lucyknada, who publishes as lucy, develops tools for working with local language models. Their repositories include a long-context retrieval test and a notebook for preparing quantized models in several formats.[1][2][3]
Start with Detective Needle →
Links
GitHub Contributions
Test whether a model can recover a buried fact
Detective Needle inserts a small fact into a long input and tests whether a model can recover it. It supports OpenAI-compatible endpoints and writes results that can be viewed in a browser. The README explicitly measures context in characters, not tokens, and explains that sampling can change results between runs. This makes it a focused retrieval diagnostic, rather than a general measure of intelligence or reasoning.[2]
Prepare different quantization formats in one notebook
QuantKitty adapts ideas from AutoQuant into a portable Jupyter notebook. It supports several quantization formats, allows multiple jobs without restarting the kernel, and can upload outputs privately for inspection. Users supply their own Hugging Face credentials; the README recommends checking the private result before publishing it.[3]
Tools, writing & talks
Tools
Start here
Detective Needlegithub.comA configurable long-context retrieval test for OpenAI-compatible model endpoints.
Read the characters-versus-tokens warning, then inspect the example configuration. Running a test needs Node.js and a compatible model endpoint, potentially with API costs. Compare several runs before interpreting a single retrieval chart.
Developer[2]- QuantKittygithub.com
A portable notebook for quantizing models and checking private uploads before release.
Developer[3]