Skip to content

Add Tier 2 Forge benchmark for wide conditional spread - #557

Open
mikkel wants to merge 4 commits into
developfrom
tests/forge-wide-conditional-spread
Open

mikkel wants to merge 4 commits into
developfrom
tests/forge-wide-conditional-spread

Conversation

@mikkel

@mikkel mikkel commented Nov 17, 2026

Copy link
Copy Markdown
Contributor

Registers a runnable Tier 2 stretch task in the discriminator stability view. It scores raw draws from a conditional GAN on 65 continuous features. Both groups have mean zero and exchange their narrow and wide axes. The scorer checks within-group mean and SD, coverage of 80% bag-summary intervals, interval score and energy outside the two signal coordinates. The trained-task controls include a correct sampler, a pure group swap, contracted variance, padding damage and checkpoint continuation.

The included full-budget CPU baseline uses the released EMPF configuration, seed 0 and the public deterministic initializer. G, D and the prior each received 1,200 updates, with 1,200 observed regularizer applications. Across the final generated samples, variability coverage was 65.7% and combined coverage was 71.4%. The worst group/coordinate coverage error was 23.4 percentage points against a 12-point tolerance. Forge graded the sustained result FAIL. The compact receipt and 24-frame training GIF are in reports/forge/wide-conditional-spread/.

The adapter currently supports the clean EMPF setup with a learned MoG prior and continuous penalty on each step. The check asks whether trained samples recover the right distribution. Companion PR #556 checks the observer directly.

Closes #554.

This branch has not been deployed

No deployments
Sign up for free to join this conversation on GitHub. Already have an account? Sign in to comment

Labels

None yet

Projects

None yet

Development

Successfully merging this pull request may close these issues.

1 participant