SciBERT-SolarPhysics-Search

Canonical model rebuilt for the major revision, trained exclusively on Nucleo_core, PIML_core, and CombFinal_core.

Method

• DAPT (MLM) on the domain-specific core corpus.
• Contrastive fine-tuning with query_side -> positive_side.
• ML_Multimodal was not used for training.

Compatibility with the paper

• Notebook 05 computes perplexity, NMI, ARI, Silhouette, MRR, Recall@K, and NearestCentroidAcc.
• Notebooks 06 through 08 compute the approved major-revision extension (generic SciBERT, BM25, core versus holdout, bootstrap confidence intervals, and the final audit).

DAPT

• eval_loss: 1.183335781097412
• perplexity: 3.2652482094089605

Class-based retrieval (technical evaluation)

model	MRR	Recall@10	Recall@50	Recall@100
SciBERT-baseline	0.608464	0.864165	0.984858	0.994514
SciBERT-SolarPhysics-Search	0.659382	0.893131	0.989686	0.996269

Clustering (technical evaluation)

model	k	NMI	ARI	Silhouette
SciBERT-baseline	16	0.0860483	0.0318237	0.123657
SciBERT-SolarPhysics-Search	16	0.0984312	0.0437347	0.150183

Centroid-based separability

model	NearestCentroidAcc
SciBERT-baseline	0.288786
SciBERT-SolarPhysics-Search	0.453149