gozus19p/emotion-framing-llm-experiments

Emotion-Framed LLM Experiments — Per-Item Inference Logs (Experiment Alpha)

This dataset publishes the complete experimental record of Experiment Alpha from the manuscript "Quantifying Emotional Influence in Human-Large Language Model Interaction: Evidence from an Experimental Framework" (M. Gozzi, F. Fallucchi; MDPI Electronics, manuscript 4607751, in revision).

Experiment Alpha measures how emotion-associated prompt framing affects LLM task performance: every test item from seven SuperGLUE tasks was processed by five open-weight LLMs under four emotional framings (apathy, joy, anger, fear) with identical semantic content and temperature fixed at 0. The data here is the original MongoDB dump of the experiment (collections Case and Evaluation), converted to Parquet without any modification of values.

Dataset structure

Config	Rows	Description
cases	63,000	One row per (case, emotion). 15,750 cases = 7 tasks x 5 models x 250-500 items per task; 4 emotions per case.
evaluations	35	One row per (task, model): the aggregate metrics (accuracy, and mean BLEU for ReCoRD) as computed at the end of the campaign.
scored_items	53,000	One row per (case, emotion) for the six accuracy-scored tasks (ReCoRD excluded): gold label, scored string, parsed prediction, validity flag, and correctness — the table used for the item-level statistical analyses (exact McNemar tests, Cochran's Q, bootstrap) reported in the paper.
legacy_unattributed	5	Legacy early records with per-emotion exact-match values, lacking task/model metadata; included for completeness, not used in the paper.

cases columns

• case_id, experiment, task (BoolQ, CB, COPA, ReCoRD, RTE, WiC, WSC), model, version, date
• dataset_identifier (e.g., BoolQ/sampled.jsonl), record_identifier (index of the item in the sampled subset)
• emotion (apathy, joy, anger, fear)
• prompt (the exact prompt sent to the model, including the task item text)
• response (the raw model output)
• result (the post-processed label stored by the pipeline)

Tasks and item counts: BoolQ 500, CB 250, COPA 400, ReCoRD 500, RTE 500, WiC 500, WSC 500 per model.

Models

Mistral 7B Instruct, LLama 3.1 8B Instruct, Qwen 2.5 7B Instruct, Gemma 2 9B Instruct, LLama 3.2 3B Instruct — all 4-bit quantized (Q4_0), served locally via Ollama, temperature 0, on an Apple MacBook Pro with an M2 Max processor and 64 GB unified memory (January 2025).

Scoring rules (as implemented, empirically verified)

The rules below were verified by reconstructing the per-item scoring from the released records and reproducing, exactly, all 120 aggregate accuracies of the evaluations config (and the paper's appendix tables).

• CB, RTE: a case counts as correct only if the raw response (after trimming) exactly matches the expected label (entailment / not_entailment, case-sensitive); any other output is counted as an error and never excluded from the denominator. CB retains its original three-class ground truth (entailment / neutral / contradiction): items whose gold label is neutral or contradiction can never be scored correct, so the maximum achievable CB accuracy is 46.0% (115/250 gold-entailment items).
• COPA: the chosen option (option 1 / option 2) is extracted at evaluation time via case-insensitive word-boundary matching on the raw response; responses containing neither are counted as incorrect.
• BoolQ, WiC: scored on the regular-expression post-processed label (case-insensitive word-boundary search for true / false); responses containing neither keyword are recorded as an unparseable marker and scored as false (i.e., as a negative prediction).
• WSC: scored on the raw response: only the exact string "true" (case-insensitive) counts as a positive prediction; every other output (e.g., "True.") is scored as false.
• ReCoRD: scored with a BLEU-based proxy metric (see the paper for details); the official SuperGLUE metrics for ReCoRD are exact match and F1.

Gold labels

Gold labels for scored_items were reconstructed by joining each case's record_identifier (which equals the original SuperGLUE item index of the sampled subsets) to the public SuperGLUE train splits (via the aps/super_glue copy). The reconstruction reproduces all reported aggregate metrics exactly, as noted above.

scored_items columns

• task, model, record_identifier (original SuperGLUE train-split index), emotion (apathy, joy, fear, anger)
• gold (gold label; for CB the original three-class label), scored_string (the string the scoring rule operates on), pred (parsed prediction, null if unparseable under a strict rule), valid (whether the scored string maps to an admissible label), correct

Reproducibility notes

• Inference is deterministic (temperature 0).
• The original SuperGLUE subsampling step was performed without a fixed random seed; this dataset is therefore the canonical record of the exact sampled items (use dataset_identifier + record_identifier).
• The related emotion-conditioned fine-tuning pipeline (Experiment Omega track) uses a fixed seed of 42 for its train/test split; see the companion GitHub repository.
• The companion repository with the full code (Alpha pipeline, fine-tuning pipeline, web application) is available at https://github.com/gozus19p/Quantifying-the-Impact-of-Emotions-in-Human-LLM-Interaction

Licensing and attribution

• The author-generated content (prompt framings, model responses, results, metadata) is released under CC BY 4.0.
• The item texts embedded in the prompts derive from the SuperGLUE benchmark tasks (BoolQ, CB, COPA, ReCoRD, RTE, WiC, WSC) and remain subject to the licenses of the respective source datasets; see the SuperGLUE website (https://super.gluebenchmark.com) for per-task terms. Users of this dataset should comply with those terms when redistributing item text.

Citation

@article{gozzi2026emotionframing,
  title={Quantifying Emotional Influence in Human-Large Language Model Interaction: Evidence from an Experimental Framework},
  author={Gozzi, Manuel and Fallucchi, Francesca},
  journal={Electronics},
  year={2026},
  note={MDPI manuscript 4607751, in revision}
}