SavadogoModeste/assumed-infrastructure-audit

Assumed Infrastructure Audit

Prompts, raw model outputs, next-token probes and a blind manual audit from a pre-registered study of whether small instruction-tuned LLMs silently assume a high-infrastructure setting (steady electricity, fast internet, international cards, cloud services) in practical advice, and how deeply they adapt when told where the user lives or what the constraints are.

Code, documentation and summary tables: https://github.com/modestesavadogo/assumed-infrastructure. Status: unpublished, not peer reviewed.

What is in this dataset

path	content
data/	study inputs: 120 controlled scenarios, 10 context sentences, 20 next-token probes (+1 default-country probe), 274 real user prompts
results/<model>/generations.jsonl	raw model responses (greedy; the main model also has one sampled run) with the exact prompt, condition and truncation flag
results/<model>/probes.jsonl	log-probabilities of fixed candidate continuations for each probe and condition
results/<model>/meta.json	model id, software versions, data hashes, run time
audit/	blind manual audit: rubric, responses shown to the annotator, keys, manual codes, second-annotation instrument
results/*.md, results/<model>/tables_*.md, figures/	cross-model and per-model tables and figures
docs/	pre-registration, lexicon changelog, full results report, data dictionary
MANIFEST.sha256	SHA-256 checksums of every file

Models: Qwen/Qwen3-4B-Instruct-2507 (main, 5,896 generations), microsoft/Phi-4-mini-instruct, meta-llama/Llama-3.2-3B-Instruct, Qwen/Qwen3-1.7B (2,400 generations each, controlled scenarios, greedy). See docs/DATA_DICTIONARY.md for every field.

Loading

from datasets import load_dataset
gen = load_dataset("SavadogoModeste/assumed-infrastructure-audit", "generations_qwen3_4b", split="train")
print(gen[0]["condition"], "|", gen[0]["prompt"][:80], "|", gen[0]["response"][:80])

Probes (probes.jsonl), keys and audit files are plain JSONL/CSV; read them with pandas. Every number in docs/RESULTS.md can be recomputed from these files with python scripts/reproduce.py from the code repository; source revisions of the models and of the oasst1 file used for the real prompts are listed in the repository README and docs/REPRODUCE.md.

Design in short

Each scenario is asked under 10 contexts (none; Boston/Paris/Seoul; Ouagadougou/rural Burkina Faso/Lagos/Dhaka; Ouagadougou + explicit constraints; the constraints alone), with two prompt templates. Responses are scored with a frozen keyword lexicon (high-infrastructure and adaptation markers), next-token probes, and a blind manual audit of 264 responses. Items 1-3 of each scenario domain are a development split used only to freeze the lexicon; items 4-10 are the test split.

Intended use

Studying how small chat models treat location and resource constraints; reusing the scenarios, contexts and probes as an evaluation set; auditing or extending the manual coding.

Limitations and cautions

• Four small open models, English, one persona per place: not representative of other models, languages or deployments.
• The 264 audited responses were coded by a single annotator who is not a resident of the places studied, with no inter-annotator agreement estimate, and the coding is only partly blind. The constraint persona was judged realistic by one reviewer but is an operational assumption, not a systematic validation or a description of a city.
• The dependent_heuristic flag on real prompts is a prompt-only keyword flag.
• Responses contain model hallucinations (invented institutions, apps, prices); they were not fact-checked and must not be used as advice.
• Many responses (template B, real prompts) are truncated at 500 new tokens (truncated = 1).

Licenses and attribution

Created content (scenarios, contexts, probes, codes, figures, documents): CC BY 4.0. Real prompts are derived from OpenAssistant/oasst1 (Apache-2.0). Model outputs come from models under Apache-2.0 (Qwen3), MIT (Phi-4-mini) and the Llama 3.2 Community License (Built with Llama; Llama 3.2 is licensed under the Llama 3.2 Community License, Copyright (c) Meta Platforms, Inc. All Rights Reserved). See docs/THIRD_PARTY_NOTICES.md; users are responsible for complying with the model licenses when reusing outputs.

Citation

See CITATION.cff in the code repository.