---
license: llama3.1

base_model:
  - meta-llama/Llama-3.1-8B-Instruct

language:
  - en

tags:
  - biomedical-entity-linking
  - entity-linking
  - entity-disambiguation
  - named-entity-linking
  - biomedical
  - healthcare
  - umls
  - medmentions
  - text-generation
  - constrained-decoding
  - causal-lm
  - llm

library_name: transformers
pipeline_tag: text-generation

datasets:
  - AnonymousARR42/MedMentions

finetuning_task:
  - entity-linking

metrics:
  - recall

model-index:
  - name: LongBEL-8B-MedMentions-ST21pv
    results:
      - task:
          type: entity-linking
          name: Biomedical Entity Linking
        dataset:
          type: AnonymousARR42/MedMentions
          name: MedMentions-ST21pv
        metrics:
          - type: recall
            name: Recall@1
            value: 0.793
---

# LongBEL: Long-Context and Document-Consistent Biomedical Entity Linking

## LongBEL

**LongBEL** is a novel document-level framework for biomedical entity linking (BEL). Instead of normalizing each mention independently, LongBEL conditions each prediction on the document context and on previous normalizations produced in the same document. This design enforces document-level consistency and is enhanced by our **robust memory** mechanism. The method is introduced in our paper, currently under review.

## LongBEL (MedMentions Edition)

This is a **finetuned version of LLaMA-3-8B** trained on **MedMentions**, applying the LongBEL framework to enable long context and robust memory predictions.

| Field | Value |
|---|---|
| Base model | `meta-llama/Llama-3.1-8B-Instruct` |
| Task | Biomedical Entity Linking |
| Dataset | MedMentions-ST21pv |
| Knowledge base | UMLS 2017AA, ST21pv subset |
| Input | BigBio-like documents with mention spans and semantic groups |
| Output | Ranked UMLS concept predictions |
| Decoding | Semantic-guided constrained decoding |
| Main metric | Recall@1 |


## Intended Use

This model is intended for research on biomedical entity linking and document-level consistency.

It assumes that mention spans and semantic groups are already provided. It does **not** perform named entity recognition. In a full pipeline, a NER model should first detect mentions and assign semantic groups, then LongBEL can normalize these mentions to UMLS concepts.

## Usage

### Loading the model

```python
import torch
from transformers import AutoModelForCausalLM

model = AutoModelForCausalLM.from_pretrained(
    "AnonymousARR42/LongBEL_8B_MedMentions_st21pv",
    trust_remote_code=True,
    device_map="auto",
)
````

### Inference example

The model expects BigBio-like documents. Each entity should include a mention text, character offsets, and a semantic group in the `type` field.

```python
num_beams = 5

bigbio_pages = [
    {
        "id": "001",
        "document_id": "doc_001",
        "passages": [
            {
                "id": "0",
                "type": "paragraph",
                "text": [
                    "A 29-year-old pregnant woman presented with severe-range hypertension, "
                    "headache, and epigastric pain. Laboratory testing showed proteinuria "
                    "and mildly elevated liver enzymes. She was admitted overnight with "
                    "suspected PET and was started on urgent treatment."
                ],
                "offsets": [[0, 257]],
            }
        ],
        "entities": [
            {
                "id": "T1",
                "type": "Living Beings",
                "text": ["pregnant woman"],
                "offsets": [[14, 28]],
            },
            {
                "id": "T2",
                "type": "Disorders",
                "text": ["severe-range hypertension"],
                "offsets": [[44, 69]],
            },
            {
                "id": "T3",
                "type": "Disorders",
                "text": ["proteinuria"],
                "offsets": [[128, 139]],
            },
            {
                "id": "T4",
                "type": "Disorders",
                "text": ["PET"],
                "offsets": [[217, 220]],
            },
        ],
        "events": [],
        "coreferences": [],
        "relations": [],
    }
]

predictions = model.sample(
    bigbio_pages=bigbio_pages,
    num_beams=num_beams,
)

for i in range(0, len(predictions), num_beams):
    mention = predictions[i]["mention"]
    print(f"## Mention {(i // num_beams) + 1}: {mention}")

    for j in range(num_beams):
        pred = predictions[i + j]
        print(
            f"   - Beam {j + 1}:\n"
            f"     Predicted concept name: {pred['pred_concept_name']}\n"
            f"     Predicted code: {pred['pred_concept_code']}\n"
            f"     Beam score: {pred['beam_score']:.3f}\n"
        )
```


