Datasets:

brandburner
/

doctorwho-s12-narrative-kg

license: cc-by-sa-4.0
task_categories:
  - graph-ml
  - text-generation
tags:
  - narrative
  - knowledge-graph
  - screenplay
  - fabula
  - neo4j
  - graph-gravity
size_categories:
  - 1K<n<10K

Doctor Who - Narrative Knowledge Graph

A rich narrative knowledge graph extracted from Doctor Who screenplays using the Fabula pipeline. Contains characters, locations, objects, organizations, events, themes, and conflict arcs with full participation semantics and Graph Gravity importance tiers.

Dataset Overview

Metric	Value
Source database	`doctorwho.s12`
Type	Season database
Episodes	20
Total nodes	3,160
Total edges	11,042
Schema version	1.2.0
Exported	2026-08-05

Entity Breakdown

Type	Count
Act	50
Agent	117
ConflictArc	60
Episode	20
Event	596
Location	164
Object	497
Organization	51
PlotBeat	1,114
SceneBoundary	397
Theme	88
Writer	6

Graph Gravity Tiers

Tier	Count	Description
anchor	11	Main characters / key locations
planet	204	Recurring entities
asteroid	614	Minor / one-off entities

Relationship Types

AFFILIATED_WITH, BELONGS_TO_EPISODE, CALLBACK, CAUSAL, CHARACTER_CONTINUITY, CONTAINS_ACT, CONTAINS_BEAT, CONTAINS_SCENE, CREDITED_ON, EMOTIONAL_ECHO, ESCALATION, EXEMPLIFIES_THEME, FORESHADOWING, INVOLVED_IN_ARC, INVOLVED_WITH, IN_EVENT, NARRATIVELY_FOLLOWS, OCCURS_IN, PARTICIPATED_AS, PART_OF ... and 7 more

Related Datasets

This is a single-season dataset containing entities and events as extracted from Season 12 screenplays.

Megagraph (all seasons unified): brandburner/doctorwho-mega-narrative-kg

Note: The megagraph is not a simple union of season datasets. Cross-season entities are reconciled through a Global Entity Registry (GER), receiving new canonical UUIDs and distilled descriptions. Graph Gravity tiers are recalculated across all episodes. Use individual season datasets for single-season analysis; use the megagraph for cross-season analysis.

Files

File	Description
`nodes.parquet`	All graph nodes with properties
`edges.parquet`	All relationships with properties
`positions.parquet`	3D layout coordinates for visualization
`meta.json`	Dataset metadata and entity counts

Schema

Nodes (`nodes.parquet`)

Column	Type	Description
`node_id`	string	Unique node identifier (UUID)
`primary_label`	string	Node type (Agent, Location, Event, etc.)
`name`	string	Display name
`description`	string	Foundational description
`tier`	string (nullable)	Graph Gravity tier: anchor / planet / asteroid
`episode_count`	int (nullable)	Number of distinct episodes entity appears in
`first_episode_seq`	int (nullable)	First appearance episode
`last_episode_seq`	int (nullable)	Last appearance episode
`properties_json`	string	Full node properties as JSON

Edges (`edges.parquet`)

Column	Type	Description
`source_node_id`	string	Source node UUID
`target_node_id`	string	Target node UUID
`relationship_type`	string	Relationship type (e.g., PARTICIPATED_AS)
`properties_json`	string	Edge properties as JSON

Positions (`positions.parquet`)

Column	Type	Description
`node_id`	string	Node UUID
`x`, `y`, `z`	float	3D coordinates
`size`	float	Node size (Graph Gravity weighted)
`r`, `g`, `b`	int	RGB color by entity type
`community`	int	Louvain community index (seeded, deterministic)
`tier`	string (nullable)	Graph Gravity tier

Layout method (schema ≥ 1.2.0): Coordinates are derived from the entities' semantic text embeddings (UMAP with a fixed seed and PCA initialization), so narratively similar entities sit near each other. Non-embedded nodes (events, scenes, episodes, etc.) are placed at the weighted barycenter of their narrative neighbours. The layout is deterministic: re-exporting an unchanged graph reproduces identical coordinates, and lightly-changed graphs keep comparable layouts. Not comparable with positions published under schema ≤ 1.1.0, which used a non-deterministic node2vec structural embedding. See meta.json → positions for the exact method and coverage stats.

Usage

from datasets import load_dataset
import pandas as pd

# Load from HuggingFace
ds = load_dataset("brandburner/doctorwho-s12-narrative-kg")

# Or load parquet directly
nodes = pd.read_parquet("nodes.parquet")
edges = pd.read_parquet("edges.parquet")

# Filter to anchor characters
anchors = nodes[(nodes['primary_label'] == 'Agent') & (nodes['tier'] == 'anchor')]

# Build a NetworkX graph
import networkx as nx
G = nx.DiGraph()
for _, n in nodes.iterrows():
    G.add_node(n['node_id'], label=n['primary_label'], name=n['name'])
for _, e in edges.iterrows():
    G.add_edge(e['source_node_id'], e['target_node_id'], type=e['relationship_type'])

Citation

@misc{fabula_doctorwho_s12,
  title = {Doctor Who Narrative Knowledge Graph},
  author = {Fabula Pipeline},
  year = {2026},
  publisher = {HuggingFace},
  howpublished = {\url{https://huggingface.co/datasets/brandburner/doctorwho-s12-narrative-kg}}
}

License

CC BY-SA 4.0