metadata
license: mit
language:
- en
tags:
- generative-agents
- emotion
- multi-agent
- simulation
- reinforcement-learning
- cognitive-science
- occ-theory
- predictive-processing
task_categories:
- reinforcement-learning
- text-classification
size_categories:
- 100K<n<1M
pretty_name: 'Emotion Engine: Emergent Emotional Appraisal in Generative Agents'
Emotion Engine Dataset
204,520 agent-step records from a five-condition controlled experiment on emergent emotional appraisal in generative agents.
Paper: Emergent Emotional Appraisal in Generative Agents via Predictive World Modeling Author: Prem Babu Kanaparthi Code: github.com/premxai/emotion_engine
What This Dataset Is
Five emotion-engine architectures were compared in Ghost Town — a 12-agent survival simulation across 6 scenario variants and 8 random seeds.
The central finding: Condition D (trained only to predict future world events) independently rediscovers OCC cognitive appraisal signatures (fear, grief, suspicion) with zero emotion rules, labels, or reward shaping.
Each record is one agent at one timestep: the full observation vector, emotion state, latent representation, action taken, 12 future-event prediction probabilities, and outcome metadata.
Dataset Structure
Splits
| Split | Condition | Scenario | Seeds | Records |
|---|---|---|---|---|
baseline |
Baseline (no emotion) | all 6 | 0–7 | 40,907 |
condition_a |
Hand-coded OCC rules | all 6 | 0–7 | 40,921 |
condition_b |
Behavioral cloning | all 6 | 0–7 | 40,838 |
condition_c |
Emotion dynamics model | all 6 | 0–7 | 40,916 |
condition_d |
Predictive world modeling | all 6 | 0–7 | 40,938 |
Total: 204,520 records
Scenarios in Every Split
| Scenario | Primary OCC Signature | Description |
|---|---|---|
standard_night |
Fear | 1 ghost per night, standard threat |
high_ghost_pressure |
Fear (extreme) | 3 simultaneous ghosts per night |
storm_scarcity |
Stress + Grief | Food supply reduced 70% by storm |
ally_death |
Grief | June Carter dies at step 1 (scripted) |
betrayal_refusal |
Suspicion | Rival pairs initialized with negative ties |
crowded_shelter |
Fear + Suspicion | Shelter capacity halved (6 slots / 12 agents) |
Fields
Each record contains the following fields:
Identity
| Field | Type | Description |
|---|---|---|
run_id |
string | Unique run identifier, e.g. condition_d_standard_night_seed0_12-agent |
step |
int | Timestep within run (0–71, 3 simulated days × 24 steps/day) |
agent_id |
string | Agent name (one of 12 fixed agents) |
condition |
string | baseline_0, condition_a, condition_b, condition_c, condition_d |
seed |
int | Random seed (0–7) |
scenario |
string | One of the 6 scenario variants |
Observation (14 dimensions)
| Field | Type | Description |
|---|---|---|
obs_visible_ghosts |
int | Number of ghosts within sensor range |
obs_visible_deaths |
int | Number of ally deaths witnessed this step |
obs_nearby_allies |
int | Allies within 5 tiles |
obs_trusted_allies |
int | Allies with positive social tie nearby |
obs_supplies_seen |
int | Supply units visible |
obs_in_shelter |
bool | Agent is inside a safe building |
obs_nearest_refuge_distance |
float | Tiles to closest safe building |
obs_steps_since_ghost_seen |
int | Steps since last ghost sighting (capped at 20) |
obs_steps_since_ally_died |
int | Steps since last witnessed death (capped at 20) |
obs_steps_since_betrayal |
int | Steps since last betrayal received (capped at 20) |
obs_ally_deaths_witnessed |
int | Cumulative ally deaths witnessed |
obs_betrayals_received |
int | Cumulative betrayals received |
obs_average_trust |
float | Mean social tie value across all agents [−1, 1] |
obs_graph_tension |
float | Mean negative tie magnitude |
Emotion / Affect (Condition D only — zeros for others)
| Field | Type | Description |
|---|---|---|
affect_fear |
float | Fear activation [0, 1] |
affect_grief |
float | Grief activation [0, 1] |
affect_trust |
float | Trust activation [0, 1] |
affect_stress |
float | Stress activation [0, 1] |
affect_relief |
float | Relief activation [0, 1] |
affect_suspicion |
