,373 message or summary turns.

Read the record_type before using an example. Most rows are synthetic. Documented calls, translations, published scripts, user-supplied transcripts, and generated simulations serve different purposes and should not be treated as interchangeable evidence of real-world behavior.

What's included

• 83 documented-call excerpts: source-reported calls from research, journalism, court records, and an anonymized publisher transcript. Some are paraphrased and seven reconstruct selected exchanges from court descriptions. A row is an excerpt, not necessarily a complete call or a unique call.
• 6 translations: five Thai translations of English examples and one English translation of a Turkish example. Each points to its parent through translation_of_id.
• 154 published scripts: spoken examples, including two-sided scripts, collector-only prompts, and voicemail. These are illustrative scripts, not evidence of actual calls. Some have only one turn.
• 1 user-supplied transcript: a 22-turn Thai debt settlement negotiation transcribed from a YouTube clip. The setting and whether the call was staged are unverified. Names and the callback number were redacted. การเจรจาต่อรองไกล่เกลี่ยทวงหนี้
• 1,000 synthetic conversations: full simulations provided by RecoveryBench; they contain three to nine turns each.
• 100 synthetic benchmark exchanges: borrower-message and agent-response pairs from RecoveryBench-100. These are two-turn scenarios, not full calls.

The English, Thai, Turkish, romanized Hindi, romanized Bengali, and code-mixed Hinglish content retains the source wording where available. Hinglish and romanized Hindi share the hi-Latn language value; the original RecoveryBench language label is retained in source_metadata for finer filtering. Where real-call material was translated or paraphrased, the transformation field describes the curation.

What the data can—and cannot—show

The dataset is useful for exploring conversation formats, building prototypes, and testing processing pipelines. It is not a representative sample of debt collection. The 1,100 synthetic rows account for about 82% of the release; the simulator uses repeated personas, templates, and generated scenarios. Its volume should not be interpreted as evidence about the prevalence of behaviors among real borrowers or collectors.

The 83 documented-call rows include English renderings of Chinese-language research transcripts, paraphrases, and court-summary reconstructions. For example, seven rows summarize exchanges described in court records rather than reproducing full recordings. Source references and transformation distinguish these from direct transcript excerpts. Calls can involve third parties or advisers, and some describe disputed or potentially abusive conduct. Source-supplied compliance labels and RecoveryBench expected-compliance values are annotations, not legal conclusions.

The complete dataset is supplied as a single train split for convenience; it is not a held-out benchmark. Translations are dependent on their parent records, multiple excerpts can come from one source, and synthetic examples may share templates. Create evaluation splits by source and parent call, and consider separating synthetic examples entirely when evaluating real-call performance. The 100 original RecoveryBench scenarios are identified by source_split=benchmark, but are included in this dataset's single train split and do not constitute a separate held-out set here.

Record structure

Field	Meaning
conversation_id	Stable row identifier.
record_type	documented_call, translation, published_script, user_supplied_transcript, synthetic_conversation, or synthetic_benchmark.
description	Brief description of the example and its provenance.
conversation	Ordered array of {role, content} objects. assistant denotes the collector or collection system; user denotes the call recipient or borrower. A recipient can also be a third party or adviser.
references	Array of source titles, URLs, local curation paths, and source locators. Local paths point to the original workspace and are not distributed here.
language, source_language	Language of the stored wording and of its source; language codes can include script information.
translation_of_id	Parent row for a translation, otherwise empty.
channel, coverage	Medium and extent of the source material. A complete simulation is not a real call.
message_count	Number of stored turns, including any summary turns.
transformation, redacted, compliance_label	Known transformations, redaction information, and supplied labels. Empty values mean unknown or not applicable.
source_split, source_metadata	Source collection and JSON-encoded source metadata for RecoveryBench records. Includes the original scenario ID, source language label, persona/context, and expected annotations where supplied.

Load and filter

from datasets import load_dataset

rows = load_dataset("saksornr/debt-collection", split="train")
real_excerpts = rows.filter(lambda row: row["record_type"] == "documented_call")
simulations = rows.filter(lambda row: row["record_type"] == "synthetic_conversation")
print(real_excerpts[0]["conversation"])
print(simulations[0]["references"])

The Parquet file and JSONL file retain nested arrays. The UTF-8 CSV file stores conversation and references as JSON within cells. No audio is included.

Sources and access

Real-call material includes excerpts discussed by Choi and colleagues, Harrington, and Muratlar and colleagues, together with other sources cited in individual records. The source papers may describe larger underlying corpora; their full private or unreleased calls are not included. See source coverage for the 142 indexed archive entries, additional sources for further research and court material, and the requested-source catalog for access limitations on other corpora.

RecoveryBench provides the simulations and benchmark exchanges in its repository under an MIT license; the license text is included here. Its separately documented single-message borrower-intent CSV dataset has a research/educational-use designation in its dataset card. Those single-message CSV records are not included: their redistribution terms are unspecified and they are not multi-turn conversations.

There is no established license covering this entire combined dataset. The other sources retain their own rights and attribution requirements; assess the source-specific terms before redistribution, training, or commercial use. Public access does not by itself establish unrestricted reuse rights.

Validation and privacy

All three exports were reopened and checked for matching nested conversations and references, valid IDs and translation links, and preservation of Thai, Turkish, and other multilingual text. The 1,000 simulated conversations have distinct complete message sequences; this does not rule out shared templates or similar individual turns. Counts and checks are in validation.json.

Real-call excerpts have the redaction and transformation status recorded during curation where known. Synthetic names and amounts belong to generated scenarios. These checks are not a comprehensive privacy review, and empty redaction fields should not be read as a guarantee that a source has been anonymized.