diff --git a/studio/backend/core/inference/llama_cpp.py b/studio/backend/core/inference/llama_cpp.py
index 37a57e4c39..dac8d380c3 100644
--- a/studio/backend/core/inference/llama_cpp.py
+++ b/studio/backend/core/inference/llama_cpp.py
@@ -249,17 +249,42 @@ class LlamaCppBackend:
                     )
 
                 # Determine the filename from the variant (e.g., "Q4_K_M" -> find matching file)
+                # For split GGUFs (e.g., *-00001-of-00003.gguf) we must download ALL shards.
                 gguf_filename = None
+                gguf_extra_shards: list[str] = []
                 if hf_variant:
                     # Try common naming patterns
                     try:
+                        import re
                         from huggingface_hub import list_repo_files
                         files = list_repo_files(hf_repo, token=hf_token)
                         variant_lower = hf_variant.lower()
-                        for f in files:
-                            if f.endswith(".gguf") and variant_lower in f.lower():
-                                gguf_filename = f
-                                break
+                        # Use word-boundary matching so "Q8_0" doesn't also
+                        # match "IQ8_0" or other superset variant names.
+                        boundary = re.compile(
+                            r'(?<![a-zA-Z0-9])' + re.escape(variant_lower) + r'(?![a-zA-Z0-9])'
+                        )
+                        gguf_files = sorted(
+                            f for f in files
+                            if f.endswith(".gguf") and boundary.search(f.lower())
+                        )
+                        if gguf_files:
+                            gguf_filename = gguf_files[0]
+                            # For split GGUFs (e.g. model-Q8_0-00001-of-00003.gguf)
+                            # discover siblings by exact basename + total match
+                            # so "model-Q8_0-v2-*" isn't pulled in as a sibling.
+                            shard_pat = re.compile(r'^(.*)-\d{5}-of-(\d{5})\.gguf$')
+                            m = shard_pat.match(gguf_filename)
+                            if m:
+                                prefix = m.group(1)
+                                total = m.group(2)
+                                sibling_pat = re.compile(
+                                    r'^' + re.escape(prefix) + r'-\d{5}-of-' + re.escape(total) + r'\.gguf$'
+                                )
+                                gguf_extra_shards = [
+                                    f for f in gguf_files[1:]
+                                    if sibling_pat.match(f)
+                                ]
                     except Exception as e:
                         logger.warning(f"Could not list repo files: {e}")
 
@@ -269,13 +294,23 @@ class LlamaCppBackend:
                         repo_name = hf_repo.split("/")[-1].replace("-GGUF", "")
                         gguf_filename = f"{repo_name}-{hf_variant}.gguf"
 
-                logger.info(f"Downloading GGUF: {hf_repo}/{gguf_filename}")
+                logger.info(f"Downloading GGUF: {hf_repo}/{gguf_filename}"
+                            + (f" (+{len(gguf_extra_shards)} shards)" if gguf_extra_shards else ""))
                 try:
                     local_path = hf_hub_download(
                         repo_id=hf_repo,
                         filename=gguf_filename,
                         token=hf_token,
                     )
+                    # Download remaining shards for split GGUFs — llama-server
+                    # auto-discovers them when they are in the same directory.
+                    for shard in gguf_extra_shards:
+                        logger.info(f"Downloading GGUF shard: {shard}")
+                        hf_hub_download(
+                            repo_id=hf_repo,
+                            filename=shard,
+                            token=hf_token,
+                        )
                 except Exception as e:
                     raise RuntimeError(
                         f"Failed to download GGUF file '{gguf_filename}' from {hf_repo}: {e}"
diff --git a/studio/backend/core/training/trainer.py b/studio/backend/core/training/trainer.py
index 31e05d0bda..2e37cad189 100644
--- a/studio/backend/core/training/trainer.py
+++ b/studio/backend/core/training/trainer.py
@@ -1866,10 +1866,34 @@ class UnslothTrainer:
 
