Instructions to use Pritish92/ner-medgemma15-4b-it-lora-b0 with libraries, inference providers, notebooks, and local apps. Follow these links to get started.

Libraries

How to use Pritish92/ner-medgemma15-4b-it-lora-b0 with PEFT:

from peft import PeftModel
from transformers import AutoModelForCausalLM

base_model = AutoModelForCausalLM.from_pretrained("google/medgemma-1.5-4b-it")
model = PeftModel.from_pretrained(base_model, "Pritish92/ner-medgemma15-4b-it-lora-b0")

Transformers

How to use Pritish92/ner-medgemma15-4b-it-lora-b0 with Transformers:

# Use a pipeline as a high-level helper
from transformers import pipeline

pipe = pipeline("text-generation", model="Pritish92/ner-medgemma15-4b-it-lora-b0")
messages = [
    {"role": "user", "content": "Who are you?"},
]
pipe(messages)

# Load model directly
from transformers import AutoModel
model = AutoModel.from_pretrained("Pritish92/ner-medgemma15-4b-it-lora-b0", dtype="auto")

Notebooks
Google Colab
Kaggle
Local Apps Settings

vLLM

How to use Pritish92/ner-medgemma15-4b-it-lora-b0 with vLLM:

Install from pip and serve model

# Install vLLM from pip:
pip install vllm
# Start the vLLM server:
vllm serve "Pritish92/ner-medgemma15-4b-it-lora-b0"
# Call the server using curl (OpenAI-compatible API):
curl -X POST "http://localhost:8000/v1/chat/completions" \
	-H "Content-Type: application/json" \
	--data '{
		"model": "Pritish92/ner-medgemma15-4b-it-lora-b0",
		"messages": [
			{
				"role": "user",
				"content": "What is the capital of France?"
			}
		]
	}'

Use Docker

docker model run hf.co/Pritish92/ner-medgemma15-4b-it-lora-b0

SGLang

How to use Pritish92/ner-medgemma15-4b-it-lora-b0 with SGLang:

Install from pip and serve model

# Install SGLang from pip:
pip install sglang
# Start the SGLang server:
python3 -m sglang.launch_server \
    --model-path "Pritish92/ner-medgemma15-4b-it-lora-b0" \
    --host 0.0.0.0 \
    --port 30000
# Call the server using curl (OpenAI-compatible API):
curl -X POST "http://localhost:30000/v1/chat/completions" \
	-H "Content-Type: application/json" \
	--data '{
		"model": "Pritish92/ner-medgemma15-4b-it-lora-b0",
		"messages": [
			{
				"role": "user",
				"content": "What is the capital of France?"
			}
		]
	}'

Use Docker images

docker run --gpus all \
    --shm-size 32g \
    -p 30000:30000 \
    -v ~/.cache/huggingface:/root/.cache/huggingface \
    --env "HF_TOKEN=<secret>" \
    --ipc=host \
    lmsysorg/sglang:latest \
    python3 -m sglang.launch_server \
        --model-path "Pritish92/ner-medgemma15-4b-it-lora-b0" \
        --host 0.0.0.0 \
        --port 30000
# Call the server using curl (OpenAI-compatible API):
curl -X POST "http://localhost:30000/v1/chat/completions" \
	-H "Content-Type: application/json" \
	--data '{
		"model": "Pritish92/ner-medgemma15-4b-it-lora-b0",
		"messages": [
			{
				"role": "user",
				"content": "What is the capital of France?"
			}
		]
	}'

Docker Model Runner
How to use Pritish92/ner-medgemma15-4b-it-lora-b0 with Docker Model Runner:
```
docker model run hf.co/Pritish92/ner-medgemma15-4b-it-lora-b0
```

Pritish92/ner-medgemma15-4b-it-lora-b0

This is a LoRA adapter fine-tuned from google/medgemma-1.5-4b-it for instruction-following NER extraction into a strict JSON list format:

[{"label":"...","text":"..."}]

This repository contains adapter weights only (not full base model weights). You must have access to google/medgemma-1.5-4b-it to run it.

