SS Suchir Salhan
Home / Research / Beetle

Flagship framework

Beetle

Small models. Big questions.

An open-source framework I built for controlled bilingual & multilingual language-model pretraining and interpretability — now released.

Now open source

Beetle is released

The software is live. Everything below points you to it.

Beetle is a framework I designed and built for controlled bilingual and multilingual pretraining. It gives you a single, declarative interface in which the tokeniser, the target language, the training budget, and the exposure structure of the data can each be manipulated independently, all configured through YAML. It grew out of computational psycholinguistics, but it is a general interpretability and controlled-pretraining infrastructure: because every model is checkpointed densely and exposure is a first-class variable, you can ask when and how abilities emerge, not just whether a final model has them. I designed the system, trained the models, and authored the libraries; it builds on the PicoDecoder backbone from PicoLM.

330 models285 bilingual45 monolingual 13 L1s 5 curricula (B1–B5) 3 data scales
Open the Beetle software Read the launch write-up Read the paper
The Beetle framework: independently manipulable tokeniser, target language, training budget, and exposure structure, all configured via YAML.
The Beetle framework. Tokeniser, target language, training budget, and exposure curriculum are independently configurable, feeding a controlled-pretraining engine that emits densely checkpointed models for analysis and interpretability.