Add LLaDA-8B Base Brain-Score model plugin

[hxu129]

Brain-Score Language model submission

This PR registers llada-8b-base (GSAI-ML/LLaDA-8B-Base) as a Brain-Score Language ArtificialSubject for neural-model evaluation.

What the plugin exposes

• Clean-input, frozen layer-24 representations with passage-local cumulative context.
• Mean pooling over current text-part tokens.
• fMRI and ECoG neural tasks; behavioral tasks explicitly raise NotImplementedError until a diffusion conditional-likelihood / decoding policy is separately registered.
• Pinned checkpoint revision and an LLADA_MODEL_PATH override for offline or shared-GPU evaluation.

Brain-Score validation

Under Brain-Score Language 0e0bb4a2d8df5d3c30fe26ec4528f27a188e1cdc:

• LLaDA and a fixed GPT-Neo-1.3B control completed the 12 registered neural identifiers across Pereira2018, Blank2014, Fedorenko2016, and Tuckute2024.
• The live Pereira-384 raw score was 0.134269, matching the precomputed reference 0.133745 within the pre-registered 0.001 tolerance.
• Plugin discovery, two-part live neural digestion, and rejection of unsupported behavioral tasks were validated.

Important scope boundary

Tuckute2024-rdm is undefined for both models in this Brain-Score revision: the supplied assembly has one neuroid, so its official row-wise RDM correlation has zero variance. It is documented but not averaged into a neural headline.

These results establish a clean representational-alignment baseline. They do not claim a diffusion-specific mechanism, behavioral alignment, or a neural temporal-process match.

Files

Adds the model package, requirements, contract tests, and a model-specific README under brainscore_language/models/llada8b/.

[hxu129]

Submission-status update: the fresh local official-protocol neural matrix completed for all 12 registered identifiers (no failed identifiers), and the plugin contract test passed (2 passed).

The remaining blocker is the submission workflow rather than model evaluation. Its first job fails before dependency installation with:

A branch or tag with the name 'llada8b-brainscore-neural-v1' could not be found\n\nThis PR head is hxu129/language:llada8b-brainscore-neural-v1, while the workflow checks out ref: github.event.pull_request.head.ref from the upstream repository without setting repository: github.event.pull_request.head.repo.full_name. Could a maintainer please either rerun this submission from an upstream-owned branch or update the checkout to support fork heads? The normal Language unit, plugin, integration, and documentation checks are all green.\n\nFor scope clarity: behavioral benchmarks remain intentionally unsupported; Tuckute2024-rdm is undefined in this Brain-Score revision because the supplied assembly has one neuroid, so its row-wise RDM correlation has zero variance.

[hxu129]

Hi Brain-Score maintainers — could you please check the status of the Jenkins scoring job triggered after merge of PR #409 for llada-8b-base?

The post-merge workflow reports that the scoring trigger succeeded, but the model has not yet appeared on the public Language leaderboard or in my Language profile. The workflow also fell back to its default email/user ID because it could not resolve my GitHub username (hxu129) to an email.

Could you please associate this submission and its eventual results with my Brain-Score account: hxu401@connect.hkust-gz.edu.cn?

For clarity, the local official-protocol neural matrix completed successfully, but I am not treating those local values as the official leaderboard result. Thank you.

[KartikP]

Hi @hxu129 thanks for submitting the model. I'll keep an eye on your model submission - it may take a few hours. If anything is acting up, please email me at kpradeep@mit.edu.