Prompt format (exact)

### Instruction:
{instruction}
Maintain the JSON key order exactly as shown.
Output format: [{"label":"...","text":"..."}]

### Input:
{input_chunk}

### Response:

How to load

import torch
from peft import PeftModel
from transformers import AutoProcessor, AutoModelForImageTextToText

adapter_id = "Pritish92/ner-medgemma15-4b-it-lora-b0"
base_id = "google/medgemma-1.5-4b-it"

processor = AutoProcessor.from_pretrained(adapter_id, use_fast=False)
base_model = AutoModelForImageTextToText.from_pretrained(
    base_id,
    dtype=torch.bfloat16,
    device_map="auto",
)
model = PeftModel.from_pretrained(base_model, adapter_id)
model.eval()

Training details

Date: 2026-02-26
Sequence length cap (max_length): 8192
Chunking strategy: entity_aware
- prompt overhead tokens reserved: 256
- output overhead tokens reserved: 1024
- max input chunk tokens: 2048
- overlap chunk tokens: 256
- min chunk tokens: 256
Batch size: 1
Gradient accumulation: 8 (effective batch: 8)
Learning rate: 2e-05
Planned epochs: 3.0
Loss masking: response-only (prompt + input chunk tokens masked with -100)
W&B run name: N/A
W&B run id: 2pjkzb4r

LoRA / PEFT

LoRA rank (r): 64
LoRA alpha: 128
LoRA dropout: 0.05
Target modules: language_model.layers.20.self_attn.q_proj, language_model.layers.25.self_attn.k_proj, language_model.layers.18.self_attn.q_proj, language_model.layers.18.self_attn.k_proj, language_model.layers.23.self_attn.v_proj, language_model.layers.19.self_attn.k_proj, language_model.layers.7.self_attn.v_proj, language_model.layers.1.self_attn.q_proj, language_model.layers.10.self_attn.q_proj, language_model.layers.16.self_attn.k_proj, language_model.layers.16.self_attn.v_proj, o_proj, language_model.layers.24.self_attn.v_proj, language_model.layers.7.self_attn.k_proj, language_model.layers.13.self_attn.v_proj, language_model.layers.10.self_attn.k_proj, 28.self_attn.k_proj, language_model.layers.0.self_attn.v_proj, language_model.layers.8.self_attn.k_proj, language_model.layers.21.self_attn.q_proj, language_model.layers.23.self_attn.k_proj, language_model.layers.9.self_attn.q_proj, language_model.layers.25.self_attn.v_proj, language_model.layers.15.self_attn.k_proj, language_model.layers.2.self_attn.k_proj, 31.self_attn.k_proj, language_model.layers.9.self_attn.v_proj, language_model.layers.26.self_attn.k_proj, language_model.layers.11.self_attn.k_proj, language_model.layers.3.self_attn.q_proj, language_model.layers.11.self_attn.q_proj, language_model.layers.4.self_attn.q_proj, language_model.layers.17.self_attn.q_proj, language_model.layers.24.self_attn.k_proj, 30.self_attn.v_proj, language_model.layers.14.self_attn.v_proj, language_model.layers.5.self_attn.q_proj, 27.self_attn.k_proj, 27.self_attn.v_proj, language_model.layers.3.self_attn.v_proj, language_model.layers.17.self_attn.k_proj, language_model.layers.3.self_attn.k_proj, up_proj, 29.self_attn.q_proj, language_model.layers.4.self_attn.k_proj, language_model.layers.14.self_attn.k_proj, 31.self_attn.v_proj, language_model.layers.19.self_attn.q_proj, 32.self_attn.k_proj, 29.self_attn.v_proj, 31.self_attn.q_proj, 30.self_attn.k_proj, 27.self_attn.q_proj, language_model.layers.22.self_attn.v_proj, language_model.layers.23.self_attn.q_proj, language_model.layers.5.self_attn.v_proj, language_model.layers.17.self_attn.v_proj, language_model.layers.24.self_attn.q_proj, 28.self_attn.q_proj, language_model.layers.16.self_attn.q_proj, language_model.layers.13.self_attn.k_proj, language_model.layers.26.self_attn.q_proj, language_model.layers.26.self_attn.v_proj, language_model.layers.5.self_attn.k_proj, 29.self_attn.k_proj, 32.self_attn.q_proj, language_model.layers.21.self_attn.v_proj, language_model.layers.19.self_attn.v_proj, language_model.layers.11.self_attn.v_proj, language_model.layers.21.self_attn.k_proj, language_model.layers.9.self_attn.k_proj, language_model.layers.14.self_attn.q_proj, language_model.layers.1.self_attn.k_proj, language_model.layers.8.self_attn.q_proj, 33.self_attn.k_proj, language_model.layers.12.self_attn.q_proj, language_model.layers.2.self_attn.q_proj, language_model.layers.25.self_attn.q_proj, language_model.layers.15.self_attn.v_proj, 28.self_attn.v_proj, language_model.layers.4.self_attn.v_proj, language_model.layers.0.self_attn.k_proj, 30.self_attn.q_proj, language_model.layers.22.self_attn.k_proj, language_model.layers.12.self_attn.v_proj, language_model.layers.2.self_attn.v_proj, language_model.layers.18.self_attn.v_proj, language_model.layers.6.self_attn.k_proj, language_model.layers.6.self_attn.v_proj, 33.self_attn.q_proj, language_model.layers.22.self_attn.q_proj, language_model.layers.20.self_attn.v_proj, down_proj, language_model.layers.12.self_attn.k_proj, language_model.layers.1.self_attn.v_proj, language_model.layers.10.self_attn.v_proj, language_model.layers.0.self_attn.q_proj, language_model.layers.8.self_attn.v_proj, 33.self_attn.v_proj, language_model.layers.20.self_attn.k_proj, language_model.layers.7.self_attn.q_proj, language_model.layers.13.self_attn.q_proj, language_model.layers.6.self_attn.q_proj, gate_proj, language_model.layers.15.self_attn.q_proj, 32.self_attn.v_proj

