Home / Projects
Active research projects
Member-led research projects — the kind where the bottleneck is the idea, not the cluster. Each one is hosted by one to three members. Projects mark themselves "open call" when they want collaborators, annotators, GPU contributors, or reviewers.
Long-trajectory replay buffers for open agents
Training a 7B agent on a 14-day partial-observability trajectory replay buffer. Looking for help with eval harness and GPU contributors.
Sparse-feature atlas for the open 7B stack
A community-contributed atlas of dictionary features across four popular open 7B checkpoints. 18k features labelled so far, 41 contributors.
Cantonese tool-use evaluation set
Built from real Hong Kong customer-service transcripts. Pays $30 per accepted annotation hour, capped at 12h per contributor.
Reproducibility audit of router-distillation papers
Replicating nine published router-distillation papers end-to-end. So far: three reproduce, two partially, one fails with original code, three in progress.
African-language code-switching prompts
Yoruba, Igbo, Hausa, Swahili and Amharic code-switching prompts for agent tool-use. Pays $40/h to verified native speakers.
`agentcheck` — schema linter for agent tools
Now standard for alphabell submissions. Looking for plugin authors for Anthropic + Gemini tool schemas.
Causal scrubbing for feature lifetimes
Tracking dictionary-feature lifetimes across pretraining checkpoints. Mid-training data and code shared monthly.
Agent trajectory dataset with paired human review
Building a 4,000-trajectory dataset of agent-human paired sessions. Pays $25 per accepted review hour.
Have a project of your own?
If your idea is bigger than a mini-grant — a multi-month research thread, a public dataset, a benchmark — propose it as an alphabell project. Project leads get default compute and the right to recruit collaborators from the member pool.