Commit History
23 days ago
-
Retrained with pointwise distillation: scores on v3's scale without calibration
Verified signatureStages 2-3 retrained with a per-candidate pointwise distillation term towards Mizan-Rerank-v3's scores, replacing the post-hoc +3.26 logit calibration (calibration.json removed), which had pushed traps and partially relevant passages to 0.6-0.9 on general/adversarial text. Held-out mean nDCG@10 0.871 (was 0.875), fiqh nDCG@10 0.748 (was 0.761), relevant_kept@8 0.735 without calibration, contradiction-first 19.3% (was 16.9%). Model card numbers and charts regenerated; training section removed.107efea
24 days ago
-
Update README.md
Verified signature87706ea -
Model card: remove the Sistani mention, link the live demo
Verified signature9db97d0 -
Model card: drop links to the local benchmark files
Verified signature39fcc71 -
Keep the benchmark harness and raw results local (not published)
Verified signature9da6d6f -
Charts: 4 models (Mizan-Rerank-V2 removed), one-row legend and labels outside the bars
Verified signature0df80a2 -
Model card: family table lists the current models only
Verified signature1966801 -
Model card: footnote on the held-out average used in the family table
Verified signature46ad1e7 -
Remove stray __pycache__
Verified signaturefbc808d -
Benchmark results: keep summaries only
Verified signatureaac2efd -
Mizan-Rerank-v3-Turbo: 150M Arabic cross-encoder (3-stage distillation from bge-reranker-v2-m3 + Mizan-Rerank-v3), model card, benchmarks and charts
Verified signature119e3d5 -
initial commit
Verified signaturecd51ced