HalluGuard: 4B-Param Hallucination Detector

• HalluGuard is a 4B-parameter small reasoning model designed to detect and mitigate hallucinations in retrieval-augmented generation pipelines.
• It delivers evidence-grounded chain-of-thought justifications that increase auditability and transparency in claim verification.
• Its training pipeline uses synthetic data and ORPO fine-tuning to achieve robust generalization and efficient deployment in critical domains.

HalluGuard is a 4B-parameter Small Reasoning Model (SRM) specifically engineered to detect and mitigate hallucinations in Retrieval-Augmented Generation (RAG) pipelines. It performs document-claim pair classification, outputs evidence-grounded justifications for transparency, and achieves competitive accuracy with models twice its size. HalluGuard’s design prioritizes robust generalization, efficiency, and chain-of-thought explainability, making it suitable for enterprise and safety-critical RAG deployments.

1. Model Architecture and Inference Dynamics

HalluGuard utilizes a 4B-parameter backbone (Qwen3-4B) fine-tuned for hallucination detection in RAG. The forward process takes as input a document $x$ and claim $c$, returning both a binary label and a justification:

• Classification function:

$$t(x, c) = \begin{cases} \text{grounded} & \text{if } c \text{ is fully supported by } x, \ \text{hallucinated} & \text{otherwise} \end{cases}$$

• Reasoning trace: During inference, HalluGuard generates intermediate reasoning steps (“think mode”) enclosed within special tags (e.g., > …) before outputting the final decision. The justification explicitly quotes the evidence in $x$ that supports or contradicts $c$.
• Output: Beyond the binary classification, HalluGuard provides an evidence-centric justification, increasing transparency and model auditability.

This architecture aligns large-model chain-of-thought reasoning with efficient deployment, striking a balance between interpretability and real-time inference.

2. Synthetic Dataset Creation and Curation

The foundation of HalluGuard’s training is a large, high-diversity synthetic dataset sourced and curated as follows:

• Base corpus: 250,000 documents sampled from the FineWeb web crawl, filtered for language, safety, formatting, and duplication.
• Multi-stage curation:
  • Unsafe, short, or gibberish-like documents are removed.
  • Near-duplicate elimination increases informational diversity.
• Data reformation: Each document is rewritten in one of 18 target linguistic styles using Llama-3.3-70B, with temperature and stylistic parameters sampled uniformly. This syntactic and discourse reformation significantly broadens the distribution over which the SRM is trained.
• Claim generation: For each stylized document, Llama-3.3-70B generates roughly equal numbers of synthetic “grounded” and “hallucinated” claims. Hallucinations include both intrinsic (fabricated) and extrinsic (misattributed) types, though these are treated as a single class during training.
• Preference data: For every (document, claim) pair, two model completions are generated:
  1. PG-Large (Qwen3-32B, high-quality response)
  2. PG-Small (Qwen3-0.6B) The preferred completion is heuristically assigned to the larger model, but filtered by model agreement and verified by independent evaluators.

This pipeline creates a scalable, domain-agnostic, and highly controlled source of training data—addressing both scarcity and quality constraints endemic to hallucination research.

3. Training and Preference-Based Fine-Tuning

HalluGuard's optimization unifies supervised and preference learning via Odds Ratio Preference Optimization (ORPO):

• Training tuples: Each (x, c) pair is associated with two candidate responses (preferred/non-preferred) along with their labels and justifications.

• Preference refinement: Initial preference assignments based on model size are vetted against label agreement and third-party consensus to avoid systematic bias.
• ORPO: This mechanism allows the model to directly optimize for both correct classification and high-quality rationale generation within a single training stage. Parameter-efficient fine-tuning (LoRA) is employed to encode preference tuples.
• Outcome: This procedure distills complex reasoning skills from larger models into the 4B backbone, supporting light-weight deployment without significant performance loss.

4. Evaluation and Performance Metrics

HalluGuard is benchmarked primarily against the LLM-AggreFact suite, specifically its RAGTruth subset:

• Balanced Accuracy (BAcc) is used:

$$\text{BAcc} = \frac{1}{2}\left(\frac{TP}{TP + FN} + \frac{TN}{TN + FP}\right)$$

where TP/FP/TN/FN are standard confusion matrix terms.

• RAGTruth (LLM-AggreFact): HalluGuard-4B achieves 84.0% BAcc, precisely matching MiniCheck-7B and outperforming Granite Guardian 3.3-8B (82.2%), using roughly half the parameter count.
• Full benchmark: 75.7% BAcc, matching GPT-4o (75.9%), thus validating its ability to scale beyond the specialized data regime.
• Justification transparency: The model’s rationales are explicitly cited, enabling audit and risk management at the claim level.

The comparative table below summarizes core results:

Model	Parameters	RAGTruth BAcc	Full AggreFact BAcc
HalluGuard-4B	4B	84.0%	75.7%
MiniCheck-7B	7B	84.0%	–
Granite Guardian 3.3	8B	82.2%	–
GPT-4o	N/A	–	75.9%

5. Evidence-Grounded Chain-of-Thought Justification

A distinctive feature of HalluGuard is its multi-step intermediate reasoning:

• Chain-of-thought (“think mode”): The model generates verbose, step-by-step justifications prior to issuing the final label.
• Evidence citation: Reasoning traces reference specific document spans, enabling traceability for both grounded and hallucinated outputs.
• Transparency implications: This explicitness reduces the risk of unjustified or spurious classifications, supporting deployment in regulated or safety-sensitive domains.

This mechanism positions HalluGuard as an interpretable alternative to opaque discriminative hallucination detectors.

6. Generalization, Deployment, and Accessibility

HalluGuard’s design emphasizes broad generalization and practical deployment:

• Domain-agnosticism: Data reformulation into diverse linguistic and discourse styles enables transfer to out-of-domain data such as reports and dialogues, not only web-native prose.
• Efficient deployment: The 4B parameter footprint allows on-premises or edge deployment, suiting enterprise or regulated use-cases.
• Open access: The authors commit to releasing both model and datasets under Apache 2.0, supporting reproducibility and adaptation.

A plausible implication is that HalluGuard’s synthetic data and distillation procedures could serve as blueprints for constructing lightweight, transparent detectors in other high-risk NLP settings.

7. Context in Retrieval-Augmented Generation and Future Directions

Within the RAG paradigm, HalluGuard addresses a primary failure mode—unsupported or hallucinated claims blended with retrieved content. Its strong balanced accuracy, interpretability through reasoning traces, and architectural efficiency position it as a practical solution for trust-critical RAG applications, including legal, financial, and compliance workflows.

Potential directions for system extension include explicit modeling of hallucination subtypes (extrinsic versus intrinsic), integration with RAG pipeline retrieval calibration, and longitudinal studies on deployment effectiveness in enterprise environments.

In summary, HalluGuard exemplifies the convergence of model scaling, synthetic data curation, preference distillation, and reasoning transparency, offering a state-of-the-art approach for hallucination mitigation in Retrieval-Augmented Generation and related LLMing scenarios (Bergeron et al., 1 Oct 2025).