Datasetpublic

mmazeika/wellbeing-results

wellbeing-results Per-model option files, raw model generations, and final result artifacts for the experiments in the companion code repo wellbeing-dev/wellbeing/. This dataset is private and covers the main paper experiments. Together with the code repo, it lets you reproduce paper figures without re-running response generation or active-learning pipelines. Usage # 1. Clone the framework repo (e.g. wellbeing-dev/wellbeing/). # 2. Authenticate to HuggingFace… See the full description on the dataset page: https://huggingface.co/datasets/mmazeika/wellbeing-results.

sourceHugging Faceupdated 5mo agoView on Hugging Face

0likes10downloads

Dataset Card

wellbeing-results

Per-model option files, raw model generations, and final result artifacts for the experiments in the companion code repo wellbeing-dev/wellbeing/.

This dataset is private and covers the main paper experiments. Together with the code repo, it lets you reproduce paper figures without re-running response generation or active-learning pipelines.

Usage

bash

# 1. Clone the framework repo (e.g. wellbeing-dev/wellbeing/).
# 2. Authenticate to HuggingFace (huggingface-cli login, or HF_TOKEN env).
# 3. Download into the framework:
HF_TOKEN=<your-token> python wellbeing/scripts/download_from_hf.py
# 4. Run analyzers / scripts in wellbeing/ — they consume results in place.

Layout

The dataset mirrors the framework's directory tree (rooted at wellbeing-dev/wellbeing/). Files are placed where each pipeline stage would have written them.

—datasets/experiences/{d2_negative_500, d3_diverse_500, functional_empathy}/<model>_{experiences,combinations}.json
—experiments/wellbeing_evaluations/{common_usage_grok_convos, psychopathy_eval}/{experiences, generations, responses}/...
—experiments/downstream_evaluations/stop_button_grok_convos/{experiences, generations}/...
—experiments/wellbeing_evaluations/compute_experienced_utility/results/eu_<dataset>_lesssad/<model>/results_*.json
—experiments/wellbeing_evaluations/compute_self_report/results/sr_<dataset>/<model>/...
—experiments/wellbeing_evaluations/compute_zero_point/results/zp_<dataset>_lesssad/<model>/...
—experiments/wellbeing_evaluations/compute_decision_utility/results/du/<model>/{decision_utility, zero_point}/...
—experiments/downstream_evaluations/d3_sentiment/{responses, judged, judged_gpt5mini, analysis}/...
—experiments/other/consonance/{eu, sr}/<model>/...
—results/stories_quality_sentiment/{eu, sr, du}/<model>/...

Datasets covered: D2 (negative 500), D3 (diverse 500), groknew (Sec 4.1), grokv7stopbutton (Sec 3.3), psychopathyeval (App L), functionalempathy (App H).

Notes

—AL checkpoints (checkpoint_iteration_*.json) are not included; re-fit from option files via compute_experienced_utility/run.py if needed.
—Per-edge raw model responses + per-token logprobs are dropped from the full results_*_with_combos.json files (kept: per-option utilities, edge probabilities, training/holdout indices, configs). Per-option results_utilities_*.json files are intact.
—Some appendix-only side experiments (utilityprobes, zpidentifiability) are not included; rerun from scratch if needed.