270M text only EmbeddingGemma-2 fork for Hinglish and Hindi retrieval RAG. Loads without vision or audio. ONNX INT8 and GGUF ready. Apache-2.0. By Eulogik.
| Check | Base | Ours |
|---|---|---|
| Hindi IndicQA NDCG@10, 768d | 0.7279 | 0.7324 |
| Hindi STS spearman | 0.6430 | 0.6423 |
| ONNX INT8 drift | - | 0.0066 |
| Dims | NDCG@10 | vs 768 |
|---|---|---|
| 768 | 0.7324 | - |
| 512 | 0.7247 | -0.008 |
| 256 | 0.7023 | -0.030 |
| 128 | 0.6294 | -0.103 |
Pick 512 for quality, 256 for balance. 128 costs real quality on Hindi.
from sentence_transformers import SentenceTransformer
model = SentenceTransformer("eulogik/bharat-embed-270m-gemma2",
config_kwargs={"vision_config": None, "audio_config": None},
model_kwargs={"torch_dtype": "bfloat16"})
q = model.encode("GST refund kaise claim karein?", prompt_name="SearchQuery",
truncate_dim=256, normalize_embeddings=True)