ProgressLM: bypass broken prompts/__init__, inline score parser

vanilla_baseline_code/Experiment/RMBench/Vanilla_Baseline/ProgressLM/run_prediction.py

@@ -68,12 +68,39 @@ if "CUDA_VISIBLE_DEVICES" not in os.environ:
         except Exception:
             os.environ["CUDA_VISIBLE_DEVICES"] = "0"
 
-sys.path.insert(0, str(Path(ARGS.progresslm_code).resolve()))
+_CODE = str(Path(ARGS.progresslm_code).resolve())
+sys.path.insert(0, _CODE)
+import re
 import numpy as np
 from PIL import Image
+import importlib.util as _ilu
+
+# core is a healthy package -> normal import
 from core.model import Qwen2VLChat
-from prompts.visual_demo_prompt import build_visual_demo_prompt_from_item, VISUAL_DEMO_SYSTEM_PROMPT
-from run_visual_demo_single import parse_visual_demo_response
+# prompts/__init__ imports deleted modules and crashes -> load the file directly
+_spec = _ilu.spec_from_file_location("plm_vdp", os.path.join(_CODE, "prompts/visual_demo_prompt.py"))
+_vdp = _ilu.module_from_spec(_spec); sys.modules["plm_vdp"] = _vdp; _spec.loader.exec_module(_vdp)
+build_visual_demo_prompt_from_item = _vdp.build_visual_demo_prompt_from_item
+VISUAL_DEMO_SYSTEM_PROMPT = _vdp.VISUAL_DEMO_SYSTEM_PROMPT
+
+
+def parse_visual_demo_response(response: str):
+    """Extract <score> as 0..1 (or 'n/a'/None). Mirrors ProgressLM's parser."""
+    if not response:
+        return {"score": None}
+    m = re.search(r"<score>(.*?)</score>", response, re.DOTALL)
+    if not m:
+        return {"score": None}
+    s = m.group(1).strip()
+    if s.lower() in ("n/a", "na"):
+        return {"score": "n/a"}
+    try:
+        v = float(s[:-1]) / 100.0 if s.endswith("%") else float(s)
+        if v > 1.0:
+            v = v / 100.0
+        return {"score": max(0.0, min(1.0, v))}
+    except ValueError:
+        return {"score": None}
 
 OUT = Path(ARGS.out_root)
 REFS = OUT / "progresslm_refs"