MemoryATHENA general-NLP release
This repository contains the shared MemoryATHENA checkpoint used for six paper-aligned general-NLP tasks: SST2, MR, CR, RT, AGN, and Yahoo Answers Topics. The six tasks reuse the same frozen memory, expert readers, and router; this is one model release rather than six duplicate copies of the same weights.
Training data
The causal training stream is a general-mixed equal-token mixture of WikiText-103, Amazon Polarity review text, CC-News, and IMDB text. Training uses 20M tokens per phase, sequence length 2048, batch size 1, and seed 42. The source-memory artifact is frozen and the expert/router artifacts are included here. The training corpora are referenced by name only and are not redistributed.
Test data and protocol
The held-out evaluation tasks are SST2 (872 examples), MR (2,000), CR (2,000), RT (1,066), AGN (7,600), and Yahoo Answers Topics (60,000). The paper-aligned scorer uses domain-conditional PMI with next-token log-probability sums over label synonyms; labels are used only for final accuracy calculation. Raw test examples and labels are not included.
Results (accuracy percent)
| Condition | SST2 | MR | CR | RT | AGN | Yahoo |
|---|---|---|---|---|---|---|
| Historical Vanilla Mistral reference | 81.08 | 75.60 | 74.00 | 74.67 | 73.24 | 55.03 |
| Engram-only (forced E path) | 84.17 | 81.00 | 82.40 | 82.36 | 72.93 | 57.51 |
| Three-model router, default tau=0 | 88.07 | 84.70 | 84.10 | 83.86 | 76.64 | 45.91 |
| Three-model router, Yahoo tau=1.0 post-hoc sweep | 88.07 | 84.70 | 84.10 | 83.86 | 76.64 | 57.43 |
The Vanilla reference uses a historical scorer and is not a fully protocol-matched headline comparison. The Yahoo tau=1.0 value is a post-hoc test-set threshold sweep and must not be read as an independently validation-selected default; the default tau=0 row is retained.
Files and links
memory/: frozen source memory and source adaptor.experts/: trained expert readers and metadata.router/: trained advantage router and metadata.metadata/: release metadata and compact six-task result summary.- Dataset: OLAResearchX/MemoryATHENA.
- Results reference: OLAResearchX/XMemTransfer-Results.
- Collection: MemoryATHENA.
No ARTIFACTS.md, raw datasets, raw test examples, labels, prediction lists, optimizer states, credentials, or internal filesystem paths are part of this release.
Model tree for OLAResearchX/memoryathena-general-nlp-20260922
Base model
mistralai/Mistral-7B-v0.3