CrowdTensor volunteer-training Beta and Draft Qwen2.5-7B GSM8K Campaign RFC

CrowdTensor, an Apache-2.0 open-source protocol for checkpointed volunteer model-training campaigns, completed a feasibility run using Qwen/Qwen2.5-7B-Instruct and openai/gsm8k, achieving an accuracy improvement from 71.875% to 74.219%. The preregistered threshold was passed, but no statistical significance or broad-capability claim is made. A smaller Founding SmolLM2/WikiText-2 campaign now serves live aggregate status, though the project still lacks independent physical multi-host evidence and permissionless adversarial safety.

CrowdTensor is an Apache-2.0 open-source protocol for checkpointed volunteer model-training Campaigns. A Campaign pins immutable model/data revisions and gives admitted CPU/GPU/TPU Cells bounded work. Accepted LoRA deltas advance one auditable checkpoint; contributor disappearance pauses or reassigns work instead of discarding committed progress. The completed feasibility run used: Qwen/Qwen2.5-7B-Instruct@a09a35458c702b33eeacc393d103063234e8bc28 openai/gsm8k@740312add88f781978c0658806c59bc2815b9866 71.875% - 74.219% The practical preregistered threshold passed; the paired bootstrap interval included zero, so no statistical-significance or broad-capability claim is made. A smaller Founding SmolLM2/WikiText-2 Campaign now serves live aggregate status. Two maintainer-operated private Kaggle GPU Cells seeded its first round through the public HTTPS contribution path. Enrollment remains controlled, and the project still lacks independent physical multi-host evidence and permissionless adversarial safety. I would value review of the Draft 7B RFC, especially the fresh-holdout design, the 256-to-1,024-step extension rule, quantized 16 GB GPU work units, delta validation, and maintainer/rollback policy.