SciBERT-SolarPhysics-Search

Modelo canonico reconstruido para o major review, treinado somente em Nucleo_core, PIML_core e CombFinal_core.

Metodo

• DAPT (MLM) sobre o corpus de dominio do core.
• Fine-tuning contrastivo com query_side -> positive_side.
• Sem uso de ML_Multimodal no treino.

Compatibilidade com o paper

• O notebook 05 calcula perplexity, NMI, ARI, Silhouette, MRR, Recall@K e NearestCentroidAcc.
• Os notebooks 06 a 08 calculam o incremento aprovado do major review (SciBERT generic, BM25, core vs holdout, bootstrap CIs e auditoria final).

DAPT

• eval_loss: 1.183335781097412
• perplexity: 3.2652482094089605

Retrieval por classe (Tecnica)

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

Clusterizacao (Tecnica)

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

Separabilidade por centroide

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