Training data

Local CSVs:

NER/NER-Data/ner_train_dataset.csv
NER/NER-Data/ner_dev_dataset.csv
NER/NER-Data/ner_test_dataset.csv

Example counts: raw train=18,115, raw val=2,010; after chunking train examples=25,570

Evaluation

Best checkpoint metric: eval_entity_f1_balanced=0.339016 (best checkpoint: step 9591)
Train runtime: 58611.6s (16h 16m 51s)
eval_entity_f1: 0.367399
eval_entity_f1_balanced: 0.339016
eval_entity_f1_clinical_status: 0.180791
eval_entity_f1_diagnosis: 0.372832
eval_entity_f1_symptom: 0.378277
eval_entity_label_macro_f1: 0.310633
eval_entity_micro_f1: 0.344190
eval_entity_parse_fail_rate: 0.906250
eval_entity_precision: 0.654518
eval_entity_precision_clinical_status: 0.405063
eval_entity_precision_diagnosis: 0.708791
eval_entity_precision_symptom: 0.721429
eval_entity_recall: 0.278551
eval_entity_recall_clinical_status: 0.116364
eval_entity_recall_diagnosis: 0.252941
eval_entity_recall_symptom: 0.256345
eval_runtime: 104.118900
eval_samples_per_second: 0.615000
eval_steps_per_second: 0.077000

Notes

MedGemma can be prompt-sensitive; keep inference prompt formatting aligned with training.
Validate JSON output before downstream use.
If google/medgemma-1.5-4b-it is gated, authenticate first.

References

MedGemma model card: https://huggingface.co/google/medgemma-1.5-4b-it
MedGemma notebooks: https://github.com/google-health/medgemma/tree/main/notebooks

Downloads last month: 1

Model tree for Pritish92/ner-medgemma15-4b-it-lora-b0

Base model

google/medgemma-1.5-4b-it

Adapter

(55)

this model