             elif dataset_source:
                 # Load from Hugging Face
-                load_kwargs = {"path": dataset_source, "split": train_split or "train"}
+                split_name = train_split or "train"
+                load_kwargs = {"path": dataset_source, "split": split_name}
                 if subset:
                     load_kwargs["name"] = subset
-                dataset = load_dataset(**load_kwargs)
+
+                _slice_start = dataset_slice_start or 0
+                if (dataset_slice_end is not None
+                        and dataset_slice_end >= 0
+                        and dataset_slice_end >= _slice_start):
+                    # Manual slice — stream only the rows we need instead of
+                    # downloading the entire dataset.
+                    rows_to_stream = dataset_slice_end + 1
+                    print(
+                        f"[dataset-slice] Manual slice specified "
+                        f"(start={dataset_slice_start}, end={dataset_slice_end}), "
+                        f"streaming {rows_to_stream} rows\n"
+                    )
+                    stream = load_dataset(**load_kwargs, streaming=True)
+                    dataset = Dataset.from_list(list(stream.take(rows_to_stream)))
+                    print(
+                        f"[dataset-slice] Downloaded {len(dataset)} rows "
+                        f"(requested {rows_to_stream})\n"
+                    )
+                    self._update_progress(
+                        status_message=f"Streamed {len(dataset)} rows from HuggingFace"
+                    )
+                else:
+                    dataset = load_dataset(**load_kwargs)
 
                 # Check if stopped during dataset loading
                 if self.should_stop:
@@ -1877,7 +1901,7 @@ class UnslothTrainer:
                     return None
 
                 self._update_progress(status_message=f"Loaded dataset from HuggingFace: {dataset_source}")
-                print(f"Loaded dataset from Hugging Face: {dataset_source}\n")
+                print(f"Loaded dataset from Hugging Face: {dataset_source} ({len(dataset)} rows)\n")
 
                 # Resolve eval split from a separate HF split (explicit or auto-detected)
                 if eval_enabled:
diff --git a/studio/frontend/src/components/ui/tooltip.tsx b/studio/frontend/src/components/ui/tooltip.tsx
index 94fba54cc2..30eb1e7b80 100644
--- a/studio/frontend/src/components/ui/tooltip.tsx
+++ b/studio/frontend/src/components/ui/tooltip.tsx
@@ -1,6 +1,6 @@
-// SPDX-License-Identifier: AGPL-3.0-only - See /studio/LICENSE.AGPL-3.0
-// Copyright © 2025 Unsloth AI
-
+// SPDX-License-Identifier: AGPL-3.0-only - See /studio/LICENSE.AGPL-3.0
+// Copyright © 2025 Unsloth AI
+
 import { Tooltip as TooltipPrimitive } from "radix-ui";
 import type * as React from "react";
 
@@ -46,14 +46,14 @@ function TooltipContent({
       <TooltipPrimitive.Content
         data-slot="tooltip-content"
         sideOffset={sideOffset}
-        className={cn(
-          "data-open:animate-in data-open:fade-in-0 data-open:zoom-in-95 data-[state=delayed-open]:animate-in data-[state=delayed-open]:fade-in-0 data-[state=delayed-open]:zoom-in-95 data-closed:animate-out data-closed:fade-out-0 data-closed:zoom-out-95 data-[side=bottom]:slide-in-from-top-2 data-[side=left]:slide-in-from-right-2 data-[side=right]:slide-in-from-left-2 data-[side=top]:slide-in-from-bottom-2 rounded-2xl corner-squircle px-3 py-1.5 text-xs **:data-[slot=kbd]:rounded-4xl bg-foreground text-background border border-foreground/40 shadow-lg z-50 w-fit max-w-xs origin-(--radix-tooltip-content-transform-origin)",
-          className,
-        )}
+        className={cn(
+          "data-open:animate-in data-open:fade-in-0 data-open:zoom-in-95 data-[state=delayed-open]:animate-in data-[state=delayed-open]:fade-in-0 data-[state=delayed-open]:zoom-in-95 data-closed:animate-out data-closed:fade-out-0 data-closed:zoom-out-95 data-[side=bottom]:slide-in-from-top-2 data-[side=left]:slide-in-from-right-2 data-[side=right]:slide-in-from-left-2 data-[side=top]:slide-in-from-bottom-2 rounded-2xl corner-squircle px-3 py-1.5 text-xs **:data-[slot=kbd]:rounded-4xl bg-foreground text-background border border-foreground/40 shadow-lg z-[999999] w-fit max-w-xs origin-(--radix-tooltip-content-transform-origin)",
+          className,
+        )}
         {...props}
       >
         {children}
-        <TooltipPrimitive.Arrow className="size-2.5 translate-y-[calc(-50%_-_2px)] rotate-45 rounded-[2px] data-[side=left]:translate-x-[-1.5px] data-[side=right]:translate-x-[1.5px] bg-foreground fill-foreground z-50 translate-y-[calc(-50%_-_2px)]" />
+        <TooltipPrimitive.Arrow className="size-2.5 translate-y-[calc(-50%_-_2px)] rotate-45 rounded-[2px] data-[side=left]:translate-x-[-1.5px] data-[side=right]:translate-x-[1.5px] bg-foreground fill-foreground z-[999999] translate-y-[calc(-50%_-_2px)]" />
       </TooltipPrimitive.Content>
     </TooltipPrimitive.Portal>
   );
diff --git a/studio/frontend/src/features/studio/sections/dataset-section.tsx b/studio/frontend/src/features/studio/sections/dataset-section.tsx
index c61565cbd9..4f20133b5c 100644
--- a/studio/frontend/src/features/studio/sections/dataset-section.tsx
+++ b/studio/frontend/src/features/studio/sections/dataset-section.tsx
@@ -89,6 +89,13 @@ function formatUpdatedDate(timestamp: number | null): string {
   return new Date(timestamp * 1000).toLocaleDateString();
 }
 