**Example Output:**

```text
## Mention 1: pregnant woman
   - Beam 1:
   - Predicted concept name:Pregnant Woman
   - Predicted code: C0033011
   - Beam score: 1.000

   - Beam 2:
   - Predicted concept name:Pregnant woman
   - Predicted code: C0033011
   - Beam score: 0.003

   - Beam 3:
   - Predicted concept name:Pregnant woman (person)
   - Predicted code: C0033011
   - Beam score: 0.001

   - Beam 4:
   - Predicted concept name:Pregnancy Partner
   - Predicted code: C3538996
   - Beam score: 0.000

   - Beam 5:
   - Predicted concept name:Pregnant woman (person)
   - Predicted code: C0033011
   - Beam score: 0.000

## Mention 2: severe-range hypertension
   - Beam 1:
   - Predicted concept name:Hypertensive disease
   - Predicted code: C0020538
   - Beam score: 0.078

   - Beam 2:
   - Predicted concept name:Hypertension (in some patients)
   - Predicted code: C3280936
   - Beam score: 0.022

   - Beam 3:
   - Predicted concept name:Hypertensive disease (disorder)
   - Predicted code: C0020538
   - Beam score: 0.010

   - Beam 4:
   - Predicted concept name:Hypertension, severe
   - Predicted code: C4013784
   - Beam score: 0.010

   - Beam 5:
   - Predicted concept name:Hypertension (patient A)
   - Predicted code: C4313262
   - Beam score: 0.004

## Mention 3: proteinuria
   - Beam 1:
   - Predicted concept name:Proteinurias
   - Predicted code: C0033687
   - Beam score: 1.000

   - Beam 2:
   - Predicted concept name:Proteinuric diabetic nephropathy (disorder)
   - Predicted code: C0403519
   - Beam score: 0.003

   - Beam 3:
   - Predicted concept name:Proteinuria
   - Predicted code: C0033687
   - Beam score: 0.003

   - Beam 4:
   - Predicted concept name:Proteinuric diabetic nephropathy
   - Predicted code: C0403519
   - Beam score: 0.002

   - Beam 5:
   - Predicted concept name:Proteinuric hypertension of pregnancy (disorder)
   - Predicted code: C0032914
   - Beam score: 0.001

## Mention 4: PET
   - Beam 1:
   - Predicted concept name:PET - Pre-eclamptic toxemia
   - Predicted code: C0032914
   - Beam score: 0.075

   - Beam 2:
   - Predicted concept name:PET - Pre-eclamptic toxaemia
   - Predicted code: C0032914
   - Beam score: 0.039

   - Beam 3:
   - Predicted concept name:Preeclamptic toxemia
   - Predicted code: C2931877
   - Beam score: 0.027

   - Beam 4:
   - Predicted concept name:Preeclampsia
   - Predicted code: C0032914
   - Beam score: 0.023

   - Beam 5:
   - Predicted concept name:Preeclampsia with Severe Features
   - Predicted code: C0341950
   - Beam score: 0.019
```

### Saliency map example

The model can also return token-level saliency maps during inference.

```python
predictions, saliency_maps = model.sample(
    bigbio_pages=bigbio_pages,
    num_beams=num_beams,
    with_saliency_maps=True,
)

model.display_saliency_map(saliency_maps[3])
````

Example saliency map for the mention `PET`:

<p align="center">
  <img src="saliency_map.png" alt="Saliency map for PET prediction" width="900">
</p>

## Evaluation

Entity linking performance is reported using Recall@1 with bootstrap confidence intervals. The best result is shown in **bold**, and the second-best result is <u>underlined</u> and ⭐ marks the main LongBEL-8B model.

