feat: add embedding model training support

Add end-to-end embedding/sentence-transformer training pipeline using
FastSentenceTransformer, SentenceTransformerTrainer, and
MultipleNegativesRankingLoss with BatchSamplers.NO_DUPLICATES.

Backend:
- Add is_embedding_model() detection via HF tags + pipeline_tag
- Add /check-embedding/ API route and EmbeddingCheckResponse
- Extend derive_model_type() to return "embeddings"
- Add _run_embedding_training() in worker.py with progress callbacks,
  stop handling, LoRA (task_type=FEATURE_EXTRACTION), and model saving
- Add is_embedding field to TrainingStartRequest and ModelDetails
- Add YAML configs for 5 models: all-MiniLM-L6-v2, bge-m3,
  embeddinggemma-300m, gte-modernbert-base, Qwen3-Embedding-0.6B

Frontend:
- Wire isEmbeddingModel flag through store, API types, and mappers
- Force packing=false, train_on_completions=false, warmup_ratio=0.03
- Hide packing and train_on_completions checkboxes for embedding models
- Auto-set modelType to "embeddings" from backend model_type response

studio/backend/utils/models/__init__.py

@@ -8,6 +8,7 @@ from .model_config import (
     ModelConfig,
     GgufVariantInfo,
     is_vision_model,
+    is_embedding_model,
     detect_audio_type,
     is_audio_input_type,
     VALID_AUDIO_TYPES,
@@ -26,6 +27,7 @@ __all__ = [
     'ModelConfig',
     'GgufVariantInfo',
     'is_vision_model',
+    'is_embedding_model',
     'detect_audio_type',
     'is_audio_input_type',
     'VALID_AUDIO_TYPES',

studio/backend/utils/models/model_config.py

@@ -25,6 +25,30 @@ logger = logging.getLogger(__name__)
 # Format: "canonical_model_name.yaml": [list of all equivalent model names]
 # Based on the model mapper provided - canonical filename is based on the first model name in the mapper
 MODEL_NAME_MAPPING = {
+    # ── Embedding models ──
+    "unsloth_all-MiniLM-L6-v2.yaml": [
+        "unsloth/all-MiniLM-L6-v2",
+        "sentence-transformers/all-MiniLM-L6-v2",
+    ],
+    "unsloth_bge-m3.yaml": [
+        "unsloth/bge-m3",
+        "BAAI/bge-m3",
+    ],
+    "unsloth_embeddinggemma-300m.yaml": [
+        "unsloth/embeddinggemma-300m",
+        "google/embeddinggemma-300m",
+    ],
+    "unsloth_gte-modernbert-base.yaml": [
+        "unsloth/gte-modernbert-base",
+        "Alibaba-NLP/gte-modernbert-base",
+    ],
+    "unsloth_Qwen3-Embedding-0.6B.yaml": [
+        "unsloth/Qwen3-Embedding-0.6B",
+        "Qwen/Qwen3-Embedding-0.6B",
+        "unsloth/Qwen3-Embedding-4B",
+        "Qwen/Qwen3-Embedding-4B",
+    ],
+    # ── Other models ──
     "unsloth_answerdotai_ModernBERT-large.yaml": [
         "answerdotai/ModernBERT-large",
     ],
@@ -894,6 +918,67 @@ def download_gguf_file(
     return local_path
 
 
+# Cache embedding detection results per session to avoid repeated HF API calls
+_embedding_detection_cache: Dict[str, bool] = {}
+
+
+def is_embedding_model(model_name: str, hf_token: Optional[str] = None) -> bool:
+    """
+    Detect embedding/sentence-transformer models using HuggingFace model metadata.
+
+    Uses a belt-and-suspenders approach combining three signals:
+      1. "sentence-transformers" in model tags
+      2. "feature-extraction" in model tags
+      3. pipeline_tag is "sentence-similarity" or "feature-extraction"
+
+    This catches all known embedding models including those like gte-modernbert
+    whose library_name is "transformers" rather than "sentence-transformers".
+
+    Args:
+        model_name: Model identifier (HF repo or local path)
+        hf_token: Optional HF token for accessing gated/private models
+
+    Returns:
+        True if the model is an embedding model, False otherwise.
+        Defaults to False for local paths or on errors.
+    """
+    if model_name in _embedding_detection_cache:
+        return _embedding_detection_cache[model_name]
+
+    # Local paths have no HF metadata to query
+    if is_local_path(model_name):
+        _embedding_detection_cache[model_name] = False
+        return False
+
+    try:
+        from huggingface_hub import model_info as hf_model_info
+
+        info = hf_model_info(model_name, token=hf_token)
+        tags = set(info.tags or [])
+        pipeline_tag = info.pipeline_tag or ""
+
+        is_emb = (
+            "sentence-transformers" in tags
+            or "feature-extraction" in tags
+            or pipeline_tag in ("sentence-similarity", "feature-extraction")
+        )
+
+        _embedding_detection_cache[model_name] = is_emb
+        if is_emb:
+            logger.info(
+                f"Model {model_name} detected as embedding model: "
+                f"pipeline_tag={pipeline_tag}, "
+                f"sentence-transformers in tags={('sentence-transformers' in tags)}, "
+                f"feature-extraction in tags={('feature-extraction' in tags)}"
+            )
+        return is_emb
+
+    except Exception as e:
+        logger.warning(f"Could not determine if {model_name} is embedding model: {e}")
+        _embedding_detection_cache[model_name] = False
+        return False
+
+
 def scan_trained_loras(outputs_dir: str = "./outputs") -> List[Tuple[str, str]]:
     """
     Scan outputs folder for trained LoRA adapters.