OpenMed/gliner-relex-base-v1.0-mlx

OpenMed GLiNER-Relex Base v1.0 MLX

This is a public OpenMed MLX artifact for zero-shot entity and relation extraction on Apple Silicon. It packages knowledgator/gliner-relex-base-v1.0 in the OpenMed MLX format so it can run locally in Python, macOS, and iOS apps through OpenMedKit.

OpenMed is the main developer experience:

• Website: openmed.life
• GitHub: github.com/maziyarpanahi/openmed
• Python: pip install "openmed[mlx]"
• Swift: add the OpenMed GitHub repository as a Swift Package and use OpenMedKit

What This Model Does

OpenMed/gliner-relex-base-v1.0-mlx performs experimental zero-shot relation extraction. You provide:

• entity labels, such as medication, condition, symptom, or procedure
• relation labels, such as treats, causes, associated with, or contraindicated for
• input text

The model returns extracted entities plus directed relation triples between entity pairs.

Artifact metadata:

• OpenMed MLX task: zero-shot-relation-extraction
• OpenMed MLX family: gliner-uni-encoder-token-relex
• Backbone: DeBERTa-v3 / deberta-v2 config family
• Max sequence length: 512
• Weight format: weights.safetensors
• Runtime APIs: GLiNERRelexMLXPipeline in Python and OpenMedRelationExtractor in Swift
• Runtime status: experimental

Python Quick Start

Install OpenMed with MLX support:

pip install -U "openmed[mlx]"

Run relation extraction from the Hub snapshot:

from huggingface_hub import snapshot_download
from openmed.mlx.inference import GLiNERRelexMLXPipeline

model_path = snapshot_download("OpenMed/gliner-relex-base-v1.0-mlx")
extractor = GLiNERRelexMLXPipeline(model_path)

result = extractor.inference(
    "Aspirin was prescribed for headache after the patient reported migraine symptoms.",
    labels=["medication", "condition", "symptom"],
    relations=["treats", "associated with"],
    threshold=0.5,
    relation_threshold=0.9,
)

print(result["entities"])
print(result["relations"])

You can also load a local snapshot directory:

from openmed.mlx.inference import create_mlx_pipeline

extractor = create_mlx_pipeline("/path/to/gliner-relex-base-v1.0-mlx")

Swift And iOS Quick Start

Add OpenMedKit to your Xcode project:

1. In Xcode, choose File > Add Package Dependencies.
2. Paste https://github.com/maziyarpanahi/openmed.
3. Add the OpenMedKit package product to your app target.
4. Download this MLX artifact once, then reuse the cached local model for offline inference.

Example:

import OpenMedKit

let modelURL = try await OpenMedModelStore.downloadMLXModel(
    repoID: "OpenMed/gliner-relex-base-v1.0-mlx"
)

let extractor = try OpenMedRelationExtractor(modelDirectoryURL: modelURL)

let result = try extractor.extract(
    "Aspirin was prescribed for headache after the patient reported migraine symptoms.",
    entityLabels: ["medication", "condition", "symptom"],
    relationLabels: ["treats", "associated with"],
    threshold: 0.5,
    relationThreshold: 0.9
)

for entity in result.entities {
    print(entity.label, entity.text, entity.score)
}

for relation in result.relations {
    print(relation.head.text, relation.label, relation.tail.text, relation.score)
}

Notes for Apple apps:

• MLX inference runs on Apple Silicon and supported Apple devices.
• The iOS Simulator is not the acceptance target for MLX inference; test on a physical iPhone or iPad.
• After the first model download, inference is local and does not require a server.

Prompt Format

The artifact includes OpenMed prompt-packing metadata so Python and Swift build the same input sequence:

{
  "kind": "gliner-relex",
  "entity_token": "<<ENT>>",
  "relation_token": "<<REL>>",
  "separator_token": "<<SEP>>",
  "class_token_index": 128001,
  "rel_token_index": 128003,
  "embed_marker_token": true,
  "split_mode": "words"
}

Practical Guidance

• Start with a conservative relation_threshold, such as 0.9, for cleaner demos.
• Use concise labels. For example, prefer medication over a long natural-language phrase.
• Tune threshold and relation_threshold separately because entity recall and relation precision behave differently.
• Keep inputs under the artifact's 512 token context window or chunk text before inference.
• Treat this runtime as experimental while OpenMed continues parity testing across Python MLX and Swift MLX.

Credits

• Base checkpoint: knowledgator/gliner-relex-base-v1.0
• OpenMed GitHub: github.com/maziyarpanahi/openmed
• OpenMed website: openmed.life
• MLX packaging, Python runtime, and Swift runtime: OpenMed