camembert-pharmacy-reranker-distil

French pharmacy / drug CrossEncoder reranker (~68M DistilCamemBERT). This is not a universal medical reranker and not a SOTA claim.

• HF repo: ymelka/camembert-pharmacy-reranker-distil
• Release tag: exp-037-step-01948
• Weights SHA256 (model.safetensors): 815637b9664e85d9196e0687e3befd07727bc221e1b45eb71d1f649a83ed4051

Architecture

• DistilCamemBERT CrossEncoder, ~68M parameters (CamembertForSequenceClassification, num_labels=1)
• max_length = 256 at train and eval
• CPU-friendly (same cost class as the parent)

Parent

antoinelouis/crossencoder-distilcamembert-mmarcoFR
revision = 9447ae56ace29839d8bf24f740abdf48ab0a7690
license = MIT

Parent training data: French mMARCO (unicamp-dl/mmarco). Backbone: cmarkea/distilcamembert-base.

Task

Rerank French pharmacy specialty cards (BDPM) for queries such as substitution (générique de …), dosage, formulation, route, population (enfant/adulte), and same-molecule wrong variant.

Training (exp-037)

experiment     = exp-037-v3-pharmacy-distil-mixture-v5
data           = pharmacy-mixture-v5 (~14.7k query groups, public BDPM only)
loss           = pairwise
epochs         = 1
lr             = 2e-6
freeze_layers  = 4 (bottom encoder layers)
max_length     = 256
selected ckpt  = step-01948

No private Wellchat text. No PARHAF / medical-QCM mix in this recipe.

How to use

from transformers import AutoModelForSequenceClassification, AutoTokenizer
import torch

repo = "ymelka/camembert-pharmacy-reranker-distil"
rev = "exp-037-step-01948"
tok = AutoTokenizer.from_pretrained(repo, revision=rev)
model = AutoModelForSequenceClassification.from_pretrained(repo, revision=rev)
model.eval()

pairs = [
    ("générique de Doliprane 500 mg", "PARACETAMOL XYZ 500 mg, comprimé"),
    ("générique de Doliprane 500 mg", "IBUPROFENE ABC 400 mg, comprimé"),
]
enc = tok(pairs, padding=True, truncation=True, max_length=256, return_tensors="pt")
with torch.no_grad():
    scores = model(**enc).logits.squeeze(-1)
print(scores)

Or sentence_transformers.CrossEncoder(repo, revision=rev, max_length=256).

Benchmarks

Frozen BDPM evaluator (RAW_EXACT_GOLD + SEMANTIC_MULTI_POSITIVE). Holdout opened once after preregistration.

split	metric	parent Distil	this model
pharmacy_dev N=627	RAW P@1	0.761	0.797
pharmacy_dev	SEMANTIC P@1	0.809	0.884
pharmacy_dev substitution	SEMANTIC P@1	0.449	0.936
pharmacy_final_holdout N=157	RAW P@1	0.834	0.873
pharmacy_final_holdout	SEMANTIC P@1	0.860	0.917
SyntecReranking	MAP	0.784	0.792
AlloprofReranking	MAP@1000	0.5810	0.5811
MIRACL-fr	nDCG@10	0.469	0.447

Holdout SEMANTIC paired bootstrap 10k, seed 42: Δ +0.057, RESCUE 14 / DESTROY 5, 95% CI [0.006, 0.115] excludes 0.

Provenance / licenses

artefact	license / source
this fine-tune (weights)	MIT (inherits parent)
parent CrossEncoder	MIT — Antoine Louis, mMARCO-FR
DistilCamemBERT	MIT — cmarkea/distilcamembert-base
pharmacy-mixture-v5 / pharmacy-ranking-v1	BDPM dump 2026-08-31, Licence Ouverte / CADA — cite base-donnees-publique.medicaments.gouv.fr

No Wellchat private queries, traces, or MAX_DATA.

Limitations (read these)

• Holdout is an unseen query / unseen slate holdout, not an unseen-drug entity holdout (cis_overlap_dev_holdout = 280/361).
• Generic-group overlap exists between train and holdout (69/92 CIS_GENER groups).
• Contraindication, interactions, pregnancy, renal, hepatic: not validated (tiny or empty N).
• Some benchmark axes are lexical shortcuts (indication, adverse_effect, parts of drug_identity).
• MIRACL-fr Δ ≈ −0.022 nDCG@10 vs parent (not an E2-style collapse; Syntec/Alloprof held).
• Not a universal French medical reranker.
• Not a SOTA claim.

Independent qrel audit: iux2019/wellchat PR #55 (VALID_WITH_QREL_BUGS). Main public metric freeze: BDPM substitution siblings + population side. Adversarial overlay additionally promotes some formulation/route same-substance other-dose positives.

Intended use

Rerank a short candidate slate of French specialty cards. Keep a first-stage retriever. Do not use as a clinical decision system.