OrcaColony
AI & ML interests
OrcaColony is a distributed-training systems research project.
Recent Activity
OrcaColony
OrcaColony studies how donated compute and careful experiments can improve language models that are small enough to train and test outside frontier labs.
The project combines:
- a volunteer training framework with bounded work units, coordinator-owned optimization, restartable campaigns, and contributor-controlled credit;
- systems research on browser workers, CPU and GPU execution, PEFT, memory placement, partial-model methods, recovery, and sparse experts;
- owner-defined training campaigns with versioned usage evaluations, sample-level evidence, reproducible model and data releases, and contributor-controlled credit.
Current status
The training system has passed local correctness, restart, recovery, and resource studies. The campaign framework accepts a research question, usage scenario, evaluator, inputs, metrics, comparisons, and analysis plan supplied by the campaign owner. It preserves exact revisions, model evaluations, evidence files, limitations, and contributor credit without assigning a mandatory success or promotion state.
Record Patch v1 is retained as a historical owner-operated prototype. Its 17.5M-parameter model reached 30 of 32 strict canonical JSON outputs after 512 updates but remained at 0 of 32 exact and semantic record matches. No donated compute was used. Its thresholds, holdouts, and proposed follow-up do not define the next OrcaColony campaign.
Negative, unchanged, and inconclusive results can be released with the same provenance and credit records as positive results. Training loss remains a diagnostic, not proof that a model became more useful.
Releases and credit
Each model release can include the exact campaign, data revision, tokenizer, checkpoint, optimizer state, owner-defined evaluation records, bundled evidence artifacts, source revision, checksums, limitations, and linked agent findings.
People who donate accepted training work choose named, pseudonymous, or anonymous credit. Releases record contributor-approved totals and hardware details without publishing worker credentials or private identifiers.