This repository publishes the research framework needed to inspect prepared
interview records, build prompts for experimental responses, evaluate model
outputs, and compare paired method scores statistically. Its actual starting
point is a local prepared_input.json file that already conforms to the
input schema.
The installable Python distribution is limited to modules under
src/interviewsim.
This release is for research purposes only and should not be used to compete with OpenAI, Anthropic, or Google.
The release supports research analysis and comparison of synthetically generated interview-response datasets and their evaluation results. The bundled examples are entirely fictional. The paper's original interview data are not distributed with this repository.
InterviewSim starts with structured interview records in the canonical
prepared_input.json format. The following activities are outside the release:
- data discovery;
- data collection;
- scraping;
- downloading;
- transcription; and
- source verification.
No component is provided to locate people, enumerate identities, retrieve media, perform transcript acquisition, follow source URLs, or decide whether a record is authentic.
The public framework is limited to:
- loading and validating the documented local record format;
- holding out the latest 20% of interview records by
sequence, rounded up to a whole interview, and using the remaining earlier records for training; - building no-context, short-profile, long-profile, chronological, random, retrieval, and hybrid grounding prompts;
- validating caller-supplied text-model and embedding interfaces;
- building and parsing strict prompts for content similarity, factual consistency, personality similarity, and MCQ evaluation;
- aggregating metric outputs and scoring MCQs; and
- performing paired sign-flip permutation tests.
External model services, credentials, access configuration, and model artifacts are not part of the package.
This repository releases no:
- real interview data or excerpts;
- real identity list or mapping between IDs and people;
- source-data links, platform identifiers, or retrieval manifests;
- raw audio, video, or transcripts;
- model weights, checkpoints, or training states;
- fine-tuning adapters;
- credentials or internal infrastructure details; or
- deployable personality agents.
Generated responses and score files are experiment artifacts. They are not verified biographical records, identity claims, or deployment-ready systems.
The synthetic examples were authored
entirely from scratch for this repository from ten generic fictional roles,
with five short interviews per role. Every personality_id, profile,
interview, question, and response is invented; none was sampled from or
transformed from research transcripts. Each personality's
prepared_input.json is actual loader input.
Across the examples, 50 interviews contain 150 Q&As: 120 train and 30 test. The split files and all reference artifacts were generated deterministically for inspection of derived formats. They are not a sample of the paper dataset, do not preserve its population or content distribution, and cannot reproduce the paper's reported numbers. A successful audit demonstrates only that the invented files, references, and counts are internally consistent.
The paper's empirical results depend on unreleased inputs and model configurations in addition to the public framework. This repository supports method inspection and experiments on separately supplied conforming inputs; it does not claim paper-number reproduction.
InterviewSim is licensed under the Creative Commons Attribution-NonCommercial 4.0 International License (CC-BY-NC 4.0).