+function normalizeSliceInput(value: string): string | null {
+  const trimmed = value.trim();
+  if (!trimmed) return null;
+  if (!/^\d+$/.test(trimmed)) return null;
+  return trimmed;
+}
+
 export function DatasetSection() {
   const {
     dataset,
@@ -641,6 +648,19 @@ export function DatasetSection() {
               datasetEvalSplit={datasetEvalSplit}
               setDatasetEvalSplit={setDatasetEvalSplit}
             />
+          ) : !selectedDatasetName ? (
+            <HfDatasetSubsetSplitSelectors
+              variant="studio"
+              enabled={false}
+              datasetName={null}
+              accessToken={hfToken || undefined}
+              datasetSubset={datasetSubset}
+              setDatasetSubset={setDatasetSubset}
+              datasetSplit={datasetSplit}
+              setDatasetSplit={setDatasetSplit}
+              datasetEvalSplit={datasetEvalSplit}
+              setDatasetEvalSplit={setDatasetEvalSplit}
+            />
           ) : datasetSource === "upload" && selectedLocalDataset ? (
             <div className="rounded-lg border bg-muted/20 px-3.5 py-3">
               <div className="mb-3 flex items-center justify-between gap-3">
@@ -770,11 +790,14 @@ export function DatasetSection() {
                       </Tooltip>
                     </span>
                     <Input
+                      type="number"
                       inputMode="numeric"
+                      min={0}
+                      step={1}
                       placeholder="0"
                       value={datasetSliceStart ?? ""}
                       onChange={(e) =>
-                        setDatasetSliceStart(e.target.value || null)
+                        setDatasetSliceStart(normalizeSliceInput(e.target.value))
                       }
                     />
                   </div>
@@ -802,11 +825,14 @@ export function DatasetSection() {
                       </Tooltip>
                     </span>
                     <Input
+                      type="number"
                       inputMode="numeric"
+                      min={0}
+                      step={1}
                       placeholder="End"
                       value={datasetSliceEnd ?? ""}
                       onChange={(e) =>
-                        setDatasetSliceEnd(e.target.value || null)
+                        setDatasetSliceEnd(normalizeSliceInput(e.target.value))
                       }
                     />
                   </div>
diff --git a/studio/frontend/src/features/training/api/mappers.ts b/studio/frontend/src/features/training/api/mappers.ts
index f9fb07d6dc..b10c282113 100644
--- a/studio/frontend/src/features/training/api/mappers.ts
+++ b/studio/frontend/src/features/training/api/mappers.ts
@@ -12,7 +12,7 @@ function parseSliceValue(value: string | null): number | null {
   const trimmed = value.trim();
   if (!trimmed) return null;
   const num = Number(trimmed);
-  if (!Number.isFinite(num) || !Number.isInteger(num)) return null;
+  if (!Number.isFinite(num) || !Number.isInteger(num) || num < 0) return null;
   return num;
 }
 
