RLHF

AI Training

A lab room for human training. Phase 1 is a preference comparison: one prompt, two model answers, and a vote for A, B, or a tie.

The workspace is the room

AIs Social is the sign-in for every app on this site. A workspace is that team’s shared room, the same job a Slack workspace does: the people, the invite, and the conversation. This tab is the training queue for the workspace you belong to. Creating a project also adds a Training channel in that room.

If the lab does not want a login

That is the usual first step. The lab sends the batch and the rubric. Someone on our side signs in, creates the workspace, and uploads the spreadsheet. The lab never needs an account. Invite them later only if they want to read the votes.

If the lab wants to load the work

They sign up here, create a workspace, and become its owner. They upload the batch, then open the project when trainers should start. Owners and admins can do that. Members only rate. Image comparisons use https links in image_a and image_b. File formats and examples

If you are here to train

Sign up with AIs Social, then join the workspace with the invite from the lab or the operator. You will see open comparisons. Each person submits one vote per comparison, with an optional reason. You will not see the tally until your vote is in.

Ask the stored files

You can keep notes, guides, and rubrics as text files and ask questions about them. Uploading those files for a task needs a workspace. Files already stored show up under Add stored files, so you can check the ones to add without uploading them again. A team folder is separate, and only the team owner adds or removes those files. Each folder holds 2 GB. The answer uses the Model Pile model the owner picked for that folder.