Supernova TeraLLM Embedding V4

Part of the Supernova TeraLLM Embedding family.

Architecture

Supernova V4

Lightweight Supernova embedding architecture designed for fast semantic retrieval.

Speciality

High-speed lightweight retrieval.

Model size

Parameters: 3,635,328

This model is designed to be lightweight and suitable for semantic retrieval experiments.

Supernova Embedding Family

Model Role
V1 Balanced retrieval
V2 Stronger semantic retrieval
V3 Alternative architecture
V4 Lightweight/high-speed retrieval

Training

The models were trained as part of the Supernova semantic retrieval research project.

The evaluation process includes:

  • Recall@1
  • Recall@3
  • Recall@5
  • MRR
  • unseen-data evaluation
  • speed testing
  • comparison against Sentence Transformers

Intended use

This model is intended for:

  • semantic search
  • Nepali information retrieval
  • retrieval-augmented generation
  • document matching
  • lightweight embedding experiments

Limitations

This is a research model.

Performance can vary depending on:

  • dataset
  • domain
  • tokenizer coverage
  • candidate distribution
  • query style

The model should therefore be evaluated on data appropriate to the intended application.

Supernova Project

Developed as part of the Supernova TeraLLM embedding research project.

to use Supernova

import torch from huggingface_hub import hf_hub_download

REPO_ID = "Supernova11c/Supernova-teraillm-Embedding-V4"

weights_path = hf_hub_download( repo_id=REPO_ID, filename="pytorch_model.bin" )

state_dict = torch.load( weights_path, map_location="cpu" )

model = SupernovaV4()

model.load_state_dict( state_dict )

model.eval()

Downloads last month
23
Safetensors
Model size
3.64M params
Tensor type
F32
·
Inference Providers NEW
This model isn't deployed by any Inference Provider. 🙋 Ask for provider support