Podcast question-answering datasets: long-context answers with traceable evidence Scope podcast QA datasets with episode editions, grounded answers, evidence intervals, unanswerable cases, benchmark context and reviewable delivery examples. Podcast question-answering data connects a question to a supported answer and evidence in an identified episode edition. It differs from a transcript archive, a retrieval relevance label or an episode summary. A useful collection preserves source identity, speaker context, answerability, reference provenance and the distinction between transcript evidence and audio evidence. This guide supplies an illustrative JSONL example and a practical design for long-context podcast evaluation. Core unit Question + supported answer + identified evidence Source rule Pin episode edition and transcript revision Unknowns Unanswerable is a valid reference state Artifact Authored text-only JSONL, not recorded audio A transcript collection is a source, not automatically a QA benchmark An archive can contain many episodes and still lack trustworthy questions, answers and evidence spans. Specify whether a model must answer from audio, a transcript, metadata, retrieved excerpts or a combination. These input conditions define different tasks. An answer that is easy to locate in corrected text may be difficult to recover from noisy audio, while a question about laughter or emphasis may not be answerable from text at all. The [SPoRC: Mapping the Podcast Ecosystem](https://aclanthology.org/2025.acl-long.1222/) and [SPoRC official dataset card](https://huggingface.co/datasets/blitt/SPoRC) provide a research reference for structured podcast analysis. Their existence does not establish that every episode has human-verified QA targets or that a source can be offered commercially. Use the actual record fields, annotation provenance and selected artifact terms to decide what additional preparation would be required. Use historical benchmarks for task design, not assumed availability The [TREC Podcasts Track and current access notice](https://trecpodcasts.github.io/) organized segment retrieval and episode summarization around podcast sources. Those are useful neighboring tasks, but a relevant segment label is not a complete answer reference. The [The Spotify Podcast Dataset research paper](https://arxiv.org/abs/2004.04270) also provides background on podcast audio and automatic transcripts; automatic transcript availability does not establish answer correctness. The current TREC page explicitly says its Spotify dataset is no longer maintained and no longer accepts access requests as of December 2023. Treat it as a historical research reference, not a promise that a buyer can obtain that archive today. Likewise, SPoRC’s current dataset card states research-use-only. A public corpus citation should never be presented as evidence that Fiund can distribute its recordings or license every downstream use. Distinguish local, episode-level and cross-source questions A local question can be answered from a short interval. An episode-level question may require combining statements separated by a long discussion. A cross-episode question requires an explicitly defined archive scope and source grouping. Decide which jobs the collection is meant to measure before generating questions because each has different evidence and leakage risks. Also distinguish direct extraction, temporal ordering, speaker attribution, multi-step reasoning and questions whose premise is unsupported. A broad topic question may reward a model’s prior knowledge instead of its use of the episode. Prefer questions that require the supplied source when that is the evaluation objective, and record whether external knowledge is allowed. Question family Reference needs Common mistake Local fact Supporting phrase or audio interval Accepting a plausible answer without source support Long-context relation All required evidence intervals and reasoning rule Citing one nearby sentence that does not establish the answer Speaker attribution Stable speaker/source reference Confusing the host’s question with the guest’s assertion Unanswerable Why the supplied scope lacks the answer Filling the gap from general knowledge Tie evidence to a specific episode edition Published episodes can differ from original sessions through edits, inserted advertisements, corrected uploads or alternate audio/video editions. A timestamp is meaningful only against the edition and time base used. Preserve show, episode, source-session and edition identifiers where available, and keep derivative relationships explicit. For audio evidence, retain the actual interval and enough surrounding context to interpret the statement. For transcript evidence, retain transcript revision, character/word span and its alignment to the identified audio edition if that alignment exists. A character span in one transcript is not automatically a measured interval in a recording. If the mapping is unavailable, say so rather than invent timestamps. An authored example with one supported and one unanswerable question The downloadable example uses a short text written specifically for this guide: an archive has forty episodes, twelve recorded remotely and twenty-eight in a studio. A question about studio-recorded episodes is supported by a defined character span. A question about overlapping speech is not supported by that source, even though a model might guess that conversations often overlap. [Download the illustrative question/evidence JSONL](/downloads/podcast-question-evidence-example.jsonl). The records contain authored explanatory text, no recorded voices, no actual podcast episode and no measured audio timing. The audio_available field is false and audio intervals are null. The file demonstrates schema and answerability decisions; it is not ready-made acoustic evaluation material. An actual long-context sample would need the permitted source, its edition, the input conditions, independently reviewed references and aligned evidence where requested. The example’s role is to make those missing requirements visible before a team treats a large text archive as a complete audio-QA dataset. Construct questions without leaking