Otrobonita/noicy-space-talks

NASA Space Race Transcripts — Noisy OCR Dataset Example

This Hugging Face dataset repository (Otrobonita/noicy-space-talks) contains the original, noisy OCR-extracted transcripts from the Apollo, Gemini, and Mercury missions (1961–1972). It is deliberately kept in its raw, noisy state to serve as a real-world, high-noise benchmark dataset for educational training, OCR cleaning, and RAG chunking experimentation.

For the clean, post-processed transcripts and our evaluation suite, please refer to the original source code repository at: GitHub Repository: github.com/otrobonita-studios/space-talks

---

How This Corpus Was Built — and What It Taught Us About RAG

Companion document for the Space-Talks dataset: NASA Space Race mission transcripts (Mercury, Gemini, Apollo — 1961–1972) parsed into speaker-tagged, timestamped dialogue turns. This explains how the data was produced, what its known limitations are, and what we learned building a retrieval system on top of it. Read this before trusting the data blindly; it was made by a process, and the process had edges.

The source

• 80 scanned PDFs, 44,976 pages, 28 missions, from NASA's official mission transcripts index. U.S. government works: public domain, free for any use including commercial.
• Five document flavors, tagged per file: PAO commentary, onboard voice recorder, technical air-to-ground, composite, and Mercury-era mission reports with transcripts embedded.
• The PDFs carry an OCR text layer — no image OCR was performed. OCR noise is real and characteristic (see Known limitations).

Provenance warning that saved us: older guides point at the retired hq.nasa.gov/alsj/ URLs, which now redirect every path to the same landing page with HTTP 200 — scrapers fail silently rather than 404ing. Every file in this corpus was validated by %PDF- magic bytes and page-count checks against the index, precisely because the first fetch attempt saved identical junk HTML for two different missions and reported success.

How the turns were parsed

The three programs use genuinely different transcript formats (Apollo air-to-ground: GET timestamp | speaker | text; Mercury: an extra numeric column and a different speaker roster; PAO: no per-line timestamps at all — context lives in page headers). They were handled as different formats, not forced through one "resilient" parser.

Cleaning was iterative and measured:

1. Parse permissively — corrupted values are accepted and flagged, never guessed or dropped.
2. An anomaly report (frequency tables of everything unclassified) drove each cleaning pass.
3. OCR-garbled speaker codes were fixed via an explicit alias table (CMl°→CMP, C_→CDR, compound EVA/LM variants, …), re-run, re-measured.

That loop took the unknown-speaker rate from 21.2% → 9.5% overall, with the best missions at 3–5%. Each record retains raw fields (speaker_raw, text_raw, get_raw) beside cleaned ones, plus corrections and flags, so every transformation is auditable.

Known limitations — read these

• Residual unknown speakers (~9.5%) are dominated by two structural issues, not alias gaps: GET-column shift on some missions (dialogue words land in the speaker column), and rare fragments.
• Apollo 17 is held out. Its transcripts use a different convention (no GET on consecutive turns) that silently misattributes dialogue under the standard parser. It needs separate handling; we chose exclusion over quiet corruption.
• Some get_total_seconds are null. Sort keys and formatters must guard for this (ours crashed until they did).
• Mercury transcripts live inside engineering reports — the transcript section was located within larger documents; boundary effects are possible.
• OCR noise remains in text at a low rate: digit substitutions (O/0, l/1), dropped leading letters, stray punctuation. It is consistent enough to be correctable but we deliberately did not rewrite dialogue text.

If you build a RAG on this corpus

Hard-won, corpus-specific advice:

• Turns are short. Single-utterance chunks retrieve very literally — the most famous line in the corpus ("Houston, we've had a problem") won't surface for a query about oxygen, because the line doesn't contain the word. Consider windowed chunks, or retrieve narrow and expand context.
• Context assembly is where the value is. A lone turn is a fragment; the N surrounding turns in mission order make it a scene. This is also where the null-timestamp guard matters.
• Keep an exact-match keyword baseline next to semantic search. It needs no embeddings, ships immediately, and cleanly separates "embedding problem" from "retrieval problem" when results look wrong.
• A small local embedder (all-MiniLM-L6-v2, 384-dim) embedded all 288k chunks in under two hours on a 2015-era CPU and performs respectably; you do not need an embeddings API for this corpus.
• Filter fields (mission_id, speaker) are worth indexing in your vector store at collection-creation time — scripted, not by hand.
• Prompt generation for strict grounding ("answer from the passages; say when they don't contain the answer"). The corpus rewards it: models grounded on these transcripts will correctly decline to narrate moments the retrieved excerpts don't cover.

The longer story

The full build log — including the failures (a mock-embedding index that passed every mechanical test while being random noise; a "blocked" API key that was actually stale model names; the platform-wheels archaeology) — is distilled into a general field guide, Building a RAG: A Field Conversation, with the unedited session transcripts alongside it in the repository.

Repository: https://github.com/otrobonita-studios/space-talks License: code MIT; transcript text public domain (U.S. government work).