---
pretty_name: IMLogic
language:
- en
license: cc-by-nc-nd-4.0
task_categories:
- question-answering
size_categories:
- 1K<n<10K
tags:
- benchmark
- personalization
- memory
- retrieval
- llm
- arxiv:2606.23283
source_datasets:
- IAAR-Shanghai/HaluMem
configs:
- config_name: qa
  default: true
  data_files:
  - split: test
    path: data/qa/test.parquet
- config_name: users
  data_files:
  - split: test
    path: data/users/test.parquet
---

# IMLogic

**IMLogic** benchmarks implicit logical memory retrieval for personalized LLMs. A query may closely match a user's apparent preference, while an appropriate personalized response depends on a different memory with limited semantic overlap.

- **Paper:** [Towards Root Memory: Benchmarking and Enhancing Implicit Logical Memory Retrieval for Personalized LLMs](https://arxiv.org/abs/2606.23283)
- **Code, prompts, and evaluation scripts:** [ThisIsCosine/IMLogic](https://github.com/ThisIsCosine/IMLogic)
- **Source histories:** [HaluMem](https://huggingface.co/datasets/IAAR-Shanghai/HaluMem)
- **License:** [CC BY-NC-ND 4.0](https://creativecommons.org/licenses/by-nc-nd/4.0/)

## Dataset summary

The release contains **2,216 QA instances** associated with **20 user trajectories**, **14,948 source memory entries**, **1,387 sessions**, and **60,146 dialogue turns**.

The two evidence roles are:

- `memory_s`: semantically relevant background that alone is insufficient to determine the most appropriate personalized response.
- `memory_l`: logically relevant information whose constraints or implications are needed for that response.

Evaluation supports memory-level retrieval using source memory entries and conversation-level evaluation using the original dialogue histories. Multiple-choice and open-ended evaluation protocols are provided in the GitHub repository.

## Load the benchmark

```python
from datasets import load_dataset

# Default configuration: one row per QA instance.
qa = load_dataset("Cosineyx/IMLogic", split="test")
# Equivalent: load_dataset("Cosineyx/IMLogic", "qa", split="test")

# User metadata and links to the original trajectories.
users = load_dataset("Cosineyx/IMLogic", "users", split="test")

example = qa[0]
print(example["query"])
print(example["user_index"], example["user_id"])
```

All instances are released in a single **test** split. No official training or validation split is introduced by this Hub packaging.

### Load aligned dialogue histories

The full nested histories are retained in the original JSONL file. Keeping large trajectories outside the preview tables makes the QA viewer easier to browse.

```python
import json
from huggingface_hub import hf_hub_download

path = hf_hub_download(
    repo_id="Cosineyx/IMLogic",
    repo_type="dataset",
    filename="original/dialogue/HaluMem-Medium.jsonl",
)
with open(path, encoding="utf-8") as handle:
    histories = [json.loads(line) for line in handle if line.strip()]

history = histories[example["user_index"]]
assert history["uuid"] == example["user_id"]
sessions = history["sessions"]
memory_bank = [memory for session in sessions for memory in session["memory_points"]]
```

For a version-pinned experiment, pass the same commit SHA as `revision=` to both `load_dataset` and `hf_hub_download`.

## Data configurations and fields

| Configuration | Split | Rows | Contents |
|---|---|---:|---|
| `qa` (default) | `test` | 2,216 | Queries, reference memories, answer options, and original annotations |
| `users` | `test` | 20 | User UUIDs, trajectory counts, and source-file linkage |

### QA fields

| Field | Meaning |
|---|---|
| `qa_id` | Stable identifier, e.g. `user00_qa0000` |
| `user_id` | The original dialogue trajectory's `uuid` |
| `user_index` | Zero-based user record index in the original JSONL file |
| `qa_index` | Zero-based question index in its original per-user QA file |
| `query` | Query to be answered |
| `memory_s` | Annotated semantic memory |
| `memory_l` | Annotated logical memory |
| `options` | Named correct and diagnostic candidate responses; see below |
| `query_time` | Original timestamp string; nullable |
| `diff_similarity_score` | Original similarity-difference annotation where present; nullable, not an evaluation score |
| `source_file` | Path to the original per-user QA file in this dataset repository |

`options` contains:

- `Correct`: the reference personalized response that accounts for the logical memory.
- `Trap_Preference`: follows the apparent preference while overlooking the logical constraint.
- `Trap_Fabrication`: introduces unsupported personal details.
- `Trap_Generic`: provides a generic response without relevant personalization.

These are **gold annotations**. The named option keys reveal correctness; for MCQ evaluation, shuffle the available candidate texts and conceal the gold keys from the answering model, as in the reference evaluator. The annotated `memory_s` and `memory_l` identify target evidence; supplying them directly is different from evaluating retrieval over a full memory bank.

### User fields and alignment

The `users` configuration contains `user_index`, `user_id`, `persona_info`, `num_sessions`, `num_memories`, `num_dialogue_turns`, `num_qa`, `dialogue_file`, and `qa_file`.

Alignment follows the original benchmark: dialogue record 0 corresponds to `qa_user00.json`, record 1 to `qa_user01.json`, and so on. Both tables include the original UUID to validate this join. Preserve the original trajectory/session order and source timestamps when reproducing the temporal evaluation protocol.

## Source data and construction

IMLogic reuses HaluMem-Medium histories and memory entries. Its construction pipeline tags memories, mines candidate logical associations, and generates QA instances through a Generator–Judger–Refiner process followed by expert screening and independent sample-based human evaluation. See the paper and repository for the construction prompts and evaluation protocols.

The original source files are included without edits under `original/`. This Hub release adds Parquet views and explicit record/user identifiers for loading and browsing. It does not regenerate any benchmark examples.

## Files and reproducibility

```text
README.md
LICENSE
NOTICE.md
release_manifest.json
data/
  qa/test.parquet
  users/test.parquet
original/
  dialogue/HaluMem-Medium.jsonl
  qa/qa_user00.json
  qa/qa_user01.json
  ...
```

`release_manifest.json` records the source repository commit, row counts, user alignment, and SHA-256 digests of all original files. Evaluation scripts and model configuration instructions are maintained in the GitHub repository.

## License and attribution

This release uses **Creative Commons Attribution-NonCommercial-NoDerivatives 4.0 International (CC BY-NC-ND 4.0)**. See [`LICENSE`](LICENSE) and [`NOTICE.md`](NOTICE.md). The original HaluMem histories and annotations retain their upstream attribution and license.

## Citation

```bibtex
@article{ding2026rootmemory,
  title={Towards Root Memory: Benchmarking and Enhancing Implicit Logical Memory Retrieval for Personalized LLMs},
  author={Ding, Hongxun and Yu, Xiang and Wang, Chengbing and Xiao, Jianfei and Bao, Keqin and Wang, Wenjie and He, Xiangnan},
  journal={arXiv preprint arXiv:2606.23283},
  year={2026}
}
```

## Contact

For questions or benchmark corrections, open an issue in the [IMLogic GitHub repository](https://github.com/ThisIsCosine/IMLogic/issues) or a discussion on this dataset page.