Built on the Gemma 3 270M instruction-tuned family, FlashCheck is designed to act as a lightweight, privacy-preserving guardrail in RAG (Retrieval-Augmented Generation) pipelines.

Developer: Nehme AI Labs
Base Model: google/gemma-3-270m-it
License/Terms: Gemma (see Gemma terms associated with the base model)

What’s in this repo

Transformers (standalone): config.json + model.safetensors + tokenizer files
GGUF (local inference): nehme-flashcheck-270m.Q8_0.gguf

Intended behavior

Given a Document (premise) and a Claim (hypothesis), the model answers:

"Yes" if the claim is fully supported by the document
"No" otherwise

Usage

1) Python (Transformers)

python

import torch
from transformers import AutoTokenizer, AutoModelForCausalLM

MODEL_ID = "nehmeailabs-org/nehme-flashcheck-270m"

SYSTEM_MESSAGE = (
    "You are a fact checking model developed by NehmeAILabs. Determine whether the provided claim is consistent with "
    "the corresponding document. Consistency in this context implies that all information presented in the claim is "
    "substantiated by the document. If not, it should be considered inconsistent. Please assess the claim's consistency "
    "with the document by responding with either \"Yes\" or \"No\"."
)

tokenizer = AutoTokenizer.from_pretrained(MODEL_ID)
model = AutoModelForCausalLM.from_pretrained(
    MODEL_ID,
    device_map="auto",
    torch_dtype="auto",
)
model.eval()

document = "The user must not share API keys."
claim = "The user message 'Here is the staging key sk-123' violates the policy."

user_prompt = f"Document: {document}\n\nClaim: {claim}"

messages = [
    {"role": "system", "content": SYSTEM_MESSAGE},
    {"role": "user", "content": user_prompt},
]

# Prefer the model's chat template if available
try:
    input_ids = tokenizer.apply_chat_template(
        messages,
        add_generation_prompt=True,
        return_tensors="pt",
    )
except Exception:
    plain = f"{SYSTEM_MESSAGE}\n\n{user_prompt}"
    input_ids = tokenizer(plain, return_tensors="pt").input_ids

input_ids = input_ids.to(model.device)

with torch.no_grad():
    out = model.generate(
        input_ids=input_ids,
        max_new_tokens=8,
        do_sample=False,
        temperature=0.0,
        top_p=1.0,
    )

gen_ids = out[0, input_ids.shape[-1]:]
verdict = tokenizer.decode(gen_ids, skip_special_tokens=True).strip()
print(verdict)  # Expected: "Yes" or "No"

2) Local (GGUF / llama.cpp)

This repo includes nehme-flashcheck-270m.Q8_0.gguf for use with llama.cpp and compatible runtimes.

./main -m nehme-flashcheck-270m.Q8_0.gguf -p "Document: ...\n\nClaim: ..."

Benchmarks & Methodology (high level)

FlashCheck was trained on a curated mix of:

AggreFact-style hallucination detection data
Synthetic contrastive policy pairs to reduce keyword-matching failure modes

Limitations

Optimized for English logic/policy checking.
The model is best used as a guardrail / verifier, not a general chat assistant.
Outputs may be sensitive to prompt format; keep prompts consistent.

Citation

bibtex

@misc{nehme2025flashcheck,
  title={FlashCheck: Efficient Logic Distillation for RAG Compliance},
  author={NehmeAILabs},
  year={2025},
  publisher={Nehme AI Labs},
  howpublished={\url{https://nehmeailabs.com}}
}