diff --git a/internal/cli/interactive.go b/internal/cli/interactive.go
index fe357947..5f28b277 100644
--- a/internal/cli/interactive.go
+++ b/internal/cli/interactive.go
@@ -141,7 +141,12 @@ func runInteractive(p *ui.Printer, pr prompter, a *runDataIngestArgs, taskSet bo
 		if err != nil {
 			return err
 		}
-		a.LocalPath = ans
+		// Trim before storing: validateDatasetPath only trims to check for
+		// emptiness, so a pasted " ~/data" (stray leading/trailing space)
+		// would otherwise survive here and defeat expandHome (first char
+		// isn't '~') — filepath.Abs then prepends cwd and the sniff / label
+		// preview read a path that doesn't exist.
+		a.LocalPath = strings.TrimSpace(ans)
 		prompted = true
 	}
 	// Expand a leading ~ now so the family sniff + label-header preview read
@@ -371,10 +376,9 @@ func defaultLabelChoice(headers []string) string {
 			return h
 		}
 	}
-	if len(headers) > 0 {
-		return headers[0]
-	}
-	return ""
+	// The only caller (promptLabelColumn) guards len(headers) > 0 before
+	// calling, so headers is never empty here.
+	return headers[0]
 }
 
 // renderReview prints the assembled ingest inputs before the confirm
diff --git a/internal/cli/interactive_test.go b/internal/cli/interactive_test.go
index 3cdfe95c..529ed740 100644
--- a/internal/cli/interactive_test.go
+++ b/internal/cli/interactive_test.go
@@ -461,6 +461,32 @@ func TestRunInteractive_RejectsEmptyPath(t *testing.T) {
 	}
 }
 
