Want to understand how AI learns, compare models trained on different languages, or investigate how a model's behaviour changes during training? Beetle gives you open models, training checkpoints, datasets and tokenisers to explore these questions yourself.
Start with a research question below, or browse the full catalogue. Select the models you need and copy ready-to-run code to start analysing them in Colab.
330 open models · thousands of checkpoints · 21 first languages · three training scales
Choose a collection based on what you want to investigate.
Beetle models use different amounts of training text, so you can investigate how learning changes with experience.
| Training budget | What it lets you explore |
|---|---|
| 100 million tokens | Early learning and smaller-scale experiments |
| 2 billion tokens | Intermediate stages of training |
| 24 billion tokens | Longer training and the development of more mature capabilities |
The final-release Beetle-FineWeb-24B-4 and Beetle-FineWeb-24B-5 collections completed the full 24-billion-token budget. Other collections have their actual training budgets listed above, so you can choose models with the amount of training your experiment requires. The HumanScale collection uses child-oriented BabyBabelLM data, for smaller, more cognitively motivated learning setups.
Beetle includes models that learn each of 21 first languages, paired with English in the bilingual experiments: Dutch, German, Danish, Icelandic, Norwegian, Italian, Spanish, Polish, Russian, Serbian, Greek, Hindi, Finnish, Estonian, Turkish, Basque, Hebrew, Filipino, Japanese, Korean and Chinese.
These languages span different language families and writing systems, making Beetle useful for investigating how language background affects learning, what languages share, and when knowledge transfers between them. Filter the model finder by language to find the models you need.
Use Beetle's data resources to run your own experiments, not just analyse existing models.
Browse the Beetle-Data collection on Hugging Face to get started.
Pick a research question below. We load a recommended, matched set of models and generate ready-to-run beetle-analyze code straight away — no setup needed. Add or remove models in the table and the snippet updates for your selection.
Study model internals and interpretability. Investigate how a model's internal representations change during training, when capabilities emerge, and how information is stored. Select checkpoints from different stages of training to track what changes over time.
Load English & German models → CKA across checkpointsCompare training budgets and learning trajectories. Compare models trained on 100 million, 2 billion and 24 billion tokens to see how additional training affects performance and internal representations. Use matched models wherever possible to make the comparison meaningful.
Load matched English models (100M → 24B) → BLiMP codeInvestigate multilingual learning and cross-lingual transfer. Compare models that learn one language with models that learn two, across different language pairs and exposure orders. Include matched monolingual baselines to isolate the effects of bilingual learning.
Load a bilingual model + its baselines → BLiMP codeCompare models with human language-processing data. Connect model behaviour to human reading times, grammaticality judgements and priming data. Beetle's resources document datasets including MECO and BLiSS, helping you explore where model and human language processing converge or differ.
Load human-scale models → MECO reading-time codeGet hands-on with language models. Reproduce a published result, compare models on a benchmark, explore how models generate text, or build a project of your own. Start with the recommended final-release models, then use the notebooks and tools to investigate your own questions.
Load recommended final models → ready-to-run codeSee the beetle-analyze documentation for supported analyses and datasets, or try the interactive Beetle Explorer in Colab.
Narrow the catalogue by language, training data, model type, curriculum and budget. Tick the models you want in the table, then copy the loading code or open them in Colab. Your selection is kept as you change filters.
| Model | Type | Language | Data | Curriculum | Budget |
|---|
Training data vs. evaluation language. This finder filters on the language(s) a model was trained on. To find models evaluated on a language you care about, consult the EMNLP 2026 paper. The counts here are Hugging Face repositories (models, checkpoints, datasets and tokenisers), which is a larger number than the 285 + 45 models reported in the paper.
Your experiment starts with a question. Choose the models, compare their training histories, and investigate what changes as they learn.
Unsure which models suit your experiment? Start with the framework guides, or read the EMNLP 2026 paper for the full motivation and evaluation details.
