---
pretty_name: LongMemEval ReMe Cleaned-S
language:
  - en
tags:
  - long-term-memory
  - benchmark
  - question-answering
configs:
  - config_name: default
    data_files:
      - split: test
        path: longmemeval_s_reme_cleaned.json
---

# LongMemEval ReMe Cleaned-S

`longmemeval_s_reme_cleaned.json` is a corrected version of the LongMemEval
Cleaned-S dataset. It keeps the original questions and haystack sessions while
replacing the answer and supporting-session ground truth with the reviewed
values from `final_groundtruth_cleaned_s.json`.

The corrections address inaccurate answers and evidence sessions, including
cases where evidence occurred after the question time and therefore leaked
future information into the evaluation.

## Sources

- Upstream dataset: [xiaowu0162/longmemeval-cleaned](https://huggingface.co/datasets/xiaowu0162/longmemeval-cleaned)
- Upstream file: `longmemeval_s_cleaned.json`
- Upstream revision: `98d7416c24c778c2fee6e6f3006e7a073259d48f`
- Upstream file SHA-256: `d6f21ea9d60a0d56f34a05b609c79c88a451d2ae03597821ea3d5a9678c3a442`
- Reviewed ground-truth SHA-256: `97dd2b0df872f12c682041c44fe8082f9fa55ab882f5c78c9237b7c77ba9bb00`
- Generated dataset SHA-256: `f8b308cc2d184d029dd2226e0c0a88eeac298e6318706d588e015beb1b83fc7d`

## Transformation

Records are matched one-to-one by `question_id`. The generated file preserves
the upstream schema and applies these mappings:

| Output field | Value |
| --- | --- |
| `answer` | Reviewed `answer` |
| `original_answer` | Upstream `answer` |
| `answer_session_ids` | Reviewed `evidence_session_ids` |
| `original_answer_session_ids` | Upstream `answer_session_ids` |

The reviewed `evidence_session_ids` are written to `answer_session_ids` so the
result remains compatible with the upstream Cleaned-S format. All other
fields, including `question`, `question_date`, `question_type`,
`haystack_sessions`, `haystack_session_ids`, and `haystack_dates`, are retained
from the upstream records.

## Change statistics

Changes are counted using exact JSON value equality.

| Metric | Records | Share |
| --- | ---: | ---: |
| Total | 500 | 100.0% |
| Any corrected field changed | 77 | 15.4% |
| No corrected field changed | 423 | 84.6% |
| `answer` changed | 57 | 11.4% |
| `answer_session_ids` changed | 58 | 11.6% |
| Both fields changed | 38 | 7.6% |
| Only `answer` changed | 19 | 3.8% |
| Only `answer_session_ids` changed | 20 | 4.0% |

Evidence-session totals changed from 948 to 819. The corrections removed 129
evidence references across 58 records and added none. Records with an empty
evidence list increased from 0 to 58; all 58 corrected records with empty
evidence use an answer beginning with `The information provided is not enough.`

| Question type | Total | Answer changed | Evidence changed | Either changed |
| --- | ---: | ---: | ---: | ---: |
| `knowledge-update` | 78 | 2 | 6 | 8 |
| `multi-session` | 133 | 7 | 11 | 17 |
| `single-session-assistant` | 56 | 0 | 0 | 0 |
| `single-session-preference` | 30 | 4 | 0 | 4 |
| `single-session-user` | 70 | 1 | 0 | 1 |
| `temporal-reasoning` | 133 | 43 | 41 | 47 |

## Validation

- Both inputs contain 500 unique `question_id` values, with identical ID sets.
- `question_type` and question time match for all 500 records.
- Every corrected evidence ID exists in the corresponding
  `haystack_session_ids` list.
- The generated file contains 500 unique records and retains every non-target
  value from the upstream dataset.

## Citation

If you use this dataset, please cite the original LongMemEval authors:

> Wu D, Wang H, Yu W, et al. Longmemeval: Benchmarking chat assistants on
> long-term interactive memory. arXiv preprint arXiv:2410.10813, 2024.