---
language:
- en
license: cc-by-nc-4.0
pipeline_tag: feature-extraction
tags:
- sentence-transformers
- mteb
- feature-extraction
- e8-lattice
- english
base_model: jinaai/jina-embeddings-v5-text-nano
base_model_relation: adapter
---

# jina-v5-text-nano — E8 Snap Head (1024-dim)

E8 lattice projection head trained on top of a **frozen**
[jinaai/jina-embeddings-v5-text-nano](https://huggingface.co/jinaai/jina-embeddings-v5-text-nano)
backbone. The head expands 768-dim float embeddings to **1024-dim E8 lattice coordinates** —
enabling O(1) semantic lookup in [LatticeMemory](https://huggingface.co/spaces/dfrokido/LatticeMemory).

## MTEB STS Results

| Task | Float baseline | E8 Snap Head | Delta |
|---|---|---|---|
| STSBenchmark | 0.9445 | **0.9185** | -0.0260 |
| STS13 | 0.8948 | 0.8711 | -0.0237 |
| SICK-R | 0.9197 | 0.8952 | -0.0245 |
| STS17 | 0.8835 | 0.8604 | -0.0231 |
| STS16 | 0.8687 | 0.8500 | -0.0187 |
| STS15 | 0.9281 | 0.9153 | -0.0128 |
| STS12 | 0.8534 | 0.8408 | -0.0126 |
| BIOSSES | 0.8744 | 0.8619 | -0.0125 |
| STS22 | 0.7030 | 0.6965 | -0.0065 |
| **STS14** | 0.8891 | **0.8929** | **+0.0038** |
| **10-task mean** | 0.8759 | **0.8602** | -0.0157 |

STSBenchmark 0.9185 places this model **top 10 globally** on the MTEB STS leaderboard
(260 submitted models as of 2026-06-04).

## What This Is

The backbone is fully frozen. The head learns to map 768-dim encoder output onto the
nearest E8 lattice point in 1024 dimensions, preserving cosine similarity while producing
addresses that can be indexed in an integer hash table.

- **Architecture:** `OutputOnlyE8Head(input_dim=768, output_dim=1024, hidden_dim=1024)`
- **Training:** STS-B train/val pairs, 5 epochs, AdamW lr=2e-4, batch 16
- **Inference:** frozen jina backbone → E8 head → L2-normalised 1024-dim output

## Usage

```python
from sentence_transformers import SentenceTransformer
from liora_core.embed.small_model_e8_head import load_head_checkpoint
from huggingface_hub import snapshot_download

head_dir = snapshot_download("dfrokido/jina-v5-text-nano-e8-snap-1024")
head, _ = load_head_checkpoint(head_dir)

backbone = SentenceTransformer(
    "jinaai/jina-embeddings-v5-text-nano",
    trust_remote_code=True,
    model_kwargs={"default_task": "text-matching"},
)

import torch
texts = ["semantic search", "vector retrieval"]
with torch.no_grad():
    embs = backbone.encode(texts, convert_to_tensor=True, task="text-matching")
    e8_embs = head(embs, snap=True)   # (N, 1024) E8-snapped
```

## Links

- Demo: [LatticeMemory](https://huggingface.co/spaces/dfrokido/LatticeMemory)
- GitHub: [sangmorg1-debug/e8-Project](https://github.com/sangmorg1-debug/e8-Project)
- Base model: [jinaai/jina-embeddings-v5-text-nano](https://huggingface.co/jinaai/jina-embeddings-v5-text-nano)