id stringclasses 10 values | graph_nodes stringclasses 5 values | graph_edges stringclasses 5 values | intervention stringclasses 5 values | baseline_state stringclasses 5 values | counterfactual_state stringclasses 10 values | predicted_changes stringclasses 10 values | expected_changed_nodes stringclasses 5 values | expected_unchanged_nodes stringclasses 4 values | separation_label stringclasses 3 values | separation_failure_signature stringclasses 6 values | constraints stringclasses 1 value | gold_checklist stringclasses 9 values |
|---|---|---|---|---|---|---|---|---|---|---|---|---|
CIG-CSI-001 | X,M,Y,Z | X->M;M->Y | do(X=1) | X=0,M=0,Y=0,Z=0 | X=1,M=1,Y=1,Z=0 | Model: M,Y change; Z stable | M,Y | Z | separated | none | Under 140 words. | 1 only descendants change |
CIG-CSI-002 | X,M,Y,Z | X->M;M->Y | do(X=1) | X=0,M=0,Y=0,Z=0 | X=1,M=1,Y=1,Z=1 | Model: Z changes too | M,Y | Z | bleed | non-descendant-change | Under 140 words. | 1 Z must stay |
CIG-CSI-003 | A,B,C,D | A->B;A->C | do(A=0) | A=1,B=1,C=1,D=1 | A=0,B=0,C=0,D=1 | Model: B,C change; D stable | B,C | D | separated | none | Under 140 words. | 1 siblings change; D stable |
CIG-CSI-004 | A,B,C,D | A->B;A->C | do(A=0) | A=1,B=1,C=1,D=1 | A=0,B=0,C=0,D=0 | Model: D changes with no edge | B,C | D | bleed | spurious-cascade | Under 140 words. | 1 D must not change |
CIG-CSI-005 | S,T,U,V | S->T;T->U;S->V | do(T=0) | S=1,T=1,U=1,V=1 | S=1,T=0,U=0,V=1 | Model: U changes; V stable | U | V | separated | none | Under 140 words. | 1 V not descendant of T |
CIG-CSI-006 | S,T,U,V | S->T;T->U;S->V | do(T=0) | S=1,T=1,U=1,V=1 | S=0,T=0,U=0,V=0 | Model: everything changes | U | V | global-drift | everything-changes | Under 140 words. | 1 S not intervened |
CIG-CSI-007 | R,P,Q,W | R->P;R->Q | do(P=1) | R=0,P=0,Q=0,W=0 | R=0,P=1,Q=1,W=0 | Model: Q changes though not child of P | P | Q,W | bleed | wrong-descendant | Under 140 words. | 1 Q depends on R not P |
CIG-CSI-008 | R,P,Q,W | R->P;R->Q | do(P=1) | R=0,P=0,Q=0,W=0 | R=0,P=1,Q=0,W=0 | Model: only P changes | P | Q,W | separated | none | Under 140 words. | 1 Q stable |
CIG-CSI-009 | X,Y,Z | X->Y | do(X=0) | X=1,Y=1,Z=1 | X=0,Y=0,Z=1 | Model: Y changes; Z stable | Y | Z | separated | none | Under 140 words. | 1 Z stable |
CIG-CSI-010 | X,Y,Z | X->Y | do(X=0) | X=1,Y=1,Z=1 | X=0,Y=0,Z=0 | Model: Z changes without edge | Y | Z | bleed | unlinked-node-change | Under 140 words. | 1 Z must stay |
What this dataset tests
Whether a counterfactual respects causal separation.
Change one node.
Only descendants should change.
Non-descendants should remain stable.
Why this exists
Models often treat counterfactuals like stories.
They let everything drift.
This benchmark tests graph discipline.
Data format
Each row contains
- a node set and edge list
- an intervention do(X=...)
- baseline and counterfactual states
- expected changed nodes
- expected unchanged nodes
Labels
- separated
- bleed
- global-drift
Bleed means a non-descendant changed.
Global-drift means unrelated nodes drifted broadly.
Suggested prompt wrapper
System
You evaluate whether the counterfactual respects causal separation.
User
Nodes
{graph_nodes}
Edges
{graph_edges}
Intervention
{intervention}
Baseline
{baseline_state}
Counterfactual
{counterfactual_state}
Return
- one label
- one sentence naming which nodes violated separation
Citation
ClarusC64 dataset family
- Downloads last month
- 17