OLAResearchX/memoryathena-ffn-only-20260922

MemoryATHENA — Parameter-matched FFN control

This repository contains the verified best checkpoint for the Parameter-matched FFN control condition in the MemoryATHENA artifact release.

• Base model: mistralai/Mistral-7B-v0.3
• Training/evaluation seed: 42
• Deployment: three-source reader/adaptor evaluation where applicable
• Results and provenance dataset: OLAResearchX/MemoryATHENA
• Related transfer results: OLAResearchX/XMemTransfer-Results
• Collection: MemoryATHENA

Files

The repository contains adaptor_best.pt, the corresponding config.json, and a compact artifact_manifest.json. Conditions with a self-contained memory artifact also contain memory.pt or memory_best.pt. The base Mistral weights and training/evaluation datasets are not redistributed here.

Reference control value

The paper reference recheck reports a five-task average of 24.916 for this condition. This is a reference-paper aggregate, not a new benchmark claim from this repository. See the dataset card and ARTIFACTS.md for the protocol, scorer, single-seed, and post-hoc-selection disclosures.

Intended use and limitations

Use this checkpoint for research reproduction and controlled comparison. Do not interpret oracle source-selection counts as deployed routing frequencies. The historical Vanilla Mistral row uses a different scorer from the E/router rows, and Yahoo tau=1.0 was selected post-hoc on the test sweep.