LLM agents in public-good and common-pool-resource experiments
This folder is an export of the experimental data used in the accompanying paper. It is designed to be copied to a Hugging Face dataset repository.
Contents
| File | Description |
|---|---|
data/ors_2022_main_decisions.csv.gz |
Sanitized decision-level records from the corrected Oechssler--Reischmann--Sofianos (ORS 2022) VCM, SCCM, and CCM sweep, including baseline and fixed-player conditions. |
data/ors_2022_main_group_periods.csv.gz |
One row per group and period for the same sweep; recommended for most analyses. |
data/ors_2022_main_manifest.csv |
Anonymized condition/run manifest and row counts. |
data/ors_2022_model_metadata.csv |
Model metadata used to describe the tested panel. |
data/prompt_ablation_decisions.csv.gz |
Decision-level records for the worked-example and coalition prompt ablations. |
data/prompt_ablation_manifest.csv |
Anonymized manifest for the prompt-ablation records. |
data/fischer_2004_exploratory_decisions.csv.gz |
Sanitized decision-level records for the exploratory Fischer (2004) CPR runs. These are single-seed exploratory data, not the queued confirmatory sweep. |
data/fischer_2004_manifest.csv |
Anonymized Fischer condition manifest. |
data/analysis_tables/ |
Small, publication-facing summary tables (CSV/JSON) copied from verified analyses. |
metadata.json |
Dataset scope, design, and disclosure notes. |
DATA_DICTIONARY.md |
Column definitions and coding conventions. |
checksums.sha256 |
SHA-256 checksums for every exported file. |
The release contains 2,469,600 ORS decision rows, 493,920 ORS group-period
rows, 322,200 prompt-ablation decision rows, and 864 Fischer decision rows.
The main ORS manifest has 486 selected result files; some files are partial
diagnostic conditions and should be filtered using status, actual_rows, and
expected_rows when reconstructing a particular analysis sample.
Important disclosure boundary
The export intentionally excludes chain-of-thought/reasoning text, raw model responses, prompts, API response identifiers, billing and reservation fields, machine fingerprints, timestamps, run-directory names, and all local or remote filesystem paths. The model outputs are represented by the parsed experimental choices and derived game outcomes. This avoids publishing private reasoning or infrastructure metadata while retaining the information needed to reproduce the paper's analyses.
Player IDs are experimental positions, not human identities. Fixed-player placement is retained because it is part of the treatment definition; no human participant data are included in this export.
Loading
import pandas as pd
decisions = pd.read_csv("data/ors_2022_main_decisions.csv.gz")
groups = pd.read_csv("data/ors_2022_main_group_periods.csv.gz")
The CSV files use UTF-8, comma separators, and explicit missing values as empty
fields. IDs are stable within this release but are synthetic and carry no source
path information. All percentages are stored as proportions in [0, 1] unless
the column name ends in _pct, in which case they are percentage points in
[0, 100].
The two experiment-level exports use gzip-compressed CSV with a header row.
Scientific-use notes
The ORS main panel is a purposive panel of model variants, not a probability sample. Human reference summaries in the paper are derived from the cited human datasets and are not reproduced here. The Fischer archive data have one inference seed and four chains per condition and should be labelled exploratory. The analysis tables are convenience products; the decision and group-period files are the authoritative exported observations.
Release date: 2026-08-28.