Model Summary
- Model name:
gemma-3-4b-it-adjuvant-extractor
- Base model:
google/gemma-3-4b-it
- Fine-tuning method: LoRA adapter training, merged into full model weights
- Primary task: Evidence-linked adjuvant extraction from title+abstract text
Prompt Used for Inference
System prompt
You are a biomedical information extraction assistant
User instruction template
Return ONLY valid JSON in this format:
[{"adjuvant": "<string>", "evidence": "<string>"}, ...]
Do not include any extra keys or explanation.
Abstract: <paper abstract>
The model receives the user instruction template followed by the title/abstract text.
The model is prompted to return a JSON array of objects with exactly two keys:
[
{
"adjuvant": "<string>",
"evidence": "<string>"
}
]
Expected behavior:
- Return a JSON array (can be empty:
[]).
- Each object must contain:
adjuvant: normalized or near-normalized adjuvant nameevidence: supporting text snippet from the same input abstract
- No extra keys and no explanatory text outside JSON.
Title: Intranasal vaccination study using alum and MPLA adjuvants in a murine influenza model.
Abstract: Mice immunized with antigen formulated with alum showed increased IgG titers. A separate group receiving MPLA-adjuvanted vaccine demonstrated stronger IFN-gamma responses and reduced viral load after challenge.
Expected Output
[
{
"adjuvant": "alum",
"evidence": "Mice immunized with antigen formulated with alum showed increased IgG titers."
},
{
"adjuvant": "MPLA",
"evidence": "A separate group receiving MPLA-adjuvanted vaccine demonstrated stronger IFN-gamma responses and reduced viral load after challenge."
}
]
Notes on Output Validity
- Output must be valid JSON.
- Output must be a JSON array (use
[] if no supported adjuvant is found).
- Each item should include only
adjuvant and evidence.
- Evidence text should come from the provided input abstract.
Intended Use
This model is intended for research workflows in biomedical literature mining, especially:
- infectious disease vaccine literature curation
- vaccine adjuvant concept extraction
- evidence-linked information extraction for downstream manual review
This model is not intended for clinical decision-making.
Training Data and Split Context
The model was trained on a curated infectious disease adjuvant corpus derived from VIOLIN ecosystem resources.
- Corpus size used in workflow: 298 abstracts
- Fixed split framework used across models:
- 256 train
- 13 validation
- 29 test
Training Configuration (fixed manuscript setting)
- LoRA rank (
r): 8
- Learning rate:
2e-4
- Epochs:
5
- Quantization during fine-tuning: 4-bit NF4 with double quantization
- Compute dtype: float16
- Per-device batch size and gradient accumulation were configured for stable updates across model families.
Usage
from transformers import AutoTokenizer, AutoModelForCausalLM
import torch
repo_id = "RehanaHasin/gemma-3-4b-it-adjuvant-extractor"
tokenizer = AutoTokenizer.from_pretrained(repo_id)
model = AutoModelForCausalLM.from_pretrained(
repo_id,
torch_dtype=torch.float16,
device_map="auto",
)
Prompting Recommendation
Use prompts that explicitly request structured JSON output containing only:
and restrict extra commentary to reduce parsing errors.
Limitations
- Evaluated on a focused infectious-disease adjuvant corpus; broader-domain generalization is not guaranteed.
- Performance depends on abstract quality and terminology variation.
- Structured output may still require post-processing and manual validation.
Ethical and Safety Notes
- Outputs can contain extraction errors or unsupported predictions.
- Human review is required before downstream knowledge integration.
- Not for diagnosis, treatment, or direct patient-care decisions.
Citation
If you use this model, please cite the associated manuscript and project repository.
For questions, please ask for provider support.