| Model | MM-ST21PV<br>(English) | QUAERO-EMEA<br>(French) | SympTEMIST<br>(Spanish) | DisTEMIST<br>(Spanish) | MedProcNER<br>(Spanish) |
| :--- | :---: | :---: | :---: | :---: | :---: |
| **Context-Free BEL** ||||| |
| SciSpacy | 53.8 ± 1.0 | 37.1 ± 4.3 | 9.8 ± 1.3 | 21.1 ± 1.9 | 10.3 ± 1.2 |
| SapBERT | 65.6 ± 1.0 | 59.7 ± 3.8 | 34.2 ± 2.0 | 38.6 ± 2.6 | 30.4 ± 2.1 |
| CODER-all | 62.9 ± 1.1 | 66.9 ± 4.0 | 42.2 ± 2.2 | 47.0 ± 2.6 | 42.7 ± 2.1 |
| SapBERT-all | 64.6 ± 1.1 | 67.9 ± 3.9 | 49.8 ± 2.4 | 49.6 ± 2.6 | 45.1 ± 2.2 |
| BERGAMOT | 60.9 ± 1.1 | 63.8 ± 4.9 | 48.0 ± 2.7 | 48.9 ± 2.4 | 42.3 ± 2.2 |
| **Local-Context BEL** ||||| |
| ArboEL | 76.9 ± 0.9 | 63.0 ± 3.9 | 55.4 ± 2.5 | 54.7 ± 2.6 | 59.7 ± 2.6 |
| GENRE / mBART-large | 69.6 ± 1.0 | 69.3 ± 5.4 | 59.8 ± 2.7 | 58.7 ± 2.7 | 66.0 ± 2.3 |
| GENRE / Llama-1B | 73.1 ± 1.0 | 75.1 ± 3.6 | 60.5 ± 2.4 | 62.5 ± 2.3 | 67.4 ± 2.1 |
| GENRE / Llama-8B | 75.0 ± 0.9 | 73.8 ± 4.0 | 61.7 ± 2.5 | 63.2 ± 2.5 | 68.3 ± 2.2 |
| **Global-Context BEL: LongBEL** ||||| |
| LongBEL-1B | 77.6 ± 0.9 | 74.5 ± 3.7 | 59.8 ± 2.5 | 61.9 ± 2.4 | 66.6 ± 2.1 |
| LongBEL-1B + Ensemble | 78.6 ± 0.8 | <u>77.2 ± 3.0</u> | 61.8 ± 2.5 | <u>64.3 ± 2.2</u> | <u>69.0 ± 2.0</u> |
| **⭐ LongBEL-8B** | <u>79.3 ± 0.8</u> | 75.4 ± 3.4 | <u>62.0 ± 2.6</u> | 63.6 ± 2.1 | <u>69.0 ± 2.1</u> |
| LongBEL-8B + Ensemble | **80.0 ± 0.8** | **77.6 ± 3.0** | **63.3 ± 2.5** | **65.8 ± 2.2** | **71.0 ± 2.0** |

The score reported for this checkpoint is the **single LongBEL-8B model**. The ensemble result requires fusing several LongBEL input configurations and is not produced by this checkpoint alone.

## Speed and Memory

Measured on a single NVIDIA H100 80GB GPU.

| Model                   | Model memory | Candidate memory |           Speed |
| ----------------------- | -----------: | ---------------: | --------------: |
| GENRE-Llama-8B baseline |      28.6 GB |           5.4 GB | 38.2 mentions/s |
| LongBEL-8B              |      28.6 GB |           5.4 GB | 15.2 mentions/s |

LongBEL has the same model memory footprint as the sentence-level Llama-8B baseline, but it is slower because it processes longer contexts and updates document-level memory during inference.

## Limitations

This model assumes that mention spans and semantic groups are given. It does not perform mention detection.

LongBEL is most useful when concepts recur within a document. When most concepts appear only once, the memory mechanism has less information to exploit.

Because LongBEL uses previous predictions as memory, early mistakes can still influence later predictions. Robust memory training reduces this risk but does not remove it completely.

This model is intended for research use. It should not be used for clinical decision-making without additional validation and human oversight.

## Reproducibility

Code and evaluation scripts are available in this [GitHub repository](https://anonymous.4open.science/r/LongBEL-31AD).

Trained model checkpoints and processed datasets are available in the anonymous Hugging Face collection associated with LongBEL.