unbluff/unbluffable-numeric-interview-questions

Unbluffable Numeric Interview Questions

600 model-generated numeric interview-practice questions from 12 parameterised families across quant, investment banking, consulting and software engineering. Each released answer was recomputed from its parameters with exact Python arithmetic by the included release validator.

This is a small reusable question corpus from Unbluffable, a spoken mock interviewer built for the Meta Global AI Developer Hackathon. It contains questions and answer-key metadata. It contains no candidate answers, user recordings, personal data, behavioural interview transcripts or private source question bank.

Content

Family	Track	Topic	Variants
qe_F001	quant	Conditional probability	50
qe_F004	quant	Discrete probability	50
qe_F005	quant	Discrete probability	50
co_F001	consulting	Profitability	50
co_F005	consulting	Breakeven	50
co_F008	consulting	Market sizing with stated assumptions	50
ib_F002	investment_banking	Accounting and three-statement linkages	50
ib_F007	investment_banking	Enterprise value and equity value bridge	50
ib_F009	investment_banking	Valuation multiples	50
sw_F005	software_engineering	Sliding window	50
sw_F020	software_engineering	Heaps	50
sw_F024	software_engineering	Prefix sums	50

All selected families have recorded zero-error solver checks in the source repository and zero declared numeric tolerance. We release every stored variant of these selected families, without changing question wording or answer values.

Provenance

• Public source commit: 5db4b185d28eeee3fa89082dee6aeeca00bb115e.
• Source question/variant generation: ChatGPT, except the selected qe_ quant families, written by Claude, according to the project's disclosures.
• Original solver author: Muse Spark Open (meta-models/muse-spark-open-1026) on Baseten.
• Original forge verification compared solver output with stored answers across all variants; the original metadata is preserved per row.
• Release validation: the included validate_dataset.py implements the 12 answer formulae and recomputes each answer exactly from its parameters. This validation was prepared with Codex.
• Source family files may say pending_builder_review. Those values are retained in source_review_status; this release does not claim completed human review.
• The original source solvers are not bundled. The included validator is a separate release implementation of the selected answer computations.
• MIT license and the original copyright notice are included from the public source repository.

Files and fields

data/questions.jsonl has one JSON object per question variant. IDs use family_id:variant_id. The public source and its file hashes are listed in manifest.json.

Field	Meaning
id, family_id, variant_id	Stable identifiers for this release
track, topic, difficulty	Source categories, not independently assessed difficulty
question, spoken_question	Original display and spoken question text
answer	Exact rational value represented as a string, e.g. 4/5 or 570
tolerance	Zero for every row in this release
parameters_json	JSON-encoded parameter object; rational strings retain exact values
wrong_methods_json	Retained source wrong-method values and names
solution_steps_json	Retained named intermediate values from the source
question_generator, solver_author_model	Disclosed generator and recorded solver model
source_review_status, source_verification_json	Original review status and machine-check metadata
release_answer_check	Exact recomputation performed for the release
source_file, source_commit	Pinned public source locations

Percentage answers, where present in a future extension, require an explicit convention. This version selects questions whose requested units are described in their text; preserve those units when interpreting numeric values.

Load locally

import json
from fractions import Fraction
from pathlib import Path

rows = [json.loads(line) for line in Path("data/questions.jsonl").read_text(encoding="utf-8").splitlines()]
example = rows[0]
print(example["question"], Fraction(example["answer"]))

After cloning or downloading the dataset repository, run:

python validate_dataset.py

This command requires only Python's standard library. It checks IDs, counts, the data checksum, exact numerical answers and wrong-answer collisions. validation_report.json records the release result.

Intended use and limitations

Use this corpus for interview-practice prototypes, exact-answer checker development and exploring parameterised questions. It is not an unseen benchmark: the questions and answers are public, synthetic and grouped into closely related families.

• There is a single questions corpus split, with no claim of a held-out test set. Any future training/evaluation split should separate whole families to reduce leakage from related variants.
• Exact recomputation establishes agreement with the included formulae. It does not establish unambiguous wording, complete domain realism, or validity of a candidate's reasoning.
• Source wrong-method names and intermediate steps are retained annotations. The release validator checks that wrong values differ from the answer; it does not reproduce every wrong-method algorithm or verify every step annotation.
• The consulting questions here are numerical. Behavioural practice and its reported user experience are outside this dataset.
• The corpus does not measure interview improvement, hiring outcomes or model superiority.
• The selection is deliberately small and Easy-labelled; it is not representative of all interviews or of the full Unbluffable bank.

Version

Version 1.0.0, prepared 4 November 2026. Family selection, source-file SHA-256 hashes and the JSONL checksum are in manifest.json.