⚡ Gevva e2b: Multimodal System 1 Decision Engine & NLI Cross-Encoder

Fast fact-checking, hallucination detection, tool routing & document verification based on Google Gemma 4.

Single-pass sequence classification • 10.2 GB download (5.1B params) • Apache 2.0 open weights

🤔 What is Gevva e2b?

When you ask an autoregressive LLM (like ChatGPT or Claude) to verify a fact, it generates text token-by-token, typically taking 1 to 3 seconds for an answer.

For discrete decision tasks, autoregressive generation is unnecessarily slow and expensive:

• "Does this retrieved context support the generated claim, or is it a hallucination?"
• "Which of these 4 tools should be called for this user request?"
• "Did the student's answer match the reference solution?"

The System 1 Approach

Psychologist Daniel Kahneman described human thought in two modes: System 1 (fast, automatic reflex) and System 2 (slow, deliberate reasoning).

Gevva e2b is a System 1 decision engine: a sequence classification cross-encoder built on Google's lightweight google/gemma-4-E2B-it foundation. Rather than generating text, it evaluates premise-hypothesis pairs in a single forward pass and outputs calibrated 3-class probability distributions:

$$\text{Class} \in {\text{Contradiction (0)}, \text{Entailment (1)}, \text{Neutral (2)}}$$

🚀 Quickstart in 30 Seconds

pip install gevva

from gevva import load

# Load model (automatically detects CUDA, MPS, or CPU)
engine = load("davidburhans/gevva-e2b")

# Check evidence against a claim
premise = "The Eiffel Tower is a wrought-iron lattice tower located on the Champ de Mars in Paris, France."
claim = "The Eiffel Tower is located in Paris."

probs = engine.predict([(premise, claim)])[0]
print(f"Entailment: {probs[1]*100:.1f}%, Contradiction: {probs[0]*100:.1f}%, Neutral: {probs[2]*100:.1f}%")
# -> Entailment: 96.1%, Contradiction: 0.9%, Neutral: 2.9%

⚡ Core Use Cases

1. 🛡️ Catch AI Hallucinations in RAG Pipelines

retrieved_document = "Patients taking Medication X showed improved sleep duration with no reported nausea."
ai_answer = "Medication X causes severe nausea in elderly patients."

probs = engine.predict([(retrieved_document, ai_answer)])[0]
# Returns [Contradiction: 75.2%, Entailment: 3.3%, Neutral: 21.5%]
if probs[0] > 0.70:
    print("🚨 Alert: AI Hallucination detected! Claim contradicts source document.")

2. 🎯 Tool & Intent Routing

available_tools = [
    "process_refund: Refund payment to customer bank account",
    "track_package: Query live shipping milestones and courier GPS",
    "reset_password: Send authentication link to user email",
    "search_help_docs: Search FAQs and documentation"
]

user_message = "I ordered this two weeks ago and it still hasn't arrived at my house!"

best_idx, scores = engine.rerank(user_message, available_tools)
print("Chosen Action:", available_tools[best_idx])
# -> "track_package: Query live shipping milestones and courier GPS"

3. 📝 Instant Answer & Rubric Grading

grade = engine.grade(
    question="What is the capital of Australia?",
    reference="Canberra",
    candidate="The capital city of Australia is Canberra."
)
print(f"Passed: {grade.is_correct} (Confidence: {grade.score*100:.1f}%)")
# -> Passed: True (Confidence: 81.8%)

💻 Hardware, Latency & Memory Profile

Gevva models are built on Google's multimodal gemma-4-E2B-it checkpoint. Here are real measured performance numbers across hardware:

Model Size & Download Footprint

• Total Model Parameters: 5.10 Billion parameters (~10.2 GB download in bfloat16). (Includes 2.3B active text backbone + SigLIP vision encoder + 262K vocabulary embedding table).

Measured Latency & Memory

Context Length Note: Gemma 4 natively supports up to 128K context via Rotary Position Embeddings (RoPE). However, fine-tuning for Gevva was trained on sequences up to 2,048 tokens. For optimal accuracy and latency in RAG pipelines, chunking retrieved evidence to ~1,000–2,000 tokens is strongly recommended.

📊 Benchmarks & Empirical Accuracy

1. Gevva 1.1 Distillation & Calibrated Validation Battery

Evaluated across our primary paired gate evaluation battery (7,015 items spanning Floor, Medium, and Hard adversarial slices) and real-world RAG hallucination benchmark (RAGTruth Served):

2. Independent Evaluation on Unseen Public Test Sets

Evaluated on 250 items per test set against a standard fact-checking baseline (nli-deberta-v3-base, ~184M parameters):

3. JevBench Public Dataset (231 Items)

Evaluated locally against the frozen open public split of JevBench (jevbench/datasets/public across 18 task families):

• Gevva 1.1 e2b: 71.43% overall accuracy (165/231):
  • Easy Tier: 100.0% (48/48)
  • Standard Tier: 90.28% (65/72)
  • Hard Tier: 46.85% (52/111)
  • 100% Perfect Families: Extraction, Fact, Intent, Ordinal, Tool selection, Trap, Routing Hard
  • P50 Latency: 29.8 ms (NVIDIA RTX 5090)

Transparency Note on JevBench:

• These scores were measured locally on the 231-item open public split and have not yet been evaluated by third-party maintainers on the private benchmark suite.
• The JevBench composite score weights Intelligence (25%), Calibration (25%), Speed (25%), and a hard-coded Cost tariff (25%).
• A temperature of $T=1.6$ was fitted specifically to minimize ECE on the 111 JevBench Hard tier items. However, on standard NLI test sets, $T=1.6$ flattens probabilities (raising ECE from 0.056 to 0.163). Therefore, Gevva ships with standard $T=1.0$ default inference.

⚠️ Known Limitations & Failure Modes

1. Numerical & Arithmetic Claims: Gevva can be unreliable with subtle arithmetic discrepancies (e.g. accepting "revenue exceeded five billion" when the document states "$4.2B" at 83% confidence). For strict financial and numeric validation, pair Gevva with programmatic regex or numeric validators.
2. Authentication & Password Routing: In tool-routing evaluations, requests involving account security, passwords, or authentication links were occasionally misrouted to general documentation rather than password-reset actions.
3. CPU Latency on Long Documents: While short queries evaluate in ~200 ms on CPU, full 900+ token documents take ~20 seconds on CPU. GPU acceleration (CUDA or Apple MPS) is strongly recommended for production RAG pipelines.
4. Confidence Thresholding: Because real-world calibration varies by domain, do not rely on a fixed 0.80 cutoff. Evaluate your specific positive/negative tradeoff curves to choose operational thresholds.

📦 Model Variants

📄 License & Terms

Gevva e2b is released under the open-source Apache 2.0 License. Underlying foundation weights inherit Google's Gemma Terms of Use.

@software{gevva2026,
  author = {Burhans, Dave and Contributors},
  title = {Gevva: Multimodal System 1 Decision Engine},
  year = {2026},
  publisher = {GitHub / Hugging Face},
  url = {https://github.com/davidburhans/gevva}
}