Update src/streamlit_app.py

src/streamlit_app.py

@@ -358,13 +358,23 @@ st.markdown("""
             border-color: rgba(255, 255, 255, 0.1);
         }
     }
-    /* Model display checkboxes: force horizontal row on mobile */
+    /* Model display checkboxes: compact row */
     .st-key-narrow_model_display [data-testid="stHorizontalBlock"] {
         flex-wrap: nowrap !important;
     }
     .st-key-narrow_model_display [data-testid="column"] {
         min-width: 0 !important;
     }
+    /* Mobile: allow wrapping so checkboxes don't overflow */
+    @media (max-width: 768px) {
+        .st-key-narrow_model_display [data-testid="stHorizontalBlock"] {
+            flex-wrap: wrap !important;
+        }
+    }
+    /* Failure mode pills: no word breaking */
+    .st-key-failure_mode_pills button {
+        white-space: nowrap !important;
+    }
     /* Mobile: stack image comparison columns vertically */
     @media (max-width: 768px) {
         .st-key-image_comparison [data-testid="stHorizontalBlock"] {
@@ -387,8 +397,17 @@ def _parse_success(value):
     return False
 
 
+def _csv_mtime():
+    """Return CSV modification time so cache invalidates when file changes."""
+    for base in (_repo_root, _script_dir):
+        candidate = base / "data" / "baseline_results_full_new.csv"
+        if candidate.exists():
+            return candidate.stat().st_mtime
+    return None
+
+
 @st.cache_data
-def load_data():
+def load_data(_mtime=None):
     """Load and clean data. Tries repo root (HF Space: /app/data/) then script dir."""
     csv_path = None
     for base in (_repo_root, _script_dir):
@@ -823,8 +842,11 @@ FAILURE_MODE_OPTIONS = [
     "Click Region Error",
     "Visual Confusion",
     "Spatial Reasoning Error",
-    "Multiple targets",
-    "Ambiguous instruction",
+    "Text Matching Bias",
+    "Instruction Misinterpretation",
+    "Goal Hallucination",
+    "Location Hallucination",
+    "Reasoning Drift",
 ]
 
 
@@ -857,11 +879,7 @@ def _build_available_samples(df_filtered, selected_variant, failure_mode="All"):
     if failure_mode != "All":
         # Get one row per sample to check failure mode (use original variant)
         df_orig = df_rel[df_rel["variant"] == "original"].drop_duplicates(subset=["task_id", "step_index"])
-        if failure_mode in ("Multiple targets", "Ambiguous instruction"):
-            # These are sub-categories of "Invalid" stored in the description column
-            match_mask = df_orig["description"].str.lower() == failure_mode.lower()
-        else:
-            match_mask = df_orig["interesting_cases"] == failure_mode
+        match_mask = df_orig["interesting_cases"] == failure_mode
         match_pairs = set(zip(df_orig.loc[match_mask, "task_id"], df_orig.loc[match_mask, "step_index"]))
         filtered_pairs &= match_pairs
 
@@ -919,7 +937,7 @@ def _render_compact_header():
     html = f"""
     <div class="gui-viewer-compact-header">
       <h2 style="margin:0;color:var(--gui-viewer-heading);font-size:1.3rem;font-weight:700;">GUI-Perturbed Baseline Result Viewer</h2>
-      <p style="margin:0.25rem 0 0.4rem;color:var(--gui-viewer-muted);font-size:0.85rem;line-height:1.4;">a baseline study of 7B GUI models using GUI-Perturbed data to stress test CUA models on GUI and understand what agents fail at and why</p>
+      <p style="margin:0.25rem 0 0.4rem;color:var(--gui-viewer-muted);font-size:0.85rem;line-height:1.4;">Explore how 7B GUI grounding models perform on original vs. perturbed screenshots from GUI-Perturbed</p>
       <div class="header-row">
         {fig_logo_html}
         {manifold_logo_html}
@@ -1110,7 +1128,7 @@ def main():
 
     # --- Load data ---
     with st.spinner("Loading results..."):
-        df = load_data()
+        df = load_data(_mtime=_csv_mtime())
     _t = _lap("load_data", _t)
     if df.empty:
         st.error("No data found")
@@ -1125,7 +1143,10 @@ def main():
     # --- Compute state from session defaults (widgets rendered later read from previous rerun) ---
     query_types = sorted(df['query_type'].unique().tolist())
     use_reasoning_options = sorted(df['use_reasoning'].unique().tolist())
-    selected_query_type = st.session_state.get("query_type_filter", query_types[0] if query_types else None)
+    _default_query_type = "relational_query" if "relational_query" in query_types else (query_types[0] if query_types else None)
+    if "query_type_filter" not in st.session_state and _default_query_type is not None:
+        st.session_state.query_type_filter = _default_query_type
+    selected_query_type = st.session_state.get("query_type_filter", _default_query_type)
     selected_use_reasoning = st.session_state.get("use_reasoning_filter", use_reasoning_options[0] if use_reasoning_options else None)
     if selected_query_type not in query_types:
         selected_query_type = query_types[0] if query_types else None
@@ -1155,8 +1176,8 @@ def main():
     # to avoid 1-rerun lag vs the manually maintained selected_models dict
     selected_models = [m for m in all_models if st.session_state.get(f"model_{m}", True)]
 