float | Suspicion activation [0, 1] |
Latent Representation (8 dimensions)
| Field | Type | Description |
|---|---|---|
latent_0 … latent_7 |
float | Internal learned representation (8-dim) |
Future-Event Predictions (Condition D only — zeros for others)
| Field | Type | Description |
|---|---|---|
pred_ghost_nearby_t3 |
float | P(ghost visible in 3 steps) |
pred_my_death_t5 |
float | P(this agent dies in 5 steps) |
pred_health_drop_t3 |
float | P(health decreases in 3 steps) |
pred_nearby_death_t5 |
float | P(an ally dies in 5 steps) |
pred_help_success_t5 |
float | P(help action succeeds in 5 steps) |
pred_refusal_received_t5 |
float | P(refusal received in 5 steps) |
pred_tie_increase_t5 |
float | P(social tie increases in 5 steps) |
pred_shelter_achieved_t2 |
float | P(agent in shelter in 2 steps) |
pred_storm_onset_t3 |
float | P(storm starts in 3 steps) |
pred_scarcity_t5 |
float | P(supply scarcity in 5 steps) |
pred_graph_tension_t3 |
float | P(social tension increases in 3 steps) |
pred_valence_t5 |
float | P(positive valence in 5 steps) |
Action & Outcome
| Field | Type | Description |
|---|---|---|
action |
string | Action taken: hide, gather_supplies, seek_safe_house, seek_hospital, refuse_help, warn, patrol |
goal |
string | Agent goal state at this step |
reward_survival |
float | Survival reward component |
reward_shelter |
float | Shelter reward component |
reward_health |
float | Health reward component |
reward_social |
float | Social reward component |
total_reward |
float | Sum of reward components |
alive |
bool | Agent survived this step |
health |
float | Health percentage (0–100) |
sheltered |
bool | Agent is in a safe building |
time_of_day |
string | day, dusk, night, dawn |
Loading the Dataset
With the datasets library (recommended)
from datasets import load_dataset
# Load a specific condition
ds = load_dataset("PremC1F/emotion-engine", "condition_d")
# Load all conditions
ds = load_dataset("PremC1F/emotion-engine", "all")
# Filter to a specific scenario
night_only = ds["train"].filter(lambda x: x["scenario"] == "standard_night")
# Filter to ghost-present steps
threat_steps = ds["train"].filter(lambda x: x["obs_visible_ghosts"] > 0)
Reproduce the Fear→Shelter result
from datasets import load_dataset
import numpy as np
from scipy.stats import fisher_exact
ds = load_dataset("PremC1F/emotion-engine", "condition_d", split="train")
df = ds.to_pandas()
ghost_present = df[df["obs_visible_ghosts"] > 0]
ghost_absent = df[df["obs_visible_ghosts"] == 0]
shelter_given_ghost = (ghost_present["action"] == "seek_safe_house").mean()
shelter_given_no_ghost = (ghost_absent["action"] == "seek_safe_house").mean()
# Contingency table
a = (ghost_present["action"] == "seek_safe_house").sum()
b = (ghost_present["action"] != "seek_safe_house").sum()
c = (ghost_absent["action"] == "seek_safe_house").sum()
d = (ghost_absent["action"] != "seek_safe_house").sum()
_, p = fisher_exact([[a, b], [c, d]])
print(f"Ghost present → shelter: {shelter_given_ghost:.1%}")
print(f"Ghost absent → shelter: {shelter_given_no_ghost:.1%}")
print(f"Fisher p = {p:.2e}")
# Expected: Ghost present → shelter: 99.0%, p = 3.3e-113
Reproduce the Suspicion Decay result
from datasets import load_dataset
ds = load_dataset("PremC1F/emotion-engine", "condition_d", split="train")
df = ds.to_pandas()
refusals = df[df["action"] == "refuse_help"].copy()
def bucket(steps):
if steps <= 5: return "recent"
if steps <= 10: return "fading"
return "forgotten"
refusals["bucket"] = refusals["obs_steps_since_betrayal"].apply(bucket)
all_social = df.copy()
all_social["bucket"] = all_social["obs_steps_since_betrayal"].apply(bucket)
for b in ["recent", "fading", "forgotten"]:
denom = (all_social["bucket"] == b).sum()
numer = ((refusals["bucket"] == b)).sum()
print(f"{b}: {numer/denom:.1%} refusal rate")
# Expected: recent 64.6%, fading 4.5%, forgotten 35.2%
Citation
@article{kanaparthi2026emotion,
author = {Kanaparthi, Prem Babu},
title = {Emergent Emotional Appraisal in Generative Agents
via Predictive World Modeling},
year = {2026},
}
License
MIT License.