rishhh's picture
Train dense-retrieval-adapted SciFact verifier
42317a6 verified
Raw
History Blame Contribute Delete
9.86 kB
{
"datasets": {
"hard_nei_source": "andreiaalexa/scifact-relevance-pairs title/train and BM25-retrieved non-cited SciFact train documents",
"test_policy": "Validation rows with train-overlapping normalized claims are removed; remaining validation rows are split by normalized claim key into validation and derived test.",
"verifier_primary": "allenai/scifact_entailment train plus allenai/scifact train retrieval-aware snippets"
},
"full_scifact_validation": {
"device": "cuda",
"end_to_end": {
"top1": {
"accuracy": 0.5711111111111111,
"confusion_matrix": {
"NOT_ENOUGH_INFO": {
"NOT_ENOUGH_INFO": 82,
"REFUTES": 7,
"SUPPORTS": 23
},
"REFUTES": {
"NOT_ENOUGH_INFO": 67,
"REFUTES": 47,
"SUPPORTS": 8
},
"SUPPORTS": {
"NOT_ENOUGH_INFO": 68,
"REFUTES": 20,
"SUPPORTS": 128
}
},
"examples": 450,
"gold_label_distribution": {
"NOT_ENOUGH_INFO": 112,
"REFUTES": 122,
"SUPPORTS": 216
},
"macro_f1": 0.5535795821874849,
"per_class": {
"NOT_ENOUGH_INFO": {
"f1-score": 0.49848024316109424,
"precision": 0.3778801843317972,
"recall": 0.7321428571428571,
"support": 112.0
},
"REFUTES": {
"f1-score": 0.47959183673469385,
"precision": 0.6351351351351351,
"recall": 0.38524590163934425,
"support": 122.0
},
"SUPPORTS": {
"f1-score": 0.6826666666666666,
"precision": 0.8050314465408805,
"recall": 0.5925925925925926,
"support": 216.0
},
"accuracy": 0.5711111111111111,
"macro avg": {
"f1-score": 0.5535795821874849,
"precision": 0.606015588669271,
"recall": 0.5699937837915979,
"support": 450.0
},
"weighted avg": {
"f1-score": 0.5817688695903894,
"precision": 0.6526574657432843,
"recall": 0.5711111111111111,
"support": 450.0
}
},
"predicted_label_distribution": {
"NOT_ENOUGH_INFO": 217,
"REFUTES": 74,
"SUPPORTS": 159
},
"strategy": "top1",
"supported_refuted_accuracy_when_evidence_retrieved_at_5": 0.6415094339622641,
"supported_refuted_examples_with_evidence_retrieved_at_5": 265
},
"top5_decisive": {
"accuracy": 0.6355555555555555,
"confusion_matrix": {
"NOT_ENOUGH_INFO": {
"NOT_ENOUGH_INFO": 61,
"REFUTES": 12,
"SUPPORTS": 39
},
"REFUTES": {
"NOT_ENOUGH_INFO": 48,
"REFUTES": 53,
"SUPPORTS": 21
},
"SUPPORTS": {
"NOT_ENOUGH_INFO": 29,
"REFUTES": 15,
"SUPPORTS": 172
}
},
"examples": 450,
"gold_label_distribution": {
"NOT_ENOUGH_INFO": 112,
"REFUTES": 122,
"SUPPORTS": 216
},
"macro_f1": 0.5935365393682225,
"per_class": {
"NOT_ENOUGH_INFO": {
"f1-score": 0.488,
"precision": 0.4420289855072464,
"recall": 0.5446428571428571,
"support": 112.0
},
"REFUTES": {
"f1-score": 0.5247524752475248,
"precision": 0.6625,
"recall": 0.4344262295081967,
"support": 122.0
},
"SUPPORTS": {
"f1-score": 0.7678571428571429,
"precision": 0.7413793103448276,
"recall": 0.7962962962962963,
"support": 216.0
},
"accuracy": 0.6355555555555555,
"macro avg": {
"f1-score": 0.5935365393682225,
"precision": 0.6153027652840247,
"recall": 0.5917884609824501,
"support": 450.0
},
"weighted avg": {
"f1-score": 0.6322954329718686,
"precision": 0.6454892831362097,
"recall": 0.6355555555555555,
"support": 450.0
}
},
"predicted_label_distribution": {
"NOT_ENOUGH_INFO": 138,
"REFUTES": 80,
"SUPPORTS": 232
},
"strategy": "top5_decisive",
"supported_refuted_accuracy_when_evidence_retrieved_at_5": 0.7660377358490567,
"supported_refuted_examples_with_evidence_retrieved_at_5": 265
}
},
"retrieval": {
"evaluated_queries": 450.0,
"mrr@5": 0.5610740740740741,
"ndcg@1": 0.4866666666666667,
"ndcg@3": 0.5487532554897722,
"ndcg@5": 0.5739882576045233,
"recall@1": 0.4866666666666667,
"recall@3": 0.6355555555555555,
"recall@5": 0.6755555555555556
}
},
"hub_model_id": "rishhh/verisci-claim-verifier-dense-adapted-seed123",
"latency_seconds": 193.2834756150114,
"limitations": [
"The derived test split is small and originates from SciFact validation, not an official hidden test set.",
"Retrieval-aware negatives are generated by BM25 over SciFact train and may overfit SciFact retrieval artifacts.",
"Full SciFact validation is public and should be treated as a promotion gate, not a hidden benchmark.",
"SciFact is CC-BY-NC-2.0, so downstream use should respect non-commercial licensing constraints."
