mine-embed-v1

mine-embed-v1

Scores how close two short English texts are in meaning with a small 57.8M embedding model trained from scratch, an early experiment for people who study or baseline embedding training, not for production search.

Made by MineAITechnology on Hugging Face. These notes are Mindket's, written from its model page and model card; the model's own files and license are on Hugging Face.

About

What it is: mine-embed-v1 is a small English text embedding model from MineAI Technology with 57.8M parameters. It turns a sentence into a vector of 512 numbers, and the cosine similarity between two vectors tells you how close the texts are in meaning. The authors call it an early experiment trained from scratch, and they do not present it as competitive or production-ready.

Who it suits: It suits researchers and students who study small-scale embedding training, or who want a baseline to compare later models against.

Title: Who uses this model
Industry | Who uses it | How they use it | Example
Research | Embedding researchers | Use it as a reference point for later versions | Compare its scores with other small embedding models
Education | Students and learners | Study how an embedding model is trained from scratch | Read the training stages and run the sample
Software | Developers | Run quick similarity experiments on short English texts | Score how close two sentences are

If you need search, clustering or anything in production, you can skip it, because the card advises against those uses.

Where to use it:

• Python scripts with torch
• Windows, macOS or Linux with Python

Title: Jobs it does
Use case | Who | What they get back
Sentence similarity | Researchers | Cosine scores between short English texts
Baselines | Embedding researchers | A weak reference model to compare against
Training study | Students | A small model whose training notes are public

How to set it up: You need Python and a few packages. The repository carries its own model.py, so you load it through its Embedder class and not through AutoModel or SentenceTransformer.

• Python with pip
• torch, safetensors and huggingface_hub
• transformers, used only for the tokenizer
• A Hugging Face token, if the repository is private

1. Install the packages listed above.
2. Download the repository and add its folder to your Python path.
3. Import Embedder and encode your sentences.

pip install -U huggingface_hub
hf download MineAITechnology/mine-embed-v1

How to use it: Download the folder, import Embedder from its model file, and call encode with a list of sentences. The result is already normalized, so multiplying the matrix by its transpose gives cosine similarities. Inputs are cut to 128 tokens by default, and no query or passage prefixes are used.

import sys
from huggingface_hub import snapshot_download          # private repo: log in with an HF token first

folder = snapshot_download("MineAITechnology/mine-embed-v1")
sys.path.insert(0, folder)
from model import Embedder

e = Embedder(folder)                                    # loads config.json + model.safetensors, GPT-2 tokenizer
vecs = e.encode(["A man is playing a guitar.", "Someone strums an instrument.", "Stocks fell today."])
print(vecs.shape)                                       # torch.Size([3, 512]), already L2-normalised
print(vecs @ vecs.T)                                    # cosine similarities

Title: Main parts of the sample
Tool | What it does
snapshot_download | Fetches the repository folder from Hugging Face
Embedder | Loads the config, weights and tokenizer from that folder
encode | Turns a list of sentences into normalized vectors

Example: You encode three sentences, two about someone playing a guitar and one about stocks falling. You get a 3 by 512 array of vectors, and the matrix product shows how close each pair of sentences is.

For newcomers and experts: An embedding is a list of numbers that stands for a sentence's meaning, so similar sentences get similar lists. A first small test is the sample script with three sentences of your own. Treat the scores as experimental.

The encoder has 10 layers, hidden size 512 and 8 heads, with mean pooling and L2 normalization on the output. It uses the GPT-2 byte-pair tokenizer, which is weak outside English. Training had two stages: masked language modelling on a corpus that is largely synthetic, then contrastive fine-tuning on 471,400 triplets, mostly AllNLI plus Quora duplicates.

Random initialization
Masked language modelling
Contrastive fine-tuning
Sentence in, 512-number vector out

Good to know: The card says that on a 28-task MTEB subset the model scores well below standard small embedding models. It also says ordinary English text makes up little of the training corpus.

Title: Reference models, MTEB 28-task average
bge-base-en-v1.5: 66.25 points
bge-small-en-v1.5: 64.37 points
gte-small: 63.95 points
all-mpnet-base-v2: 62.66 points
all-MiniLM-L12-v2: 61.6 points

• License: Apache 2.0.
• Cost to run: none named.
• Data it reaches: the files come from Hugging Face, and the text you encode stays on your own computer.
• Not for: finding documents, grouping texts or running a live search service, which the card does not recommend.

On Hugging Face, 13 November 2026

• Task: sentence similarity
• Library: pytorch
• Parameters: 57.8M
• Weights: about 231 MB in F32; about 29 MB at 4-bit (estimate)
• Formats: Safetensors
• Downloads in the last 30 days: 77
• Likes: 6
• License: apache-2.0
• Languages: en
• Tags: pytorch, safetensors, embeddings, sentence-embeddings, encoder, custom-architecture, experimental, sentence-similarity

Links

• Model: https://huggingface.co/MineAITechnology/mine-embed-v1
• Files: https://huggingface.co/MineAITechnology/mine-embed-v1/tree/main

Get the model

Download its files with Hugging Face's command line:

pip install -U huggingface_hub
hf download MineAITechnology/mine-embed-v1

Then follow the usage in its model card. Nothing in this kit runs by itself.