--- license: mit language: fr task_categories: [token-classification] tags: [gliner, biomedical, french, lymphoma, ner, ecrf, synthetic, span-localized] --- # lymphome-synth-v5 — gold TAPDL span-localisé Jeu d'**évaluation** (NER open-vocab / extraction eCRF) : **2050 dossiers cliniques lymphome synthétiques** multi-comptes-rendus (~89 champs eCRF), gold **span-localisé**. Split déterministe **seed 0 : 1640 train / 410 test**. ## Quoi de neuf vs `rntc/lymphome-synth-v4` Le gold v4 était une **re-extraction LLM libre** depuis les CR → spans partagés entre champs (~26%), valeurs hors de leur span (~62%), role-binding (1L/2L) deviné → **plafond mécanique** de l'éval span-overlap (un span tagué à K champs ne peut matcher qu'1 prédiction). **v5 = reconstruction TAPDL** (*Truth-Anchored Per-Document Localization*) : - on **ancre sur la timeline de référence** (vérité structurée du pipeline de génération) et on **localise par compte-rendu** : chaque champ est routé à son CR daté via `focus_event_indices` → **role-binding structurel** (le CR fixe la ligne, pas de devinette) ; - extraction per-CR (Qwen3-235B, guided JSON) de `{champ, span}` puis **validation programmatique** : span = **sous-chaîne exacte** (100%), **anti-partage** (1 span ↔ 1 champ garanti), dédup. - Un span peut **impliquer** la valeur sans la contenir (« patiente »→sexe, « Marseille »→habitat) : c'est volontaire et aligné sur le fonctionnement de GLiNER (span-overlap). **Statut** : *silver-as-gold* (annotateur LLM), destiné au travail avant une annotation humaine. ## Stats v5 - 2050 dossiers (1640 train / 410 test) ; **~237k spans gold** ; **89/89 champs** couverts. - Gates : **substring exacte 100%**, **0 offset partagé entre champs**. ## Format JSONL, une ligne = un dossier : ```json {"id": "lym_8", "sample_index": 8, "split": "test", "text": "", "gold": {"demography_sexe": [[start, end, "patiente"]], ...}, "field_labels": {"demography_sexe": {"label": "sexe", "desc": "...", "section": "demographie"}}, "timeline_json": "...", "documents_json": "...", "cohort_label": "LBDGC", "annotation_v4_json": ""} ``` - `gold` : `{clé_eCRF: [[start_char, end_char, texte_du_span], ...]}` (offsets dans `text`). - `timeline_json` / `documents_json` : provenance (vérité structurée + CR), pour audit/regénération. - `annotation_v4_json` : l'ancien gold v4, conservé pour comparaison. ## Méthode Voir le code : `gen/lymphome/tapdl/` (route → annotate → assemble) et `GOLD_V2_DESIGN.md` dans [github.com/Rian-T/french-medgliner](https://github.com/Rian-T/french-medgliner). MIT, recherche.