Research with Cloudflare Clef

A German-language benchmark comparing Cloudflare Clef Flash 9B, Jev and a 4B model, with test inputs, reference labels, native outputs and evaluation.

Has Clef score retrieved chunks against the question and keeps only the evidence, shrinking noisy RAG context before it reaches the LLM.

Toolkit for judging outputs cheaply with Clef and auditing whether its confidence numbers are calibrated, with a rendered demo clip.

Common-ground comparison of Jev, Clef, Kev and others against embeddings, rerankers, NLI and prompted LLMs.

Benchmarks Jev, Clef-flash and Laya against an LLM and six rerankers on BEIR with paired confidence intervals.

Fine-tunes Clef decision models on a 12 GB GPU for real-time entity resolution with TRL comparisons.

Dockerised research prototype that explores language-model support for emergency triage using synthetic narratives and a local GGUF model.

Runs Clef 27B and Flash 9B through llama.cpp and MLX, with benchmarks including render QA for 3D models and prompt-injection robustness.

A SmolVLM2-500M based decision model with a Jev/CLEF-compatible API, plus data pipelines, synthetic generators, LoRA training and an eval harness.

Code for a Portuguese article comparing Clef against two generative LLMs on 24 hand-labelled support tickets, using the probability as a certainty threshold for automation.

Exploratory study comparing quantisation levels of Clef-Flash playing Snake, with per-seed results.

Rerunnable tests of state poisoning, prompt injection and truncation against Jev and Clef.

Combines an LLM for issue decomposition with Clef to classify court decisions at scale and surface arguments and precedents.

Reproducible code checking whether the models' confidence numbers match reality.

Launch-week battery comparing Jev with Clef and Clef-flash, with harness and raw evidence.

Uses Clef to steer ingest and retrieval across vector databases, hosted or fully local.

LoRA fine-tuning recipe with a joint schema head for vertical decisions on the open Clef weights.

Evaluates decision models on whether an agent's knowledge-base write should be admitted.