Update start.sh

start.sh

@@ -1,19 +1,49 @@
 #!/usr/bin/env bash
 set -euo pipefail
 
-MODEL_DIR="/app/bert_tf"
-mkdir -p "$MODEL_DIR"
+MODEL_ROOT="/app/bert_tf"
+mkdir -p "$MODEL_ROOT"
 
 if [ -z "${WEIGHTS_URL_TAR_GZ:-}" ]; then
-  echo "ERROR: Set WEIGHTS_URL_TAR_GZ in Space → Settings → Variables."
+  echo "ERROR: Set WEIGHTS_URL_TAR_GZ (direct .tar.gz link) in Settings → Variables."
   exit 1
 fi
 
 echo "[start] downloading model tarball…"
-curl -L "$WEIGHTS_URL_TAR_GZ" -o /tmp/bert_min.tar.gz
+curl -L "$WEIGHTS_URL_TAR_GZ" -o /tmp/model.tar.gz
 
-echo "[start] extracting…"
-tar -xzf /tmp/bert_min.tar.gz -C "$MODEL_DIR"
+echo "[start] extracting to $MODEL_ROOT…"
+tar -xzf /tmp/model.tar.gz -C "$MODEL_ROOT"
+
+# --- normalize layout ---
+# If there is a single subfolder and it has vocab.txt, use it as model dir
+SUB="$(find "$MODEL_ROOT" -mindepth 1 -maxdepth 1 -type d | head -n 1 || true)"
+if [ -n "${SUB:-}" ] && [ -f "$SUB/vocab.txt" ]; then
+  export MODEL_DIR="$SUB"
+else
+  export MODEL_DIR="$MODEL_ROOT"
+fi
+
+# Pull vocab up if it's inside a 'vocab' folder
+if [ -d "$MODEL_DIR/vocab" ] && [ -f "$MODEL_DIR/vocab/vocab.txt" ]; then
+  mv "$MODEL_DIR/vocab/vocab.txt" "$MODEL_DIR/vocab.txt"
+fi
+
+# Some dumps have checkpoint.txt → rename to checkpoint
+[ -f "$MODEL_DIR/checkpoint.txt" ] && mv "$MODEL_DIR/checkpoint.txt" "$MODEL_DIR/checkpoint"
+
+# Some dumps have bert_config.json → duplicate to config.json
+[ -f "$MODEL_DIR/bert_config.json" ] && cp "$MODEL_DIR/bert_config.json" "$MODEL_DIR/config.json"
+
+echo "[debug] contents of MODEL_DIR=$MODEL_DIR"
+ls -l "$MODEL_DIR" | sed 's/^/[debug] /'
+
+# sanity check
+for f in vocab.txt config.json checkpoint model.ckpt-150000.index model.ckpt-150000.data-00000-of-00001; do
+  if [ ! -f "$MODEL_DIR/$f" ]; then
+    echo "[fatal] missing $f in $MODEL_DIR"; exit 2
+  fi
+done
 
 echo "[start] launching API…"
 exec uvicorn app:app --host 0.0.0.0 --port "${PORT:-7860}"