Align Space with local HyperText config and R2-only persistence.

Switch to sectioned clinical-note labeling, drop baked-in seed DB in favor of R2 sync, and update docs for aggregated pre-annotations.

Dockerfile

@@ -10,11 +10,10 @@ COPY scripts/r2_storage.py /app/scripts/r2_storage.py
 COPY config/label_studio.xml /app/config/label_studio.xml
 
 RUN chmod +x /entrypoint.sh \
-    && mkdir -p /label-studio/data /label-studio/logs /app/seed/label-studio /app/scripts /app/config
+    && mkdir -p /label-studio/data /label-studio/logs /app/scripts /app/config
 
 ENV LABEL_STUDIO_BASE_DATA_DIR=/label-studio/data \
     LABEL_STUDIO_LOG_DIR=/label-studio/logs \
-    LABEL_STUDIO_SEED_DIR=/app/seed/label-studio \
     STORAGE_PERSISTENCE=1 \
     PYTHONPATH=/app/scripts \
     R2_SYNC_ON_START=1 \

README.md

@@ -13,7 +13,7 @@ app_port: 8080
 
 # AntioBioNLP — expert adjudication
 
-Label Studio Space for **human adjudication** of antibiotic decisions on the MIMIC sepsis pilot cohort, with **three overlapping LLM pre-annotation layers** (Qwen3.6-27B, Gemma-4-31B-it, Qwen3.6-35B-A3B) stored on **Cloudflare R2**.
+Label Studio Space for **human adjudication** of antibiotic decisions on the MIMIC sepsis pilot cohort, with **aggregated LLM pre-annotations** (majority vote across three models) on **Cloudflare R2**.
 
 On boot the Space **pulls your Label Studio database from R2** (project, tasks, user accounts) and **pushes** changes back on shutdown / every 5 minutes.
 
@@ -50,16 +50,15 @@ Optional:
 
 | Prefix | Contents |
 |--------|----------|
-| `tasks/` | `tasks.json`, `notes_for_annotation.parquet` |
-| `llm/<model>/` | `llm_annotations.jsonl` per model |
+| `tasks/` | `tasks.json`, `aggregated_annotations.jsonl`, `notes_for_annotation.parquet` |
 | `label-studio/` | SQLite DB, media, exports (synced from the Space) |
 
 ## Using the UI
 
 1. Log in with the **same email/password** you created locally.
 2. Open project **Antibiotic decision adjudication**.
-3. Use **Settings → Predictions** to toggle Qwen / Gemma / Qwen-35B layers.
-4. Span colors = decision type; region **meta** shows LLM name and rationale categories.
+3. Enable **aggregated (majority vote)** under **Settings → Predictions**.
+4. Span colors = decision type; region **meta** shows per-model votes and rationale categories.
 
 Public signup is disabled (`LABEL_STUDIO_DISABLE_SIGNUP_WITHOUT_LINK=true`). Invite collaborators from **Organization → Invite**.
 

