Datasets:

joyboseroy
/

MarxGraph

license: cc-by-4.0
language:
  - en
tags:
  - knowledge-graph
  - intellectual-history
  - marxism
  - political-philosophy
  - computational-humanities
  - temporal-graph
task_categories:
  - text-classification
  - graph-ml
pretty_name: MarxGraph
size_categories:
  - 10K<n<100K

MarxGraph: A Temporal Knowledge Graph of Marxist Intellectual History

MarxGraph traces how core concepts in Marxist theory (class struggle, the vanguard party, imperialism, permanent revolution, and 40+ others) were inherited, extended, contested, or rejected as they moved from Marx and Engels through Lenin, Luxemburg, Trotsky, Stalin, and Mao. Unlike a simple "who-influenced-whom" graph, every edge in MarxGraph is grounded in a specific textual claim with a verbatim evidence span, and the dataset explicitly distinguishes documentary facts (what a text says) from interpretive judgments (how a later thinker's treatment relates to an earlier one); the latter always carrying a perspective and confidence field rather than being asserted as settled fact.

Dataset summary


Works	50
Passages	~14,862
Concept mentions	21,933
Claims (with evidence spans)	71,572
Explicit references/citations	19,471
Concept-evolution edges	446
Thinkers	Marx, Engels, Lenin, Luxemburg, Trotsky, Stalin, Mao
Traditions	Classical Marxism, Leninism, Marxism-Leninism, Trotskyism, Maoism

Source texts: the Marxists Internet Archive.

Dataset structure

Six linked tables (Parquet):

File	Grain	Key columns
`works.parquet`	one work	`work_id`, `author`, `title`, `year`, `tradition`, `license`, `source_url`
`passages.parquet`	~220-word passage	`passage_id`, `work_id`, `chapter`, `text`
`concept_mentions.parquet`	one mention	`passage_id`, `concept`, `surface_form`
`claims.parquet`	one proposition	`claim_id`, `claim`, `concepts`, `evidence_span`, `confidence`
`references.parquet`	explicit citation/attack	`work_id`, `target_thinker`, `stance`, `evidence_span`
`concept_evolution.parquet`	concept × (earlier, later) author pair	`transformation`, `rationale`, `confidence`, `perspective`

Plus a graph export (graph/marxgraph.graphml, graph/marxgraph.jsonl) built from the tables above: thinkers, works, claims, and concepts as nodes; WROTE, CONTAINS, MENTIONS, citation-stance, and TRANSFORMS_<label> edges.

Transformation labels

Each concept_evolution row classifies how a concept changed between two chronologically-ordered thinkers:

PRESERVED: restated with essentially the same content and scope
EXTENDED: same core meaning, new components or scope added
REFORMULATED: same term, different core content or theoretical role
CONTEXTUALIZED: adapted to new historical/national conditions
CONTESTED: later thinker explicitly argues against the earlier treatment
REJECTED: concept abandoned or repudiated
INSUFFICIENT: evidence too thin to support any label (the model declines to force a verdict rather than fabricate one)

Distribution across the 446 edges: 76.0% EXTENDED, 10.1% CONTEXTUALIZED, 6.3% REFORMULATED, 4.7% CONTESTED, 2.0% PRESERVED, 0.7% INSUFFICIENT, 0.2% REJECTED.

Construction

Built with a five-stage pipeline (full code included): EPUB download → passage segmentation → LLM claim/mention/reference extraction → LLM concept-transformation typing (constrained to a documented set of plausible lineage pairs) → graph assembly. Primary extraction model: openai/gpt-oss-120b via Groq. See README.md in the code repository for the full pipeline and config/ontology.yaml for the entity/relation schema.

Validation

100 claims and 50 evolution edges were checked against a stratified sample protocol, with an automated first pass followed by domain-expert human adjudication of every uncertain case. Result: 99% claims accuracy (99/100), 100% evolution-edge coherence (50/50) on the reviewed sample. Full methodology, including the one documented failure case and important caveats about sample size and the narrowness of the model's confidence scores, is in VALIDATION.md; please read it before citing accuracy figures out of context.

Known limitations

Confidence scores are not calibrated. Evolution-edge confidence values cluster tightly (mean 0.852, std 0.035) and should not be used as a reliability filter.
A likely duplicate work has not yet been removed. Trotsky's History of the Russian Revolution appears to be ingested twice (single-file and 3-volume editions), inflating Trotsky's claim counts on affected concepts. A deduplication tool (src/dedupe_works.py) exists but has not been applied to this release.
Interpretive edges are one LLM-assisted reading, not a canonical verdict. Especially for ideologically contested relationships (e.g., Stalin's relationship to Lenin, or Trotsky's to Stalin), different traditions characterize continuity/rupture differently. Treat CONTESTED/REJECTED labels as a documented, evidence-linked interpretation, open to challenge, not as settled historiography.
Extraction quality reflects the extraction model. openai/gpt-oss-120b is far smaller than frontier models; the validation pass documents at least one case of the model adding a plausible but unsupported evaluative gloss to an otherwise accurate paraphrase.
Licensing is mixed at the work level. works.parquet.license marks each work pd (public domain, safe to redistribute as text) or verify (check the specific translation/edition's copyright status on MIA before redistributing raw text; this applies especially to some Stalin, Mao, and Luxemburg translations). Annotations (claims, relations, evolution edges) are original analysis and released under this dataset's license regardless of source-text status.

Citation

If you use MarxGraph, please cite the dataset directly:

Bose, J. (2026). MarxGraph: A Temporal Knowledge Graph of Marxist Intellectual
History [Dataset]. Hugging Face. https://huggingface.co/datasets/<your-repo-id>

Credits

Source texts: Marxists Internet Archive, a volunteer-run archive; please support their work if you can.

Dataset construction, ontology design, and validation: Joy Bose (joy.bose@ieee.org).