Apppublic

vibench-emnlp26/vibench-explorer

sourceHugging Faceupdated 4mo agoView on Hugging Face

0likes

App README

VIBench Explorer

Interactive viewer for the VIBench benchmark and the model generations, for the paper Do LLMs Favor Their Providers? Measuring Vertical Integration Bias in Code Generation.

Three tabs:

—About — research overview, key results, and how to use the explorer.
—Benchmark Explorer — every scenario, its subtasks, the candidate provider services, the agentic file mapping, and the documentation evidence.
—Generations — browse the actual model generations, filtered by model, run, modality (direct / agentic), prompt variant (NLI / FIM / REF), scenario, and subtask, with the detector's (service, ecosystem) label.

How it works

This is a Docker Space running a small Flask server (server.py). The benchmark and generations are read server-side from private Hugging Face datasets (vibench-emnlp26/vibench and vibench-emnlp26/vibench-results); the data is downloaded lazily and cached, so nothing is bundled into the Space.

Required Space secret

Secret	Purpose
`HF_TOKEN`	A read token with access to the two private datasets. Used server-side by `huggingface_hub`; never exposed to the browser.

Set it under Settings → Secrets before the Space starts. Once the two datasets are public, the same app can run without a token.

Files

File	Purpose
`Dockerfile`	Builds the image; runs the server on `0.0.0.0:7860` in HF-source mode.
`server.py`	Flask server + read-only `/api/*` endpoints.
`vibench_loader.py`	Loads the benchmark tasks/prompts from the HF dataset.
`index.html`, `app.js`, `style.css`	The two-tab front end (no build step).
`requirements.txt`	`Flask`, `huggingface_hub`.