config/label_studio.xml

@@ -1,21 +1,90 @@
 <View>
-  <Header value="Antibiotic decision adjudication (expert)"/>
-  <View style="display:flex; gap:12px; margin-bottom:8px; flex-wrap:wrap;">
-    <Text name="note_id" value="Note ID: $note_id"/>
-    <Text name="subject" value="Subject: $subject_id | HADM: $hadm_id"/>
+  <Style>
+    .labeling-clinical-layout {
+      display: flex;
+      gap: 1em;
+      align-items: stretch;
+      max-height: calc(100vh - 190px);
+      min-height: 560px;
+      font-family: Arial, sans-serif;
+    }
+
+    .labeling-label-panel {
+      width: 220px;
+      min-width: 220px;
+      flex-shrink: 0;
+      padding: 1em;
+      background: #f8fafc;
+      border: 1px solid #d6dde8;
+      border-radius: 4px;
+      position: sticky;
+      top: 0;
+      align-self: flex-start;
+      max-height: calc(100vh - 190px);
+      overflow-y: auto;
+    }
+
+    .labeling-note-panel {
+      flex: 1;
+      min-width: 0;
+      overflow-y: auto;
+      overflow-x: hidden;
+      padding: 0.75em 1em;
+      border: 1px solid #d6dde8;
+      border-radius: 4px;
+      background: #fff;
+    }
+
+    .clinical-note {
+      max-width: 1120px;
+      margin: 0 auto;
+      color: #111827;
+      background: #fff;
+      font-family: Arial, sans-serif;
+    }
+
+    .note-document {
+      min-width: 0;
+    }
+
+    .note-section {
+      margin: 0 0 22px;
+    }
+
+    .note-section-title {
+      font-size: 13px;
+      font-weight: 700;
+      text-transform: uppercase;
+      letter-spacing: 0.02em;
+      color: #1f2937;
+      padding-bottom: 5px;
+      margin-bottom: 8px;
+      border-bottom: 1px solid #d9dee8;
+    }
+
+    .note-body {
+      white-space: pre-wrap;
+      font-size: 14px;
+      line-height: 1.42;
+    }
+  </Style>
+
+  <View className="labeling-clinical-layout">
+    <View className="labeling-label-panel">
+      <Header value="Decision type" size="4"/>
+      <HyperTextLabels name="decision" toName="note" showInline="false">
+        <Label value="initiation" background="#2563eb"/>
+        <Label value="escalation" background="#f97316"/>
+        <Label value="de-escalation" background="#16a34a"/>
+        <Label value="discontinuation" background="#7c3aed"/>
+        <Label value="within_class_switching" background="#0891b2"/>
+        <Label value="other" background="#64748b"/>
+        <Label value="undetermined" background="#dc2626"/>
+      </HyperTextLabels>
+    </View>
+
+    <View className="labeling-note-panel">
+      <HyperText name="note" value="$note_html" inline="true" clickableLinks="false" saveTextResult="yes" granularity="word"/>
+    </View>
   </View>
-  <Header value="LLM pre-annotations" size="4"/>
-  <Text name="legend" value="Toggle prediction layers (Settings → Predictions) to compare models. Colors encode decision type; meta shows rationale categories per span."/>
-  <Text name="text" value="$text"/>
-  <Labels name="decision" toName="text" showInline="true">
-    <Label value="initiation" background="#2563eb"/>
-    <Label value="escalation" background="#dc2626"/>
-    <Label value="de-escalation" background="#059669"/>
-    <Label value="discontinuation" background="#7c3aed"/>
-    <Label value="within_class_switching" background="#d97706"/>
-    <Label value="other" background="#64748b"/>
-    <Label value="none" background="#94a3b8"/>
-  </Labels>
-  <Header value="Expert notes" size="4"/>
-  <TextArea name="expert_notes" toName="text" rows="3" placeholder="Quality flags, adjudication notes…" maxSubmissions="1"/>
 </View>

scripts/entrypoint.sh

@@ -3,30 +3,15 @@ set -euo pipefail
 
 DATA_DIR="${LABEL_STUDIO_BASE_DATA_DIR:-/label-studio/data}"
 LOG_DIR="${LABEL_STUDIO_LOG_DIR:-/label-studio/logs}"
-SEED_DIR="${LABEL_STUDIO_SEED_DIR:-/app/seed/label-studio}"
 mkdir -p "$DATA_DIR" "$LOG_DIR"
 
 export PYTHONPATH="/app/scripts:${PYTHONPATH:-}"
 
-# 1) Restore from R2 (primary persistence)
 if [[ "${R2_SYNC_ON_START:-1}" == "1" ]] && [[ -n "${R2_ACCESS_KEY_ID:-}" ]]; then
   python3 /app/scripts/sync_label_studio_r2.py pull --data-dir "$DATA_DIR" || true
 fi
 
-# 2) First boot: copy baked-in seed (project, tasks, user accounts)
-if [[ ! -f "$DATA_DIR/label_studio.sqlite3" ]] && [[ -d "$SEED_DIR" ]]; then
-  echo "Seeding Label Studio data from $SEED_DIR"
-  shopt -s dotglob nullglob
-  for item in "$SEED_DIR"/*; do
-    base="$(basename "$item")"
-    [[ "$base" == ".env" ]] && continue
-    cp -a "$item" "$DATA_DIR/"
-  done
-  shopt -u dotglob nullglob
-fi
-
-# SECRET_KEY: prefer Space secret over seed .env
-if [[ -n "${SECRET_KEY:-}" ]] && [[ -d "$DATA_DIR" ]]; then
+if [[ -n "${SECRET_KEY:-}" ]]; then
   printf 'SECRET_KEY=%s\n' "$SECRET_KEY" > "$DATA_DIR/.env"
 fi