GETTING STARTED & TRAINING GUIDE

Your evidence. Your judgment. A useful starting point.

GrantTune is a research prototype for exploring scientific questions and developing grant drafts with expert feedback. We are pressure-testing it in one domain—translation, tRNAs, and proteostasis—in one shared workspace. The aim is to learn where models help, where they fail, and what expert corrections can improve.

Start here

1. Join and sign in. Open your invitation link and choose Sign in with Google. Use the browser tab that opened the invitation. Already-approved members can go straight to Home. A Google account alone does not grant workspace access.
2. Pick one item on Home. Home shows comparisons and discovery dossiers awaiting review. The sidebar has Home, Lab Review, Discovery, and Drafting; smaller Home tiles open the Document Library, Invites, and this guide.
3. Leave specific feedback. Start with one answer or research idea. Identify a claim that is supported, incorrect, or uncertain; explain why and suggest a correction. One careful review is more useful than a quick thumbs-up.

What each part does

• Lab Review: compare answers from the original model (baseline) and a literature-adapted version. Labels are visible. Check the science, citations, missing qualifications, and whether the answer actually addresses the question. Save notes and a preference when useful; you can also add an answer from another AI as a reference.
• Discovery: inspect research questions, candidate ideas, sources, and cited claims. Separate evidence from inference, flag unsupported leaps, and record corrections. Automated quality flags are prompts for review, not a scientific verdict.
• Document Library: keep papers, source versions, and research decisions together. Library inclusion does not mean a document has been used for training. Training history records the experiments.
• Drafting: provide a scientific question, evidence, and constraints; generate a starting draft, then edit it for accuracy and usefulness. Please distinguish real observations from hypotheses or hypothetical results.

This is a shared collaborator workspace: other members can see saved research and reviews. Use material you are comfortable sharing with the group. Model generation runs on a local research machine, so jobs may wait while it is unavailable.

How the model learns

We start with an existing Gemma 4 12B model. Literature adaptation trains it further on selected scientific text. Lab Review compares baseline and adapted answers so experts can assess the effect. Better text-prediction scores alone do not establish better science or grant writing.

Expert-corrected grant training is a separate, planned stage. Carefully revised aims and answers can provide examples of the behavior we want: faithful use of evidence, clear argument, appropriate uncertainty, and useful experimental design. Saving or approving a review does not start training or automatically add it to a training run.

1. Collect and correct. Review complete answers and record scientific errors, usefulness, and editing effort. Choose examples with an explicit permission and source-use decision.
2. Keep evaluation separate. Assign related projects and papers together to training, validation, or held-out evaluation. Keep held-out examples out of training and avoid repeatedly choosing settings against an inspected test.
3. Run a separately approved experiment. Check sources, duplicates, and exact input/answer pairs; freeze the dataset and run a small compatibility check before a training pilot.
4. Compare and decide. Use the same evidence and generation settings for baseline and adapted models. Assess scientific fidelity, unsupported claims, and editing effort. Keep the baseline available; a completed experiment is not automatically promoted into the workspace.

What we are testing now—and what comes next

For now, the scientific scope stays narrow while we pressure-test the workflow and model behavior. We want feedback on both the research output and the interface: what is confusing, what saves time, and what you would need before relying on a draft.

Once the approach is working, we can open new workspaces for other domains or teams and move to a larger model for post-training. Those are future directions; separate workspaces and a larger serving model are not available in this prototype. Each expansion will need its own evidence, expert feedback, and evaluation.

Members can create a seven-day link from Invites to bring in a collaborator through Google sign-in. Anyone with that link can join, so share it deliberately. Revoking a link prevents future joins; members who already joined retain access.

Questions or feedback: hello@granttune.com. Privacy · Terms