],
"max_length": 512,
"model_name": "rishhh/verisci-claim-verifier-retrieval-adapted-seed123",
"seed": 123,
"test": {
"accuracy": 0.893491124260355,
"confusion_matrix": {
"NOT_ENOUGH_INFO": {
"NOT_ENOUGH_INFO": 70,
"REFUTES": 0,
"SUPPORTS": 1
},
"REFUTES": {
"NOT_ENOUGH_INFO": 4,
"REFUTES": 20,
"SUPPORTS": 7
},
"SUPPORTS": {
"NOT_ENOUGH_INFO": 1,
"REFUTES": 5,
"SUPPORTS": 61
}
},
"examples": 169,
"expected_calibration_error": 0.09891798686699052,
"gold_label_distribution": {
"NOT_ENOUGH_INFO": 71,
"REFUTES": 31,
"SUPPORTS": 67
},
"macro_f1": 0.8567495491347223,
"per_class": {
"NOT_ENOUGH_INFO": {
"f1-score": 0.958904109589041,
"precision": 0.9333333333333333,
"recall": 0.9859154929577465,
"support": 71.0
},
"REFUTES": {
"f1-score": 0.7142857142857143,
"precision": 0.8,
"recall": 0.6451612903225806,
"support": 31.0
},
"SUPPORTS": {
"f1-score": 0.8970588235294118,
"precision": 0.8840579710144928,
"recall": 0.9104477611940298,
"support": 67.0
},
"accuracy": 0.893491124260355,
"macro avg": {
"f1-score": 0.8567495491347223,
"precision": 0.8724637681159421,
"recall": 0.847174848158119,
"support": 169.0
},
"weighted avg": {
"f1-score": 0.8895147343204121,
"precision": 0.8893405368321755,
"recall": 0.893491124260355,
"support": 169.0
}
},
"predicted_label_distribution": {
"NOT_ENOUGH_INFO": 75,
"REFUTES": 25,
"SUPPORTS": 69
},
"strategy": "test"
},
"training": {
"base_scifact_entailment_train_examples": 916,
"class_weights": {
"NOT_ENOUGH_INFO": 0.9040744304656982,
"REFUTES": 1.4886423349380493,
"SUPPORTS": 0.8182346224784851
},
"dense_retrieved_nei_max": 350,
"dense_retriever_id": "rishhh/verisci-scifact-e5-retriever",
"effective_batch_size": 16,
"epochs": 2,
"hard_title_nei_max": 150,
"hybrid_alpha": 0.75,
"learning_rate": 6e-06,
"retrieval_aware_examples_added": 2514,
"retrieved_nei_max": 250,
"test_examples": 169,
"train_examples": 2818,
"train_label_distribution": {
"NOT_ENOUGH_INFO": 1039,
"REFUTES": 631,
"SUPPORTS": 1148
},
"validation_examples": 168,
"validation_rows_removed_for_train_claim_overlap": 2
},
"validation": {
"accuracy": 0.8630952380952381,
"confusion_matrix": {
"NOT_ENOUGH_INFO": {
"NOT_ENOUGH_INFO": 55,
"REFUTES": 0,
"SUPPORTS": 2
},
"REFUTES": {
"NOT_ENOUGH_INFO": 5,
"REFUTES": 28,
"SUPPORTS": 7
},
"SUPPORTS": {
"NOT_ENOUGH_INFO": 3,
"REFUTES": 6,
"SUPPORTS": 62
}
},
"examples": 168,
"expected_calibration_error": 0.11418107826085311,
"gold_label_distribution": {
"NOT_ENOUGH_INFO": 57,
"REFUTES": 40,
"SUPPORTS": 71
},
"macro_f1": 0.8488876200143807,
"per_class": {
"NOT_ENOUGH_INFO": {
"f1-score": 0.9166666666666666,
"precision": 0.873015873015873,
"recall": 0.9649122807017544,
"support": 57.0
},
"REFUTES": {
"f1-score": 0.7567567567567568,
"precision": 0.8235294117647058,
"recall": 0.7,
"support": 40.0
},
"SUPPORTS": {
"f1-score": 0.8732394366197183,
"precision": 0.8732394366197183,
"recall": 0.8732394366197183,
"support": 71.0
},
"accuracy": 0.8630952380952381,
"macro avg": {
"f1-score": 0.8488876200143807,
"precision": 0.8565949071334323,
"recall": 0.8460505724404909,
"support": 168.0
},
"weighted avg": {
"f1-score": 0.860239703989704,
"precision": 0.8613278644791249,
"recall": 0.8630952380952381,
"support": 168.0
}
},
"predicted_label_distribution": {
"NOT_ENOUGH_INFO": 63,
"REFUTES": 34,
"SUPPORTS": 71
},
"strategy": "validation"
}
}