Trim metric dropdown and shorten landing copy

Drop strict_accuracy / mc2 / accuracy_norm from the curated metric list
(unfamiliar to most users) and force-include "accuracy" so the default
value actually appears in the dropdown instead of getting filtered out
by the id=0 == unknown_metric_id collision.

Also remove the two filler sentences from the landing markdown — the
status line below the table already explains the modality filter.

app.py

@@ -30,19 +30,19 @@ print(f"Loaded recommender: {len(RECOMMENDER.model_names)} candidate models, "
 
 # Sort the dropdown choices for a sane UX.
 TASK_CHOICES = sorted(RECOMMENDER.task2id.keys(), key=lambda x: x.lower())
-# Metric vocab is huge (3k+) and noisy — restrict to the most common bare metric names.
+# Metric vocab is huge (9k+) and noisy — restrict to the most common bare metric names.
 COMMON_METRICS = [
     "accuracy", "f1", "exact_match", "rouge_l", "bleu", "mean_iou",
     "mean_average_precision", "top_1_accuracy", "top_5_accuracy",
     "perplexity", "wer", "auc", "spearman", "pearson", "mse", "rmse",
-    "mc2", "accuracy_norm", "strict_accuracy",
 ]
 # Keep only those actually present in the metric vocab (with loose alias matching).
+# "accuracy" is always kept — it's the canonical default; even when it resolves
+# to the unknown_metric_id sentinel the UNK path is a sensible catchall.
 METRIC_CHOICES = sorted(
-    {m for m in COMMON_METRICS if RECOMMENDER.resolve_metric(m) != RECOMMENDER.model.unknown_metric_id}
+    {m for m in COMMON_METRICS
+     if m == "accuracy" or RECOMMENDER.resolve_metric(m) != RECOMMENDER.model.unknown_metric_id}
 )
-if "accuracy" in COMMON_METRICS and not METRIC_CHOICES:
-    METRIC_CHOICES = COMMON_METRICS  # fallback
 
 
 EXAMPLE_DESCRIPTIONS = [
@@ -139,14 +139,7 @@ with gr.Blocks(title="ModelLens · Finding the Best Model for Your Task", theme=
         """
         # ModelLens: Finding the Best for Your Task from Myriads of Models
         Describe your dataset, pick a task type and a metric, and ModelLens returns
-        the top candidates from a pool of **47k+** HuggingFace models. Backed by the
-        `MLPMetricFull` checkpoint trained on the cleaned + expanded
-        `unified_augmented_v2` corpus.
-
-        Results are post-filtered by a modality sanity check so that e.g.
-        *Image Generation* won't surface text-only LLMs. The status line below
-        the table shows which capability your task requires and how many
-        candidates passed the filter.
+        the top candidates from a pool of **47k+** HuggingFace models.
 
         > **BYO OpenAI key.** This Space embeds your dataset description with
         > `text-embedding-3-small`.