---
license: apache-2.0
language:
  - en
pretty_name: XMemTransfer Results
size_categories:
  - n<1K
tags:
  - xmemtransfer
  - external-memory
  - experimental-results
  - arxiv:2608.17050
configs:
  - config_name: default
    data_files:
      - split: train
        path: data/release_artifacts.jsonl
---

# XMemTransfer Results

This dataset is the machine-readable release manifest and result summary for
**Cross-Model Memory Transfer via Target-Side Reader Adaptation**. Each row
describes one released XMemTransfer source-memory or target-adaptor repository.
When a released artifact has a `results.json`, the row also includes its
validation-perplexity and gate-statistic summary.

- Paper: [arXiv:2608.17050](https://arxiv.org/abs/2608.17050)
- Project page: [XMemTransfer](https://olaresearch.github.io/XMemTransfer)
- Code: [OLAResearch/XMemTransfer](https://github.com/OLAResearch/XMemTransfer)
- Models: [XMemTransfer collection](https://huggingface.co/collections/OLAResearchX/xmemtransfer-6a4a0c34ef03a51c927d3389)

## Dataset contents

The `train` split contains 20 release-manifest rows. Twelve rows include result
summaries copied from the corresponding released model artifacts. Null metric
fields mean that the artifact is a source memory without a target-adaptor
result summary.

Important fields:

- `artifact_id`, `artifact_kind`, and `hf_repo`: released artifact identity.
- `base_model` and `source_memory_repo`: model provenance.
- `token_budget`: target-side fitting budget for adaptors, or source-memory
  training budget for source artifacts.
- `best_val_ppl`, training-step fields, and `gate_*`: values recorded by the
  released experiment's `results.json`.

## Scope and limitations

This is experiment and release metadata, not a language-model training corpus
or a benchmark. It does **not** redistribute WikiText, evaluation examples, or
model weights. Metrics are run summaries and should be interpreted with the
experimental setup in the paper and repository. Local filesystem paths in the
original run outputs are deliberately excluded.

## Datasets used by the paper

The experiments use the following public corpora and evaluation datasets. They
are listed here as references only; this release does not mirror their data:

- **Training and intrinsic evaluation:**
  [WikiText-103](https://huggingface.co/datasets/Salesforce/wikitext)
  (`wikitext-103-raw-v1`),
  [Wikipedia-2021](https://huggingface.co/datasets/Rubin-Wei/enwiki-dec2021-preprocessed-mistral)
  (December 2021 English Wikipedia, Mistral-tokenizer preprocessed release),
  [LAMBADA](https://huggingface.co/datasets/EleutherAI/lambada_openai),
  [C4](https://huggingface.co/datasets/allenai/c4) (`en`), and
  [FineWeb-Edu](https://huggingface.co/datasets/HuggingFaceFW/fineweb-edu)
  (`sample-10BT`).
- **Open-domain QA and transfer evaluation:**
  [Natural Questions Open (NQ)](https://huggingface.co/datasets/google-research-datasets/nq_open),
  [WebQuestions (WebQA)](https://huggingface.co/datasets/Stanford/web_questions),
  [TriviaQA](https://huggingface.co/datasets/mandarjoshi/trivia_qa)
  (the `rc.nocontext` RC no-context subset; validation split),
  [TruthfulQA](https://huggingface.co/datasets/truthfulqa/truthful_qa)
  (`multiple_choice`), and
  [HotpotQA](https://huggingface.co/datasets/hotpotqa/hotpot_qa) (`distractor`).
- **Multiple-choice and classification evaluation:**
  [HellaSwag](https://huggingface.co/datasets/Rowan/hellaswag),
  [PIQA](https://huggingface.co/datasets/ybisk/piqa),
  [ARC](https://huggingface.co/datasets/allenai/ai2_arc) (Easy and Challenge),
  [Winogrande](https://huggingface.co/datasets/allenai/winogrande)
  (`winogrande_xl`), [BoolQ](https://huggingface.co/datasets/google/boolq),
  [SuperGLUE RTE](https://huggingface.co/datasets/aps/super_glue) (`rte`),
  [OpenBookQA](https://huggingface.co/datasets/allenai/openbookqa),
  [SciQ](https://huggingface.co/datasets/allenai/sciq), and
  [RACE](https://huggingface.co/datasets/ehovy/race) (`high`).

See the paper's dataset and evaluation protocol for the exact splits and
preprocessing. Wikipedia-2021 is the factual/encyclopedic corpus used for the
QA memory-contribution experiments.

`OLAResearchX/XMemTransfer-Results` is a release-results and metadata index.
Its `arxiv:2608.17050` tag is what makes it appear under **Datasets citing this
paper** on the paper page; that Hugging Face panel lists repositories that
declare a paper link, rather than automatically listing every external source
dataset referenced in this card.

## License

Apache-2.0, matching the XMemTransfer code release.