scifact

VerifAI initiative to build open-source easy-to-deploy generative question-answering engine that can reference and verify answers for correctness (using posteriori model)

🩺🔍 Combating Biomedical Misinformation through Multi-modal Claim Detection and Evidence-based Verification

A genral RAG Search chatbot, with SoTA RAG techniques such as HyDE, Hybrid retrieval with BM25 + RRF and Cross encoder reranking. Evaluated on the BEIR scifact dataset and compared all the different pipelines i tried along the way

🩺⚡ SynthMed: Generating and Detecting Multimodal Deepfakes for Healthcare Communication

Public release of SciFCheX system developed for COM3610 Dissertation Project at the University of Sheffield. The pipeline is designed to perform fact-checking on scientific claims.

Domain adaption for an embedding model using unsupervised and supervised finetuning on scientific texts for the SciFact retrieval task.

Evaluating retrieval strategy bias in biomedical scientific claim verification using RAG — comparing dense, BM25, hybrid, and query-reformulation retrieval across Support Recall, Contradiction Recall, and Balance Score on SciFact.

RAG evaluation on BEIR SciFact: BM25, dense and hybrid retrieval with LLM answers.

Evidence retrieval and claim verification on SciFact: BM25 from scratch, a retrieval-vs-reasoning failure decomposition, and a fine-tuned cross-encoder. Leakage audits throughout; every number states its split and confidence interval.

Scientific retrieval and citation auditing: disjoint supervised groups, 300 verdicts, 1128 paired stress cases and a blinded unreviewed packet.

Independent benchmark: can decision models (Jev 1.13, Span-02, D1 Ultra) replace a dedicated RAG reranker (BGE)? 300 SciFact claims, human labels as the answer key, bootstrap CIs, offline explorer.

Production-style benchmark for RAG retrieval: dense, BM25, hybrid, RRF, metadata filters & query rewriting on BEIR SciFact. Docker, Qdrant, FastAPI, GPU embeddings.

A claim-level fact-checking system that uses RoBERTa to spot unsupported or contradictory scientific claims.

Production-style RAG answering API with hybrid retrieval, grounded generation, citations, validation, caching, and tracing.

SciFact retrieval experiments with BM25, hard negatives, custom InfoNCE dual encoders, and reranking.

Chiral Narrative Synthesis workspace for Thinker/Tinker LoRA pipelines, semantic fact-checking, telemetry, and reviewer-ready CNS docs.

Verifies that each citation in a draft actually supports the claim it's attached to. Faithfulness, not existence.

Reproducible SciFact evidence-retrieval benchmark for scientific RAG

Banc d'évaluation du retrieval sur BEIR SciFact : métriques codées à la main, campagnes avec prédiction écrite avant mesure, runs suivis dans MLflow.

Hybrid search benchmark for BM25 + dense vector + cross-encoder rerank on distractor-augmented BEIR scifact.