Eklav passage-reranking ablation & baseline checkpoints: answer-only SFT and the mask-only loss-masking ablation.
Adarsh Singh
AdarshSingh7647
AI & ML interests
Information Retrieval over structured data and retrieval-augmented question answering. Focuses on retrieval, reranking, and representation learning methods that enable robust and scalable reasoning over structured data.
Recent Activity
updated a collection 6 days ago
ablation updated a collection 6 days ago
ablation updated a collection 6 days ago
ablationOrganizations
None yet
Eklav checkpoints for passage reranking (BRIGHT, NevIR), plus the CotGen baseline for comparison. Same base model and data, different training method.
-
AdarshSingh7647/Eklav-0.6B-Reranker
Text Generation • 0.6B • Updated • 450 -
AdarshSingh7647/Eklav-0.6B-Reranker-CotGen
Text Generation • 0.6B • Updated • 473 -
AdarshSingh7647/Eklav-4B-Reranker
Text Generation • 4B • Updated • 446 -
AdarshSingh7647/Eklav-4B-Reranker-CotGen
Text Generation • 4B • Updated • 431
Qwen3-8B listwise table rerankers (arXiv:2607.25182). 3 variants: Naive (fast), Standard SFT (reasoning), TabRank (best). Start with TabRank.
Eklav passage-reranking ablation & baseline checkpoints: answer-only SFT and the mask-only loss-masking ablation.
Eklav checkpoints for math reasoning, plus the CotGen baseline for comparison. Same base model and data, different training method.
Eklav checkpoints for passage reranking (BRIGHT, NevIR), plus the CotGen baseline for comparison. Same base model and data, different training method.
-
AdarshSingh7647/Eklav-0.6B-Reranker
Text Generation • 0.6B • Updated • 450 -
AdarshSingh7647/Eklav-0.6B-Reranker-CotGen
Text Generation • 0.6B • Updated • 473 -
AdarshSingh7647/Eklav-4B-Reranker
Text Generation • 4B • Updated • 446 -
AdarshSingh7647/Eklav-4B-Reranker-CotGen
Text Generation • 4B • Updated • 431
Data generated as augmentations to NQ-Tables and OTT-QA for Table Retrieval
Qwen3-8B listwise table rerankers (arXiv:2607.25182). 3 variants: Naive (fast), Standard SFT (reasoning), TabRank (best). Start with TabRank.