-    if 'failure_mode_filter' not in st.session_state:
-        st.session_state.failure_mode_filter = "All"
+    if 'failure_mode_pills' not in st.session_state:
+        st.session_state.failure_mode_pills = "Spatial Reasoning Error"
 
     # Initialize navigation state
     if 'current_sample_index' not in st.session_state:
@@ -1171,16 +1192,20 @@ def main():
         st.session_state.sample_nav_input = st.session_state.current_sample_index + 1
 
     # Build available samples
+    _active_failure_mode = st.session_state.get("failure_mode_pills", "All")
+    if _active_failure_mode is None:
+        _active_failure_mode = "All"
     available_samples, available_samples_all, full_list_index_by_sample = _build_available_samples(
         df_filtered,
         st.session_state.selected_variant,
-        st.session_state.failure_mode_filter,
+        _active_failure_mode,
     )
     _t = _lap("_build_available_samples", _t)
 
     if not available_samples:
         st.error(
-            f"No samples found with both original and {st.session_state.selected_variant} perturbation"
+            f"No samples found with both original and {st.session_state.selected_variant} perturbation "
+            f"for failure mode \"{_active_failure_mode}\""
         )
         return
 
@@ -1189,7 +1214,7 @@ def main():
         selected_query_type,
         selected_use_reasoning,
         st.session_state.selected_variant,
-        st.session_state.failure_mode_filter,
+        _active_failure_mode,
     )
     filters_changed = st.session_state.get("_filter_signature") != current_filter_signature
     if filters_changed:
@@ -1231,7 +1256,7 @@ def main():
     _t = _lap("build rows by model", _t)
 
     # ==========================================
-    # RENDER: Images first (hero content)
+    # RENDER: Images (hero content)
     # ==========================================
     _render_images(
         original_rows_by_model,
@@ -1243,19 +1268,44 @@ def main():
     _t = _lap("render images", _t)
 
     # ==========================================
-    # RENDER: Model display checkboxes (under images)
+    # RENDER: Sample navigation (right under images)
+    # ==========================================
+    def on_sample_change():
+        new_val = st.session_state.sample_nav_input
+        if new_val - 1 != st.session_state.current_sample_index:
+            st.session_state.current_sample_index = new_val - 1
+
+    with st.container(key="narrow_sample_nav"):
+        position_in_full_list = full_list_index_by_sample.get(
+            (current_sample['task_id'], current_sample['step_index'])
+        )
+        total_in_full_list = len(available_samples_all)
+        nav_label = f"Sample ({st.session_state.current_sample_index + 1} of {len(available_samples)}"
+        if position_in_full_list is not None and total_in_full_list != len(available_samples):
+            nav_label += f" | {position_in_full_list} of {total_in_full_list} total"
+        nav_label += ")"
+        st.number_input(
+            nav_label,
+            min_value=1,
+            max_value=len(available_samples),
+            key="sample_nav_input",
+            on_change=on_sample_change,
+        )
+
+    # ==========================================
+    # RENDER: Model display checkboxes (collapsible, default collapsed)
     # ==========================================
     with st.container(key="narrow_model_display"):
-        st.markdown("**Model Display**")
-        model_display_cols = st.columns(len(all_models)) if all_models else []
-        for i, model in enumerate(all_models):
-            with model_display_cols[i]:
-                checked = st.checkbox(
-                    _model_label(model),
-                    value=st.session_state.selected_models.get(model, True),
-                    key=f"model_{model}",
-                )
-                st.session_state.selected_models[model] = checked
+        with st.expander("Model Display", expanded=False):
+            model_display_cols = st.columns(len(all_models)) if all_models else []
+            for i, model in enumerate(all_models):
+                with model_display_cols[i]:
+                    checked = st.checkbox(
+                        _model_label(model),
+                        value=st.session_state.selected_models.get(model, True),
+                        key=f"model_{model}",
+                    )
+                    st.session_state.selected_models[model] = checked
 
     # Recompute selected_models from current checkbox widget state
     selected_models = [m for m in all_models if st.session_state.get(f"model_{m}", True)]
@@ -1272,61 +1322,21 @@ def main():
     _t = _lap("render model results", _t)
 
