ymelka/camembert-med-reranker-v2-distil — v2 E1 gold+replay (Distil 68M)

DistilCamemBERT 68M fine-tuné sur 30 846 query groups (171k paires) : dataset-v1 train (17 843) + gold-extra v2_dev (3 003, splits officiels inutilisés) + mmarco-replay 10k (MS MARCO FR, reservoir 100k, BM25). Tier mix 71% A / 29% C.

Medical (legacy 8 965 queries)

• P@1 0.4632 MRR 0.6561 nDCG 0.7352

mMARCO forgetting (200 queries dev slice, mêmes candidates)

• P@1 0.970 (vs baseline distil 0.865, vs v1 medical 0.610) — forgetting corrigé

Training: 2 epochs, lr 2e-5, batch 16, max_len 256, FP16 T4, 22min, val_loss 0.419->0.426.

Limitation: benchmark QCM examens, pas RAG documentaire réel. Shadow holdout (972 queries, locked) non utilisé pour ce checkpoint.

Parent: antoinelouis/crossencoder-distilcamembert-mmarcoFR