ner-medgemma15-4b-it-lora

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

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"
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: 2026-02-18
• Sequence length cap (max_length): 6144
• Chunking strategy: entity_aware
  • prompt overhead tokens reserved: 256
  • output overhead tokens reserved: 1024
  • max input chunk tokens: 1536
  • overlap chunk tokens: 256
  • min chunk tokens: 256
• Batch size: 2
• Gradient accumulation: 4 (effective batch: 8)
• Learning rate: 2e-05
• Planned epochs: 3.0
• Loss masking: response-only (prompt + input chunk tokens masked with -100)

LoRA / PEFT

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

Evaluation

• Best checkpoint metric: N/A
• Train runtime: 13203.8s (3h 40m 3s)
• eval_entity_f1: 0.426995
• eval_entity_micro_f1: 0.395010
• eval_entity_parse_fail_rate: 0.828125
• eval_entity_precision: 0.669197
• eval_entity_recall: 0.357311
• eval_runtime: 3224.531100
• eval_samples_per_second: 0.020000
• eval_steps_per_second: 0.002000

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