Rohit Kumar Salla commited on
Commit
1310d8c
·
1 Parent(s): 2d35edf

move files to repo root

Browse files
cross_query_benchmark/benchmark_dev_v2.json ADDED
The diff for this file is too large to render. See raw diff
 
cross_query_benchmark/benchmark_full_v2.json ADDED
The diff for this file is too large to render. See raw diff
 
cross_query_benchmark/benchmark_test_v2.json ADDED
The diff for this file is too large to render. See raw diff
 
cross_query_benchmark/benchmark_train_v2.json ADDED
The diff for this file is too large to render. See raw diff
 
cross_query_benchmark/evaluation_schema_v2.json ADDED
@@ -0,0 +1,23 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ {
2
+ "per_query_metrics": {
3
+ "accuracy": "Fraction of correct labels",
4
+ "macro_f1": "Macro F1 across 3 classes (ENTAILED, CONTRADICTED, UNKNOWN)",
5
+ "unknown_f1": "F1 for UNKNOWN class specifically"
6
+ },
7
+ "set_level_metrics": {
8
+ "SetConsRate": "Fraction of bundles where final belief state Bn is satisfiable",
9
+ "AUC_PrefixCons": "1/n * sum I[Bt is SAT] for sequential setting",
10
+ "RevisionCost": "Average commitments revised to restore satisfiability",
11
+ "ContradictionDensity": "Fraction of steps where B_{t-1} SAT but B_t UNSAT"
12
+ },
13
+ "compute_metrics": {
14
+ "overhead": "Wall-clock time normalized to No-CoT baseline",
15
+ "solver_calls_per_bundle": "Average SAT/SMT calls per bundle"
16
+ },
17
+ "quality_metadata": {
18
+ "labels_verified_by": "z3 solver (SAT/SMT)",
19
+ "unique_query_ratio": 0.824,
20
+ "cross_query_dependencies": true,
21
+ "formal_representations": true
22
+ }
23
+ }
cross_query_benchmark/prompt_templates_v2.json ADDED
@@ -0,0 +1,14 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ {
2
+ "commitment_extraction": {
3
+ "system": "You are a logical commitment extractor. Given a case file, query, and label, extract the symbolic commitment.",
4
+ "template": "Task: Extract the logical commitment from this answer.\n\nInstructions:\n- ENTAILED: identify the key atom \u03c7 that must be true.\n- CONTRADICTED: identify the key atom \u03c7 that must be false.\n- UNKNOWN: output 'NONE' (no commitment).\n\nCase File: {premises}\nQuery: {query}\nLabel: {label}\nReasoning: {reasoning}\n\nExtracted Commitment:"
5
+ },
6
+ "repair_prompt": {
7
+ "system": "You are a belief-state repair agent. Propose minimal repairs for conflicting commitments.",
8
+ "template": "Conflicting commitments:\n{conflicting_commitments}\n\nCase File: {premises}\nCurrent Query: {query}\nCurrent Label: {label}\n\nPropose minimal repair:\n1. Revise current label\n2. Weaken commitment\n3. Retract prior commitment\n\nRepair:"
9
+ },
10
+ "multi_query_reasoning": {
11
+ "system": "You are a logical reasoner. Answer each query about the case file with ENTAILED, CONTRADICTED, or UNKNOWN.",
12
+ "template": "Case File:\n{premises}\n\nFor each query, determine:\n- ENTAILED: the premises logically imply the statement\n- CONTRADICTED: the premises logically contradict the statement\n- UNKNOWN: the premises neither prove nor refute the statement\n\nQueries:\n{queries}\n\nAnswer each with: Reasoning: ... Answer: ENTAILED/CONTRADICTED/UNKNOWN"
13
+ }
14
+ }