the answer Keep the question-generation process separate from final reference review. A generator may invent a premise, copy an answer into the question or construct a trivial query around a distinctive phrase. Record which source material and metadata it saw. Human review should check whether the question tests the intended capability and whether the answer is established by the supplied input. For a long episode, include evidence dispersed across the conversation, corrections and changes of position. A speaker may revise a number later; the reference needs an explicit rule for the final, earlier or context-specific answer. Avoid turning an unresolved disagreement between speakers into a single authoritative fact unless the question clearly concerns what a named speaker said. Review references and adjudicate competing answers Give reviewers the same defined input scope and ask them to locate evidence independently before comparing labels. Retain alternative acceptable phrasing where appropriate. Separate a difference in wording from a factual disagreement, an unsupported inference or a missing qualifier. A short correct answer can be better supported than a fluent paragraph that adds claims the episode never makes. For ambiguous cases, record the adjudication rationale and any change to the question. If the answer cannot be established under the intended input condition, mark it unanswerable, revise it or exclude it according to the protocol. Do not hide uncertainty by demanding agreement with the first generated reference. The review record should show which target is authored, generated, checked or still unresolved. Split by sources that can carry repeated information Recurring hosts, repeated intros, syndicated segments, advertisements and alternate editions can create overlap across apparently different episode files. Preserve these relationships when constructing development and evaluation sets. A model that recognizes a repeated segment has not necessarily demonstrated long-context reasoning over an unfamiliar discussion. Choose held-out shows, speakers, topics or collection periods according to the claim you intend to evaluate. Topic separation alone may not remove repeated speakers or source derivatives. Record unavailable lineage as a limitation and do not claim a leakage-free benchmark solely because filenames or transcript strings differ. Define metrics around answer support and the input condition Useful dimensions can include factual correctness, evidence sufficiency, speaker attribution, qualifier preservation and appropriate abstention. Report the question types and denominators separately. A high score on local extraction should not be presented as equivalent to multi-interval reasoning or audio-grounded understanding. Choose the grading method and tolerances before comparing systems. Automated judges can assist a review but should not silently become unquestioned ground truth. Retain a sample of difficult and borderline cases, the reference evidence and the decision rationale so the measurement can be inspected. This guide reports no model benchmark or measured Fiund corpus performance. Price the source and reference work as different layers A quote should distinguish archive access/rights, transcript preparation, edition reconciliation, audio alignment, question construction, independent review, adjudication and packaging. One thousand questions do not have a comparable cost when one set requires a short local fact and another requires multiple distant evidence intervals and specialist interpretation. For an illustrative pilot, twelve episodes with five proposed questions each yield sixty candidate questions before review. If some are unanswerable, duplicated or poorly scoped, accepted output is smaller. Count candidate and accepted questions separately, and do not infer a commercial hourly rate or dataset-wide acceptance fraction from this arithmetic. No representative market-average price is asserted. A reviewable next step with Fiund Send the intended input—audio, transcript, retrieved clips or a combination—along with question types, episode/source requirements, reference format, intended use and sample acceptance. Ask which source material and documentation can be reviewed, and which preparation layers remain to be scoped. Avoid sending private participant records or recordings in an initial inquiry. Fiund can discuss a podcast QA sourcing requirement at hello@fiund.com. Availability, source terms, preparation feasibility, accepted output and commercial scope are confirmed for the selected project. For archive identity and transcript layers, continue to the related retrieval-metadata and time-aligned transcript guides rather than treating a QA schema as a substitute for the underlying source review. Is podcast retrieval data the same as question-answering data? No. Retrieval relevance identifies useful source material; QA additionally needs a supported answer, defined input scope and sufficient evidence. Can the JSONL example be used to evaluate an audio model? Not as acoustic evidence. It contains authored text and no recordings or measured audio intervals. Real audio evaluation requires an appropriately reviewed audio source and references. Are the Spotify/TREC recordings currently available through this page? No. The TREC source states that its underlying Spotify archive is no longer taking access requests. It is cited here for historical task design, not offered as Fiund inventory. Do publicly available transcripts establish commercial training rights? Public access and the intended licensed use are separate questions. Review the selected source and artifact terms; this guide does not establish a permission scope for a public corpus. Research method: Fiund research desk; public sources, explicit worked-example labels and project-specific availability review. Editorial contact: hello@fiund.com. Sources SPoRC: Mapping the Podcast Ecosystem: https://aclanthology.org/2025.acl-long.1222/ SPoRC official dataset card: https://huggingface.co/datasets/blitt/SPoRC TREC Podcasts Track and current access notice: https://trecpodcasts.github.io/ The Spotify Podcast Dataset research paper: https://arxiv.org/abs/2004.04270