sgolkar/ai-mindmeld-data

AI Mindmeld data

AI Mindmeld connects scientific specialist models to a frozen large language model through the Latent Interconnect Protocol (LIP), whose small, independently trained connectors bring specialist representations directly into the LLM's internal states. This repository supplies the inputs for training and evaluating all 25 release connectors in sgolkar/ai-mindmeld-connectors.

It holds 38 full-data bundles and two minimal eight-object evaluation sets (40 bundles, 30.83 GB), plus one small DEMO bundle for software examples: protein (UniProtKB/Swiss-Prot) and molecule (ChEBI) domain tables, training text, evaluation items, pooled embeddings at the layers the connectors read, and ridge starts. Not included: per-residue arrays outside the DEMO bundle, all-layer scans, step-0 checkpoints, one unused ChemFM-3B evaluation embedding set, and BindingDB, Rhea, PFMbench and ProteinGym data. The inventory gives rebuild details where available.

Download

Install the release and set LIP_DATA as shown in the Quickstart. These commands use the checkout’s pinned release revisions by default.

python -m latent_interconnect.release_download --all-data --dry-run   # lists bundles and total size; installs no artifacts
python -m latent_interconnect.release_download --all-data --dest $LIP_DATA

Download only what you need with --data NAME. The inventory lists every bundle and the bundles each connector needs.

Links

• Code and documentation: https://github.com/PolymathicAI/AI-Mindmeld
• Paper: https://polymathic-ai.org/papers/aimm.pdf
• LIP library and ranking details

License

The AI Mindmeld authors license the ridge starts and the two minimal evaluation sets under CC BY 4.0. Please credit:

Siavash Golkar, Jake Kovalic, and Shirley Ho. AI Mindmeld: A Latent Interconnect Protocol. 2026. Manuscript.

Software: Siavash Golkar and the AI Mindmeld contributors, AI Mindmeld, https://github.com/PolymathicAI/AI-Mindmeld.

The other bundles, including the embedding bundles and the DEMO bundle, are outside this grant and keep their source terms. Upstream source, receiver and specialist terms apply throughout. For ChemFM-1B and ChemFM-3B, this release follows the ChemFM project license, CC BY-NC 4.0; their Hugging Face metadata instead lists MIT. The PGLM-1B model card specifies CC BY-NC 4.0. The supplied MIMIC embeddings support the released training and evaluation paths. Encoding new inputs requires the checkpoint and extraction recipe identified in the bundle. The bundle does not state a public license for that checkpoint; contact the authors for the applicable terms. See the inventory's license table and the non-commercial items.

Citation

Use the credit line above: cite the paper first, then the repository, with the software revision you used (CITATION.cff).