+// TestRunInteractive_TrimsPath: a path answer with surrounding whitespace
+// (a common paste artifact) is trimmed before it's stored, so expandHome
+// and the family sniff read the real path rather than a cwd-prefixed
+// mangle. Without the trim, " <dir>" defeats expandHome and the sniff
+// would land in the wrong place.
+func TestRunInteractive_TrimsPath(t *testing.T) {
+	dir := tabularDir(t)
+	f := &fakePrompter{answers: map[string]string{
+		"What should we call this dataset?":           "t",
+		"Where is your data? (the folder holding it)": "  " + dir + "  ",
+		"Which column holds the class?":               "churned",
+	}}
+	a := &runDataIngestArgs{Spec: push.SpecArgs{Intent: "train"}}
+	if err := runInteractive(discardPrinter(), f, a, false); err != nil {
+		t.Fatalf("runInteractive: %v", err)
+	}
+	if a.LocalPath != dir {
+		t.Errorf("LocalPath = %q, want %q (surrounding whitespace not trimmed)", a.LocalPath, dir)
+	}
+	// The trimmed path must have sniffed cleanly as tabular (not landed in a
+	// cwd-prefixed nonexistent dir that would force the family question).
+	if a.Spec.Category != "tabular_classification" {
+		t.Errorf("Category = %q, want tabular_classification (sniff read the trimmed path)", a.Spec.Category)
+	}
+}
+
 // TestRunInteractive_ShowsExampleHints: the name and path prompts carry a
 // visible example, so the guided flow teaches as it goes.
 func TestRunInteractive_ShowsExampleHints(t *testing.T) {
diff --git a/internal/push/category.go b/internal/push/category.go
index e22aebc2..9edec1f8 100644
--- a/internal/push/category.go
+++ b/internal/push/category.go
@@ -35,6 +35,13 @@ type CategorySpec struct {
 	// therefore need label.policy (object label form) so the raw target
 	// never ships to the central backend by default.
 	RegressionClass bool
+	// SelfSupervised marks text categories that train without an explicit
+	// label column — the target is derived from the text itself (MLM masks
+	// tokens; CLM predicts the next token), so the interactive flow skips
+	// the "which column is the label?" question. A registry fact rather
+	// than a hardcoded id list so a new self-supervised task can't be added
+	// without deciding this (SelfSupervisedText reads it).
+	SelfSupervised bool
 	// CLISupported reports whether `dataset push` implements the category
 	// today. semantic_segmentation is known (the schema defines it) but
 	// not yet pushable.
@@ -77,7 +84,7 @@ var categoryRegistry = []CategorySpec{
 		Blurb: "locate landmark points on an image (e.g. pose)"},
 	{ID: "text_classification", Family: FamilyText, Label: "Text classification", CLISupported: true,
 		Blurb: "sort text snippets into classes"},
-	{ID: "masked_language_modeling", Family: FamilyText, Label: "Masked language modeling", Gloss: "fill-mask", CLISupported: true,
+	{ID: "masked_language_modeling", Family: FamilyText, Label: "Masked language modeling", Gloss: "fill-mask", CLISupported: true, SelfSupervised: true,
 		Blurb: "predict masked-out words — no labels needed"},
 	{ID: "tabular_classification", Family: FamilyTabular, Label: "Tabular classification", CLISupported: true,
 		Blurb: "predict a class from table columns"},
@@ -90,7 +97,7 @@ var categoryRegistry = []CategorySpec{
 	{ID: "semantic_segmentation", Family: FamilyImage, Label: "Semantic segmentation", CLISupported: false,
 		Blurb:           "label every pixel in an image",
 		UnsupportedNote: "blocked on the ingestor's mask-sidecar support (data-ingestors#136)"},
-	{ID: "causal_language_modeling", Family: FamilyText, Label: "Causal language modeling", CLISupported: false,
+	{ID: "causal_language_modeling", Family: FamilyText, Label: "Causal language modeling", CLISupported: false, SelfSupervised: true,
 		Blurb:           "predict the next word in a sequence",
 		UnsupportedNote: "schema-recognized (data-ingestors#805); `tracebloc ingest` discover/build for its raw-.txt / prompt\\tcompletion `texts` layout is pending"},
 	{ID: "seq2seq", Family: FamilyText, Label: "Sequence-to-sequence", Gloss: "translation / summarization", CLISupported: false,
@@ -201,9 +208,13 @@ func FamilyNouns() []string {
 // SelfSupervisedText reports whether a text category trains without an
 // explicit label column — the target is derived from the text itself, so
 // the CLI skips the "which column is the label?" question. MLM masks
-// tokens; CLM predicts the next token; neither reads a labels column.
+// tokens; CLM predicts the next token; neither reads a labels column. The
+// answer is the registry's SelfSupervised flag, so a new self-supervised
+// task is handled the moment it's added to the registry — not when someone
+// remembers to edit this function.
 func SelfSupervisedText(category string) bool {
-	return category == "masked_language_modeling" || category == "causal_language_modeling"
+	c, ok := categoryByID[category]
+	return ok && c.SelfSupervised
 }
 
 // IsKnown reports whether category is a recognized task category (in the
diff --git a/internal/push/preview.go b/internal/push/preview.go
index bf195e7e..47c05d1e 100644
--- a/internal/push/preview.go
+++ b/internal/push/preview.go
@@ -117,8 +117,15 @@ func SniffFamily(path string) FamilySniff {
 		if hasSequences {
 			dir = "sequences/"
 		}
+		// "looks like", not "is": the family is confident, but texts/ and
+		// sequences/ map to DIFFERENT tasks (DiscoverText keys the sidecar
+		// dir off TextSidecarDir — texts/ for classification, sequences/ for
+		// MLM). So a confident text sniff must not imply the task the user
+		// then picks will load; the picker offers the whole text family, and
+		// the walk gives the authoritative error if the layout and the
+		// chosen task disagree.
 		return FamilySniff{Family: FamilyText, Confident: true,
-			Echo: fmt.Sprintf("Found labels.csv and a %s folder — this is text data.", dir)}
+			Echo: fmt.Sprintf("Found labels.csv and a %s folder — this looks like text data.", dir)}
 	case !hasImages && !hasText && csvCount == 1:
 		// Exactly one CSV, mirroring DiscoverTabular's findSingleCSV rule.
 		// Two or more CSVs is a directory the tabular walk rejects, so stay