Open volunteer model training

CrowdTensor

Train a model together.

Ordinary machines contribute one bounded update at a time. When compute arrives, training advances. When it leaves, the shared checkpoint waits.

See the campaign Contribute compute
Completed proof Qwen2.5-7B 256 elastic steps +2.34 point GSM8K holdout gain

Founding campaign

Founding systems campaign

The first public Coordinator hardens the contribution loop on a pinned, supported workload before the community-selected 7B Campaign opens.

Preparing enrollment Waiting for Coordinator snapshot
Round progress 0 / 100
Model
HuggingFaceTB/SmolLM2-135M
Dataset
Salesforce/wikitext
Accepted updates 0
Training tokens 0
Active contributors 0
Adapter version v0
Open dashboard

The premise

A little compute should still count.

Distributed training usually assumes stable clusters. CrowdTensor is built around the opposite condition: contributors appear, complete a short local update, and disappear.

The Coordinator admits bounded work, validates LoRA deltas, and commits an auditable checkpoint. No eligible compute means pause, not failure.

One contribution

Join briefly. Move the checkpoint.

  1. 01

    Claim bounded work

    The Cell checks the pinned model, data shard, limits, and device fit before training.

  2. 02

    Train locally

    CPU, CUDA, or JAX/TPU completes a small LoRA update without streaming raw data.

  3. 03

    Validate and commit

    The Coordinator checks the delta contract, aggregates accepted work, and advances lineage.

  4. 04

    Leave safely

    The checkpoint persists centrally. A different contributor can continue the next round.

Completed 7B proof run

The checkpoint survived the crowd changing underneath it.

A pinned Qwen2.5-7B LoRA run completed 256 exactly-once steps. Halfway through, all old GPU Miners were deleted, training paused at zero Miners, and fresh runtimes resumed from the central checkpoint.

Read the evidence
Model
7.62B
Training
256 steps
Holdout
+2.34 pts
Supervised tokens
146,659
The preregistered practical gate passed. The bootstrap interval includes zero, so this is not a statistical-significance claim.

Campaign pipeline

Start proven. Scale by public decision.

The website will remain the public record for proposals, progress, evaluation, and checkpoints.

01
Coordinator live

Founding systems campaign

Pinned SmolLM2-135M and WikiText-2. Controlled enrollment while the Internet contribution path is hardened.

View
02
Proposal phase

Community 7B campaign

A public 7B instruction or domain LoRA campaign. Model, dataset, benchmark, moderation, and rollback must pass an RFC first.

Discuss

Contribute compute

Give one update, not a permanent machine.

Enrollment is controlled during beta. Approved contributors receive a private Campaign invite; identities and credentials never appear in the public dashboard.

After receiving an invite
crowdtensor volunteer join campaign-invite.json --device auto
Install CrowdTensor Request access

Open progress. Honest boundaries.

CrowdTensor is an engineering beta. It has real checkpointed LoRA training, heterogeneous scheduling, recovery, and retained Kaggle evidence. It does not yet have permissionless admission, poisoning resistance, independent physical multi-host proof, or a production SLA.

Read the threat model