MicroGuard — Gemma-1B
A LoRA-adapted faithfulness classifier for RAG systems. Detects whether a generated answer is faithful to the retrieved context.
Performance
| Metric | Value |
|---|---|
| Balanced Accuracy | 69.4% |
| F1 Score | 0.721 |
| Cohen's Kappa | 0.447 |
| Inference Latency | 88ms |
Evaluated on a combined test set of 15,976 examples from RAGBench, RAGTruth, and HaluBench.
Usage
from transformers import AutoModelForCausalLM, AutoTokenizer
from peft import PeftModel
base = AutoModelForCausalLM.from_pretrained("google/gemma-3-1b-it")
model = PeftModel.from_pretrained(base, "tarun5986/MicroGuard-Gemma-1B")
tokenizer = AutoTokenizer.from_pretrained("google/gemma-3-1b-it")
# Or use the MicroGuard package
from microguard import MicroGuard
guard = MicroGuard(model="tarun5986/MicroGuard-Gemma-1B", base_model="google/gemma-3-1b-it")
result = guard.check(
context="The Eiffel Tower was built in 1889 by Gustave Eiffel.",
question="Who built the Eiffel Tower?",
answer="The Eiffel Tower was built by Gustave Eiffel in 1889."
)
print(result) # {'verdict': 'FAITHFUL', 'confidence': 74.2, 'latency_ms': 64.0}
Training
- Method: LoRA (r=16, alpha=32, targets: q,k,v,o projections)
- Data: 127,932 examples from RAGBench + RAGTruth + HaluBench
- Evaluation: Constrained decoding via logit comparison (0% garbage outputs)
Paper
MicroGuard: Sub-Billion Parameter Faithfulness Classification for Real-Time RAG QA
Citation
@article{microguard2026,
title={MicroGuard: Sub-Billion Parameter Faithfulness Classification for Real-Time RAG QA},
author={Sharma, Tarun},
journal={IEEE Access},
year={2026}
}