     # ==========================================
-    # RENDER: Navigation & Filters
+    # RENDER: Filters & Search
     # ==========================================
     with st.container(key="narrow_controls"):
         st.markdown("---")
-        st.markdown("#### Navigation & Filters")
-
-        # Navigation callbacks (use -/+ on number input for prev/next)
-        def on_sample_change():
-            new_val = st.session_state.sample_nav_input
-            if new_val - 1 != st.session_state.current_sample_index:
-                st.session_state.current_sample_index = new_val - 1
-
-        def on_search():
-            query = st.session_state.get("instruction_search", "").strip().lower()
-            if not query:
-                return
-            # Search from current position + 1, wrapping around
-            n = len(available_samples)
-            start = (st.session_state.current_sample_index + 1) % n
-            for offset in range(n):
-                idx = (start + offset) % n
-                instr = available_samples[idx].get("instruction", "").lower()
-                if query in instr:
-                    st.session_state.current_sample_index = idx
-                    st.session_state.sample_nav_input = idx + 1
-                    return
+        st.markdown("#### Filters & Search")
 
-        # Row 1: Sample number + search
-        nav_col, search_col = st.columns([1, 2])
-        with nav_col:
-            position_in_full_list = full_list_index_by_sample.get(
-                (current_sample['task_id'], current_sample['step_index'])
-            )
-            total_in_full_list = len(available_samples_all)
-            nav_label = f"Sample ({st.session_state.current_sample_index + 1} of {len(available_samples)}"
-            if position_in_full_list is not None and total_in_full_list != len(available_samples):
-                nav_label += f" | {position_in_full_list} of {total_in_full_list} total"
-            nav_label += ")"
-            st.number_input(
-                nav_label,
-                min_value=1,
-                max_value=len(available_samples),
-                key="sample_nav_input",
-                on_change=on_sample_change,
-            )
-        with search_col:
-            st.text_input(
-                "Search instructions",
-                key="instruction_search",
-                on_change=on_search,
-                placeholder="Type to search task instructions...",
-            )
+        # Failure mode pills (key pre-initialized in session state)
+        st.pills(
+            "Failure Mode",
+            FAILURE_MODE_OPTIONS,
+            key="failure_mode_pills",
+        )
 
-        # Row 2: Filters
-        f = st.columns(4)
+        # Filter dropdowns
+        f = st.columns(3)
         with f[0]:
             new_variant = st.selectbox(
                 "Visual Variant",
@@ -1337,7 +1347,7 @@ def main():
                 help="Precision: viewport zoom. Style: visual randomization. Text Shrink: font size reduced.",
             )
         with f[1]:
-            selected_query_type = st.selectbox(
+            st.selectbox(
                 "Instruction Variant",
                 query_types,
                 key="query_type_filter",
@@ -1345,33 +1355,43 @@ def main():
                 help="Direct Instruction vs Relational Instruction",
             )
         with f[2]:
-            selected_use_reasoning = st.selectbox(
+            st.selectbox(
                 "Reasoning",
                 use_reasoning_options,
                 key="use_reasoning_filter",
                 format_func=lambda x: "Yes" if x else "No",
                 help="Whether chain-of-thought reasoning was used",
             )
-        with f[3]:
-            new_failure_mode = st.selectbox(
-                "Failure Mode",
-                FAILURE_MODE_OPTIONS,
-                index=FAILURE_MODE_OPTIONS.index(st.session_state.failure_mode_filter),
-                key="failure_mode_select",
-                help="Filter samples by failure mode category",
-            )
+
+        # Search instructions
+        def on_search():
+            query = st.session_state.get("instruction_search", "").strip().lower()
+            if not query:
+                return
+            n = len(available_samples)
+            if n == 0:
+                return
+            start = (st.session_state.current_sample_index + 1) % n
+            for offset in range(n):
+                idx = (start + offset) % n
+                instr = available_samples[idx].get("instruction", "").lower()
+                if query in instr:
+                    st.session_state.current_sample_index = idx
+                    st.session_state.sample_nav_input = idx + 1
+                    return
+
+        st.text_input(
+            "Search instructions",
+            key="instruction_search",
+            on_change=on_search,
+            placeholder="Type to search task instructions...",
+        )
 
     _t = _lap("control panel widgets", _t)
 
-    # Handle variant or failure mode change
-    needs_rerun = False
+    # Handle variant change
     if new_variant != st.session_state.selected_variant:
         st.session_state.selected_variant = new_variant
-        needs_rerun = True
-    if new_failure_mode != st.session_state.failure_mode_filter:
-        st.session_state.failure_mode_filter = new_failure_mode
-        needs_rerun = True
-    if needs_rerun:
         st.rerun()
 
     _lap("main() total", _t0)