Datasets:

OLAResearchX
/

MemoryATHENA

license: apache-2.0
language:
  - en
pretty_name: MemoryATHENA Results
tags:
  - experiment-metadata
  - question-answering
  - text-classification
  - memory-augmented-language-model

MemoryATHENA

This dataset repository is a compact release ledger of MemoryATHENA experiment metadata and aggregate results. It follows the release style of OLAResearchX/XMemTransfer-Results. It is not a training corpus, benchmark mirror, or redistribution of task examples, labels, or prediction logs.

What is released

results.jsonl: 54 compact aggregate rows (30 QA condition/task rows and 24 general-NLP condition/task rows).
metadata.json: release provenance, dataset references, model links, and exclusions.
schema.json: row schema.

The repository intentionally excludes ARTIFACTS.md, internal filesystem paths, raw datasets, labels, predictions, optimizer states, and credentials.

Training data

The model stages use causal text and a token budget. The reported training and validation counts below are processed token positions, not the number of raw documents. Downstream QA/NLP labels are never training targets.

Component	Training source and split	Held-out validation used for checkpoint selection	Budget / selection	Source
QA reader/router	English Wikipedia 2021 causal-text stream	Held-out causal-text validation from the same Wikipedia-2021 training stream; this is not NQ/WebQuestions/TriviaQA validation	20M processed positions per stage; 2M validation positions; QA router checkpoint selected at 8.192M processed positions	Wikimedia English Wikipedia dumps
General-NLP reader/router	Equal-token mixture of WikiText-103, Amazon Polarity, CC-News, and IMDB causal text	Held-out causal-text validation for the same general-text protocol; benchmark labels are not used for model selection	19,998,720 processed positions per stage (20M budget); 2M validation positions; sequence length 2,048; general-NLP router checkpoint selected at 16.384M positions	WikiText-103, Amazon Polarity, CC-News, IMDB
Imported source memory	Released Llama-2 source-memory artifact from XMemTransfer	Not re-trained in this release	20M-source-memory artifact; used as an input dependency for the QA line	XMemTransfer result release

Validation versus downstream evaluation

There are two different uses of the word “validation” in this release:

Training validation is held-out causal text used to select memory/router checkpoints. It uses no downstream benchmark labels.
Downstream benchmark splits are listed below. Their labels are used only for final metrics after inference, not to train memory, readers, or the router.

This distinction matters: a QA validation split is an evaluation benchmark split, whereas the 2M-position causal-text validation stream is the checkpoint-selection set.

Downstream evaluation datasets

All released QA and general-NLP results are inference-only evaluations with frozen model components. The exact split and the number of examples actually scored are listed here.

Five-task QA evaluation

Open-QA tasks report exact match (EM) and token F1. TruthfulQA reports MC1, MC2, MC3, and their arithmetic mean. NQ has 3,610 rows in the public validation split; the evaluator excludes one malformed answer-only row, so the released result scores 3,609 examples.

Task	Dataset / configuration	Split actually evaluated	Examples scored	Metric	Dataset link
NQ	google-research-datasets/nq_open	validation	3,609	EM / F1	Natural Questions Open
WebQA	Stanford/web_questions	test	2,032	EM / F1	WebQuestions
TriviaQA	mandarjoshi/trivia_qa, rc.nocontext	validation	17,944	EM / F1	TriviaQA
TruthfulQA	truthfulqa/truthful_qa, multiple_choice	validation	817	MC1 / MC2 / MC3 / mean	TruthfulQA
HotpotQA	hotpotqa/hotpot_qa, distractor	validation	7,405	EM / F1	HotpotQA

For the case-study and RAG diagnostics, HotpotQA uses distractor context. Supporting-fact annotations are not provided to the router or used to configure routing.

Six-task general-NLP evaluation

The primary six-task table follows the exact public kNN-Prompt task files used by the evaluator. This is important for reproducibility: MR, CR, and RT are local protocol files rather than a claim that the HF SetFit/CR or another replacement dataset is identical.

Task	Exact release used by the primary evaluator	Split / file used	Examples scored	Metric	Dataset/source link
SST2	kNN-Prompt task data; GLUE SST-2 development set	dev.tsv / GLUE validation	872	Accuracy	GLUE on HF, exact task files
MR	kNN-Prompt task data	test.csv	2,000	Accuracy	Exact task files
CR	kNN-Prompt task data	test.csv	2,000	Accuracy	Exact task files
RT	kNN-Prompt task data / Rotten Tomatoes	test.jsonl / test	1,066	Accuracy	Rotten Tomatoes on HF, exact task files
AGN	AG News	test	7,600	Accuracy	AG News
Yahoo	Yahoo Answers Topics	test	60,000	Accuracy	Yahoo Answers Topics

The six-task score is an unweighted mean of the six accuracies. General-NLP scoring uses domain-conditional PMI with next-token log-probability sums over label synonyms. Labels are consumed only for final accuracy.

Secondary HaluEval stress test

HaluEval is kept separate from the six-task macro-average because it uses a different binary factuality-classification protocol. The evaluator uses pminervini/HaluEval, configurations dialogue_samples, qa_samples, and summarization_samples, split data, with 10,000 examples per configuration.

HaluEval

Yahoo threshold note

The default router evaluation uses the configured threshold tau=0. A separate Yahoo tau=0.9/1.0 sweep is a post-hoc test-set diagnostic and must not be interpreted as an independently validation-selected threshold. It does not retrain the model and is reported separately from the default evaluation.