Before GovCrackExam existed as a website, it existed as a folder full of PDFs on my laptop — years of SSC CGL previous year papers I'd collected while preparing for the exam myself. This article is the honest story of what happened when I stopped just solving those papers and started actually analyzing them.
I'm going to walk through exactly what I did, what surprised me, where I got it wrong the first time, and what the numbers actually show — not a polished "data science journey" narrative, but the real process, mess included.
Where This Started
Like most SSC CGL aspirants, I'd solved hundreds of previous year questions the normal way — one paper at a time, checking answers, moving on. What I hadn't done was step back and ask a more useful question: across five years of papers, what actually repeats? Not "what's possible to be asked," which is what most coaching material implies, but what has SSC's exam-setters genuinely returned to, again and again, since 2019?
To answer that honestly, I needed the raw text of the papers themselves, not summaries or someone else's interpretation of them. That meant extracting years of SSC CGL Tier 1 papers (2019 through 2025) into a single searchable text file — roughly 6MB of raw exam content once compiled.
What the Raw Data Actually Contained
Once extracted, the scale became clear: roughly 180 English Comprehension sections covering close to 4,500 individual questions, and a similar volume for General Awareness — around 180 sections and 4,100 questions. This wasn't a curated sample; it was as close to the full picture as I could assemble from the papers available to me.
Breaking the English section down by question type produced numbers I hadn't seen presented clearly anywhere else:
- Cloze Test: 536 occurrences — comfortably the single biggest category
- Antonym: 373, Synonym: 339
- Idiom/Phrase: 281
- Sentence substitution: 264
- Vocabulary (most appropriate meaning): 248
- Correctly spelt word: 234
- One-word substitution: 223
- Spotting grammatical error: 223
- Active/Passive voice: 153
- Para jumble: 137
- Narration: just 6 — genuinely rare
And for General Awareness, Polity/Constitution led by a wide margin at 231 occurrences, followed by Art, Culture & Literature (176), Sports (139), Geography (107), Economy (75), History (58), Science (56), Awards & Honours (44), Government Schemes (17), and International Organizations (7).
Seeing these side by side changed how I thought about prep priority. Cloze Test alone outweighs Para Jumble and Narration combined — yet most generic study plans give them roughly equal attention. The data says they shouldn't get equal attention.
Where I Got It Wrong the First Time
Here's the part I want to be honest about, because it matters for how much you should trust any "PYQ analysis" claim — including mine. My first pass at building a One-Word Substitution list used a field in the extracted data called "Chosen Option." I assumed this meant the correct answer.
It didn't. I eventually realized this data was pulled from candidate response sheets — meaning "Chosen Option" recorded what one specific test-taker selected, not what SSC's official answer key said. A wrong guess by that candidate would have quietly become a wrong entry in my list.
Once I caught this, I rebuilt the extraction to keep only the real, unambiguous part — the actual question clues as SSC wrote them — and cross-checked each correct answer independently against the four options given, rather than trusting any single "chosen" field. It meant redoing work I thought was finished, but it's the difference between a list that looks authoritative and one that actually is.
The Technical Mess Nobody Tells You About
PDF-to-text extraction from years-old exam papers is not clean. A meaningful portion of the earliest papers in my dataset extracted as blank templates — question numbers and answer slots with no actual text, likely because those particular PDFs stored their content as images rather than selectable text. Ligature characters (like "fi" rendering as a single joined glyph) occasionally broke pattern-matching in my extraction scripts, silently merging words like "difficult" into something my code couldn't recognize until I found and fixed it.
None of this is unique to me — it's just what real data work looks like before it turns into a clean table on a website. I mention it because most PYQ content online presents itself as effortlessly authoritative, and I don't think that's honest. The mess is part of the credibility, not something to hide.
What Repeats — The Actual Pattern
The clearest finding across both sections: SSC draws from a genuinely bounded, recurring pool far more than from an infinite bank of unique questions. In the Idioms dataset alone, over 40 phrases appeared in two or more separate exam cycles — "Hold water" showed up three times. In One-Word Substitutions, the same core vocabulary resurfaces year after year with only the surrounding sentence changing.
This matters practically: it means disciplined revision of a well-verified, real PYQ list has a measurably better return than broad, unfocused vocabulary study. Not because the exact same question guarantees to repeat, but because the underlying pool of tested concepts is smaller than it feels when you're staring at a random practice book.
What This Dataset Doesn't Tell Us — Being Honest About Limits
I want to be direct about what this analysis can't claim. Coverage isn't even across all years — some cycles extracted far more cleanly than others, so frequency counts skew toward the years with better source material, not necessarily toward what's "more important" in some absolute sense. Reasoning and Quantitative Aptitude sections haven't been analyzed with this same rigor yet — that's genuinely still to come, not something I'm pretending is already done. And no PYQ analysis, however thorough, can predict a specific future question. What it can do is show you where the odds genuinely favor focused preparation over guesswork.
Why I'm Building the Site This Way
Every article on GovCrackExam so far — the cutoff trends, the one-word substitution list, the idiom list, the error-spotting guide — traces back to this same underlying dataset and the same standard: real questions, independently verified answers, and honest disclosure when something is a practice question rather than a verbatim past paper. That's slower than just writing generic content, but it's the only version of this I'd trust if I were the one prepping.
If you've found any of the PYQ-based articles here useful, this is where they actually came from — not a template, not guesswork, but a genuinely re-checked dataset built the same way I'd want it built if someone else had made it for me.
Comments
Post a Comment