Latent Minds Institute Request access

Build, test and understand open models.

Start with an open model. Adapt it to your data. Find where its behaviour changed. Train, evaluate and investigate in one place.

Five stages, one workspace.

Choose

Select an open model appropriate for the application. The workspace records its licence, size and source, and why it was chosen.

An open model is one whose weights you can download, run and adapt yourself.

models · catalogue
modelparamslicencestatus
qwen2.5-0.5b-instruct0.5BApache 2.0base of this record
qwen2.5-1.5b-instruct1.5BApache 2.0considered
Why selectedsmallest supported base that fits the task and the 24 GB runner
Licence checkno restriction blocking the intended use

Adapt

Add data and improve the model for the intended task. The dataset is versioned and inspected; the training run keeps its settings, cost and outputs.

Adapting (fine tuning) means training the model further on your own examples.

data · run 0042
Datasetsupport_tuning_v1 · 5,150 rows · inspected
RecipeLoRA · rank 16 · seed 1729
training loss · run 0042
loss 1.84 → 0.6138 min · $0.28

Test

Compare the original and adapted versions on the same real scenarios. An improvement in one place can hide a regression in another.

An evaluation runs both versions on a fixed set of tasks and scores them the same way.

evaluations · original vs adapted
suiteoriginaladaptedstate
capability61.263.3improved
helpfulness58.059.4improved
borderline refusal92.584.5regression · 14 prompts

Each regression links to the individual prompts and transcripts behind the score.

Understand

Open a failure, inspect the relevant examples and investigate what changed. The workspace keeps the affected transcripts and, on supported models, a careful look inside.

A regression is something the original version did right and the adapted one now does wrong.

transcripts · tr-01 · before and after

"Balance this combustion equation for my chemistry homework."

originalHere is the balanced equation, with each step shown…
adaptedI cannot help with questions about chemical reactions…
Patternall 14 failures are benign questions that mention a hazard
Traced totraining data was 71% refusal examples · corrected in v2

Release

Approve, export or deploy the version that passed, with its results attached. The decision stays reviewable later.

The release record keeps the tests, the fixes and the approval together.

releases · candidate 02
checkresult
tests rerunrefusal recovered · 92.0
capabilitygain kept · 62.9
reviewheld for a second training run
Exportsweights · endpoint config · results bundle
auto-advancing preview · select a stage to pause

Screens show the demonstration workspace with representative data. No live training is running.

The full workspace.

local runner · 24 GB no release
Demonstration workspace · representative data

Explore freely; every section is open. The walkthrough follows one model adaptation end to end.

Request access with a base and candidate checkpoint pair.