diff --git a/studio/frontend/src/features/training/api/models-api.ts b/studio/frontend/src/features/training/api/models-api.ts
index 64ac3884e3..ca5bdc7bf7 100644
--- a/studio/frontend/src/features/training/api/models-api.ts
+++ b/studio/frontend/src/features/training/api/models-api.ts
@@ -67,8 +67,8 @@ export interface ModelConfigResponse {
   config?: BackendModelConfig | null;
   is_vision: boolean;
   is_embedding?: boolean;
+  is_audio: boolean;
   is_lora: boolean;
-  is_audio?: boolean;
   base_model?: string | null;
   model_type?: "text" | "vision" | "audio" | "embeddings" | null;
 }
diff --git a/studio/frontend/src/features/training/components/hf-dataset-subset-split-selectors.tsx b/studio/frontend/src/features/training/components/hf-dataset-subset-split-selectors.tsx
index 0dab4685be..b177a41999 100644
--- a/studio/frontend/src/features/training/components/hf-dataset-subset-split-selectors.tsx
+++ b/studio/frontend/src/features/training/components/hf-dataset-subset-split-selectors.tsx
@@ -56,6 +56,8 @@ export function HfDatasetSubsetSplitSelectors({
   } = useHfDatasetSplits(enabled ? datasetName : null, datasetSubset, {
     accessToken,
   });
+  const showPlaceholderDropdowns =
+    variant === "studio" && !enabled && !datasetName;
 
   // Auto-select subset and split in one pass to avoid racing effects
   useEffect(() => {
@@ -83,12 +85,48 @@ export function HfDatasetSubsetSplitSelectors({
     setDatasetSplit,
   ]);
 
-  if (!enabled || !datasetName) return null;
-
   const showDropdowns = !isLoading && !error && hfSubsets.length > 0;
 
   return (
     <>
+      {showPlaceholderDropdowns && (
+        <>
+          <div className="grid gap-3 sm:grid-cols-2">
+            <SelectorDropdown
+              variant={variant}
+              label="Subset"
+              tooltip="Select which subset (config) of the dataset to use."
+              value={null}
+              onChange={setDatasetSubset}
+              options={[]}
+              placeholder="Select a subset..."
+              disabled={true}
+            />
+            <SelectorDropdown
+              variant={variant}
+              label="Train Split"
+              tooltip="Select which split to use for training."
+              value={null}
+              onChange={setDatasetSplit}
+              options={[]}
+              placeholder="Select a split..."
+              disabled={true}
+            />
+          </div>
+          <SelectorDropdown
+            variant={variant}
+            label="Eval Split"
+            tooltip="Select which split to use for evaluation. None means no evaluation during training."
+            value={null}
+            onChange={setDatasetEvalSplit}
+            options={[]}
+            placeholder="None"
+            allowNone
+            disabled={true}
+          />
+        </>
+      )}
+
       {isLoading && (
         <div
           className={
@@ -184,6 +222,7 @@ function SelectorDropdown({
   options,
   placeholder,
   allowNone = false,
+  disabled = false,
 }: {
   variant: "wizard" | "studio";
   label: string;
@@ -193,7 +232,11 @@ function SelectorDropdown({
   options: string[];
   placeholder: string;
   allowNone?: boolean;
+  disabled?: boolean;
 }) {
+  const selectValue =
+    value ?? (allowNone && !disabled ? "_none" : undefined);
+
   if (variant === "wizard") {
     return (
       <Field>
@@ -217,8 +260,9 @@ function SelectorDropdown({
           </Tooltip>
         </FieldLabel>
         <Select
-          value={value ?? "_none"}
+          value={selectValue}
           onValueChange={(v) => onChange(v === "_none" ? null : v)}
+          disabled={disabled}
         >
           <SelectTrigger className="w-full">
             <SelectValue placeholder={placeholder} />
@@ -260,8 +304,9 @@ function SelectorDropdown({
         </Tooltip>
       </span>
       <Select
-        value={value ?? "_none"}
+        value={selectValue}
         onValueChange={(v) => onChange(v === "_none" ? null : v)}
+        disabled={disabled}
       >
         <SelectTrigger className="w-full">
           <SelectValue placeholder={placeholder} />
diff --git a/studio/frontend/src/features/training/stores/training-config-store.ts b/studio/frontend/src/features/training/stores/training-config-store.ts
index 4ee25ca668..56460137a9 100644
--- a/studio/frontend/src/features/training/stores/training-config-store.ts
+++ b/studio/frontend/src/features/training/stores/training-config-store.ts
@@ -37,6 +37,7 @@ const initialState: TrainingConfigState = {
   isCheckingVision: false,
   isVisionModel: false,
   isEmbeddingModel: false,
+  isAudioModel: false,
   isLoadingModelDefaults: false,
   modelDefaultsError: null,
   modelDefaultsAppliedFor: null,
@@ -60,6 +61,7 @@ const NON_PERSISTED_STATE_KEYS: ReadonlySet<keyof TrainingConfigState> = new Set
   "modelType",
   "isCheckingVision",
   "isEmbeddingModel",
+  "isAudioModel",
   "isLoadingModelDefaults",
   "modelDefaultsError",
   "modelDefaultsAppliedFor",
@@ -129,6 +131,16 @@ export const useTrainingConfigStore = create<TrainingConfigStore>()(
               patch.trainOnCompletions = false;
             }
 
