README.md

---
language:
- en
pipeline_tag: sentence-similarity
---

# e5-large-v2-4096 (LSG-Patched)

A [Local-Sparse-Global (LSG)](https://arxiv.org/abs/2210.15497) version of [intfloat/e5-large-v2](https://huggingface.co/intfloat/e5-large-v2), capable of handling up to **4096 tokens**. This model has been patched to work with more recent versions of Transformers—ensuring `is_decoder=False` for pure encoder mode and carefully handling attention masks at float precision.

## Quickstart Usage

Below is an example to encode queries and passages from the MS-MARCO passage ranking dataset.

```python
from sentence_transformers import SentenceTransformer

model = SentenceTransformer(
    "guymorganb/e5-large-v2-4096-lsg-patched",
    {"trust_remote_code": True}
)

input_texts = [
    'query: how much protein should a female eat',
    'query: summit define',
    "passage: As a general guideline, the CDC's average requirement of protein for women ages 19 to 70 is 46 grams per day. But, as you can see from this chart, you'll need to increase that if you're expecting or training for a marathon. Check out the chart below to see how much protein you should be eating each day.",
    "passage: Definition of summit for English Language Learners. : 1  the highest point of a mountain : the top of a mountain. : 2  the highest level. : 3  a meeting or series of meetings between the leaders of two or more governments."
]

embeddings = model.encode(input_texts, normalize_embeddings=True)
```

or...

```python
import torch
import torch.nn.functional as F
from torch import Tensor
from transformers import AutoTokenizer, AutoConfig, AutoModel

def average_pool(last_hidden_states: Tensor, attention_mask: Tensor) -> Tensor:
    """
    E5-style average pooling:
    1) Zero out masked tokens
    2) Sum the token embeddings
    3) Divide by the number of valid tokens
    """
    last_hidden = last_hidden_states.masked_fill(~attention_mask[..., None].bool(), 0.0)
    summed = last_hidden.sum(dim=1)
    counts = attention_mask.sum(dim=1)
    return summed / counts.unsqueeze(1)

model_name = "guymorganb/e5-large-v2-4096-lsg-patched"

# 1) Load config (force encoder mode)
config = AutoConfig.from_pretrained(model_name, trust_remote_code=True)
config.is_decoder = False

# 2) Load tokenizer & model
tokenizer = AutoTokenizer.from_pretrained(model_name, trust_remote_code=True)
model = AutoModel.from_pretrained(model_name, config=config, trust_remote_code=True)

# 3) Some example queries & passages
input_texts = [
  'query: how much protein should a female eat',
  'query: summit define',
  "passage: As a general guideline, the CDC's average requirement of protein for women ages 19 to 70 is 46 grams per day. But, as you can see from this chart, you'll need to increase that if you're expecting or training for a marathon. Check out the chart below to see how much protein you should be eating each day.",
  "passage: Definition of summit for English Language Learners. : 1  the highest point of a mountain : the top of a mountain. : 2  the highest level. : 3  a meeting or series of meetings between the leaders of two or more governments."
]

# 4) Tokenize
batch_dict = tokenizer(
    input_texts,
    max_length=4096,
    padding=True,
    truncation=True,
    return_tensors="pt"
)

# Convert attention mask to float (needed for LSG code)
batch_dict["attention_mask"] = batch_dict["attention_mask"].float()

# 5) Forward pass (return_dict=True gives a dictionary output)
outputs = model(**batch_dict, return_dict=True)

# 6) Pool
embeddings = average_pool(outputs.last_hidden_state, batch_dict["attention_mask"])

# 7) Normalize (optional)
embeddings = F.normalize(embeddings, p=2, dim=1)

# 8) Example similarity: compare first two (queries) vs. last two (passages)
scores = (embeddings[:2] @ embeddings[2:].T) * 100
print("Similarity scores:\n", scores.tolist())

```
@article{wang2022text,
  title={Text Embeddings by Weakly-Supervised Contrastive Pre-training},
  author={Wang, Liang and Yang, Nan and Huang, Xiaolong and Jiao, Binxing and Yang, Linjun and Jiang, Daxin and Majumder, Rangan and Wei, Furu},
  journal={arXiv preprint arXiv:2212.03533},
  year={2022}
}

@misc{guymorganb_e5_lsg_patch,
  title={{LSG Patch for E5-large-v2 up to 4096 tokens}},
  author={{GuyMorganB}},
  howpublished={\url{https://huggingface.co/guymorganb/e5-large-v2-4096-lsg-patched}},
  year={2023}
}
```