Research-Stack/shared-data/pist_trace_routing_benchmark.json

326 lines
No EOL
7.7 KiB
JSON

{
"n_predictions": 20,
"n_correct": 6,
"tactic_family_accuracy": 0.3,
"average_complexity_reduction": -16.399,
"average_size_reduction": 2.0,
"per_family_accuracy": {
"discharge": {
"correct": 5,
"total": 11,
"accuracy": 0.4545
},
"unknown": {
"correct": 0,
"total": 1,
"accuracy": 0.0
},
"lemma_introduction": {
"correct": 0,
"total": 1,
"accuracy": 0.0
},
"introduction": {
"correct": 0,
"total": 3,
"accuracy": 0.0
},
"rewrite": {
"correct": 0,
"total": 3,
"accuracy": 0.0
},
"normalization": {
"correct": 1,
"total": 1,
"accuracy": 1.0
}
},
"confusion_matrix": {
"discharge": {
"normalization": 4,
"discharge": 5,
"case_analysis": 2,
"introduction": 0,
"lemma_introduction": 0,
"rewrite": 0,
"unknown": 0
},
"unknown": {
"discharge": 1,
"case_analysis": 0,
"introduction": 0,
"lemma_introduction": 0,
"normalization": 0,
"rewrite": 0,
"unknown": 0
},
"lemma_introduction": {
"normalization": 1,
"case_analysis": 0,
"discharge": 0,
"introduction": 0,
"lemma_introduction": 0,
"rewrite": 0,
"unknown": 0
},
"introduction": {
"normalization": 2,
"discharge": 1,
"case_analysis": 0,
"introduction": 0,
"lemma_introduction": 0,
"rewrite": 0,
"unknown": 0
},
"rewrite": {
"normalization": 2,
"discharge": 1,
"case_analysis": 0,
"introduction": 0,
"lemma_introduction": 0,
"rewrite": 0,
"unknown": 0
},
"normalization": {
"normalization": 1,
"case_analysis": 0,
"discharge": 0,
"introduction": 0,
"lemma_introduction": 0,
"rewrite": 0,
"unknown": 0
},
"case_analysis": {
"case_analysis": 0,
"discharge": 0,
"introduction": 0,
"lemma_introduction": 0,
"normalization": 0,
"rewrite": 0,
"unknown": 0
}
},
"predictions": [
{
"theorem": "apply_chain",
"step": 1,
"actual_tactic": "apply h1",
"actual_family": "discharge",
"predicted_family": "normalization",
"correct": false,
"complexity_delta": -14.5643,
"partial_size": 2,
"full_size": 4
},
{
"theorem": "apply_chain",
"step": 2,
"actual_tactic": "exact hA",
"actual_family": "discharge",
"predicted_family": "discharge",
"correct": true,
"complexity_delta": -18.0138,
"partial_size": 4,
"full_size": 6
},
{
"theorem": "cases_and_elim",
"step": 1,
"actual_tactic": "apply h2",
"actual_family": "discharge",
"predicted_family": "normalization",
"correct": false,
"complexity_delta": -14.5643,
"partial_size": 2,
"full_size": 4
},
{
"theorem": "cases_and_elim",
"step": 2,
"actual_tactic": "assumption",
"actual_family": "unknown",
"predicted_family": "discharge",
"correct": false,
"complexity_delta": -18.0138,
"partial_size": 4,
"full_size": 6
},
{
"theorem": "cases_or_swap",
"step": 1,
"actual_tactic": "\u00b7 right; exact h",
"actual_family": "discharge",
"predicted_family": "normalization",
"correct": false,
"complexity_delta": -14.5643,
"partial_size": 2,
"full_size": 4
},
{
"theorem": "cases_or_swap",
"step": 2,
"actual_tactic": "\u00b7 left; exact h",
"actual_family": "discharge",
"predicted_family": "discharge",
"correct": true,
"complexity_delta": -18.0138,
"partial_size": 4,
"full_size": 6
},
{
"theorem": "constructor_example",
"step": 1,
"actual_tactic": "\u00b7 exact hA",
"actual_family": "discharge",
"predicted_family": "normalization",
"correct": false,
"complexity_delta": -14.5643,
"partial_size": 2,
"full_size": 4
},
{
"theorem": "constructor_example",
"step": 2,
"actual_tactic": "\u00b7 exact hB",
"actual_family": "discharge",
"predicted_family": "discharge",
"correct": true,
"complexity_delta": -18.0138,
"partial_size": 4,
"full_size": 6
},
{
"theorem": "have_chain",
"step": 1,
"actual_tactic": "have hC : C := hBC hB",
"actual_family": "lemma_introduction",
"predicted_family": "normalization",
"correct": false,
"complexity_delta": -14.5643,
"partial_size": 2,
"full_size": 4
},
{
"theorem": "have_chain",
"step": 2,
"actual_tactic": "exact hC",
"actual_family": "discharge",
"predicted_family": "discharge",
"correct": true,
"complexity_delta": -18.0138,
"partial_size": 4,
"full_size": 6
},
{
"theorem": "intro_all",
"step": 1,
"actual_tactic": "intro hB",
"actual_family": "introduction",
"predicted_family": "normalization",
"correct": false,
"complexity_delta": -14.5643,
"partial_size": 2,
"full_size": 4
},
{
"theorem": "intro_all",
"step": 2,
"actual_tactic": "intro hC",
"actual_family": "introduction",
"predicted_family": "discharge",
"correct": false,
"complexity_delta": -18.0138,
"partial_size": 4,
"full_size": 6
},
{
"theorem": "intro_all",
"step": 3,
"actual_tactic": "exact hA",
"actual_family": "discharge",
"predicted_family": "case_analysis",
"correct": false,
"complexity_delta": -19.113,
"partial_size": 6,
"full_size": 8
},
{
"theorem": "intro_apply",
"step": 1,
"actual_tactic": "intro hA",
"actual_family": "introduction",
"predicted_family": "normalization",
"correct": false,
"complexity_delta": -14.5643,
"partial_size": 2,
"full_size": 4
},
{
"theorem": "intro_apply",
"step": 2,
"actual_tactic": "apply h",
"actual_family": "discharge",
"predicted_family": "discharge",
"correct": true,
"complexity_delta": -18.0138,
"partial_size": 4,
"full_size": 6
},
{
"theorem": "intro_apply",
"step": 3,
"actual_tactic": "exact hA",
"actual_family": "discharge",
"predicted_family": "case_analysis",
"correct": false,
"complexity_delta": -19.113,
"partial_size": 6,
"full_size": 8
},
{
"theorem": "rw_chain_3step",
"step": 1,
"actual_tactic": "rw [h2]",
"actual_family": "rewrite",
"predicted_family": "normalization",
"correct": false,
"complexity_delta": -14.5643,
"partial_size": 2,
"full_size": 4
},
{
"theorem": "rw_chain_3step",
"step": 2,
"actual_tactic": "rw [h3]",
"actual_family": "rewrite",
"predicted_family": "discharge",
"correct": false,
"complexity_delta": -18.0138,
"partial_size": 4,
"full_size": 6
},
{
"theorem": "rw_chain_eq",
"step": 1,
"actual_tactic": "rw [h2]",
"actual_family": "rewrite",
"predicted_family": "normalization",
"correct": false,
"complexity_delta": -14.5643,
"partial_size": 2,
"full_size": 4
},
{
"theorem": "rw_chain_mixed",
"step": 1,
"actual_tactic": "simp",
"actual_family": "normalization",
"predicted_family": "normalization",
"correct": true,
"complexity_delta": -14.5643,
"partial_size": 2,
"full_size": 4
}
]
}