+            const isAudio = !!modelDetails.is_audio;
+            // Pure audio model → always uncheck trainOnCompletions.
+            if (isAudio && !modelDetails.is_vision) {
+              patch.trainOnCompletions = false;
+            }
+            // Audio-capable vision model (e.g. gemma3n) + audio dataset → uncheck.
+            if (isAudio && modelDetails.is_vision && get().isDatasetAudio) {
+              patch.trainOnCompletions = false;
+            }
+
             // Use backend-provided model_type when available, otherwise
             // infer from capability flags.
             const isEmbedding = !!modelDetails.is_embedding;
@@ -140,6 +152,7 @@ export const useTrainingConfigStore = create<TrainingConfigStore>()(
               modelType: inferredModelType,
               isVisionModel: modelDetails.is_vision,
               isEmbeddingModel: isEmbedding,
+              isAudioModel: isAudio,
               isLoadingModelDefaults: false,
               isCheckingVision: false,
               modelDefaultsError: null,
@@ -152,6 +165,7 @@ export const useTrainingConfigStore = create<TrainingConfigStore>()(
 
             set({
               isLoadingModelDefaults: false,
+              isAudioModel: false,
               modelDefaultsError:
                 error instanceof Error
                   ? error.message
@@ -165,12 +179,13 @@ export const useTrainingConfigStore = create<TrainingConfigStore>()(
                 set({
                   modelType: isVision ? "vision" : "text",
                   isVisionModel: isVision,
+                  isAudioModel: false,
                   isCheckingVision: false,
                 });
               })
               .catch(() => {
                 if (get().selectedModel !== modelName) return;
-                set({ isCheckingVision: false });
+                set({ isCheckingVision: false, isAudioModel: false });
               });
           });
       };
@@ -198,10 +213,18 @@ export const useTrainingConfigStore = create<TrainingConfigStore>()(
               isCheckingDataset: false,
             };
             if (!_trainOnCompletionsManuallySet) {
-              const { isVisionModel } = get();
+              const { isVisionModel, isAudioModel } = get();
               if (isVisionModel && isImage) {
                 updates.trainOnCompletions = false;
               }
+              // Pure audio model → always uncheck regardless of dataset.
+              if (isAudioModel && !isVisionModel) {
+                updates.trainOnCompletions = false;
+              }
+              // Audio-capable vision model (e.g. gemma3n) + audio dataset → uncheck.
+              if (isAudioModel && isVisionModel && isAudio) {
+                updates.trainOnCompletions = false;
+              }
             }
             set(updates);
           })
@@ -238,6 +261,7 @@ export const useTrainingConfigStore = create<TrainingConfigStore>()(
             isCheckingVision: false,
             isVisionModel: false,
             isEmbeddingModel: false,
+            isAudioModel: false,
             isDatasetAudio: false,
             isLoadingModelDefaults: false,
             modelDefaultsError: null,
@@ -255,6 +279,7 @@ export const useTrainingConfigStore = create<TrainingConfigStore>()(
               isCheckingVision: false,
               isVisionModel: false,
               isEmbeddingModel: false,
+              isAudioModel: false,
               isDatasetAudio: false,
               isLoadingModelDefaults: false,
               modelDefaultsError: null,
diff --git a/studio/frontend/src/features/training/types/config.ts b/studio/frontend/src/features/training/types/config.ts
index 66d5e339ce..32069baa68 100644
--- a/studio/frontend/src/features/training/types/config.ts
+++ b/studio/frontend/src/features/training/types/config.ts
@@ -61,6 +61,7 @@ export interface TrainingConfigState {
   isCheckingVision: boolean;
   isVisionModel: boolean;
   isEmbeddingModel: boolean;
+  isAudioModel: boolean;
   isLoadingModelDefaults: boolean;
   modelDefaultsError: string | null;
   modelDefaultsAppliedFor: string | null;