File size: 3,154 Bytes
e18eade
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
50
51
52
53
54
55
56
57
58
59
60
61
62
63
64
65
66
67
68
69
70
71
72
73
74
75
76
77
78
79
80
81
82
83
84
85
86
87
88
89
90
91
92
93
94
95
96
97
98
99
100
101
102
103
104
105
106
107
108
109
110
111
112
113
114
115
116
117
118
119
120
121
122
123
124
125
126
127
128
129
130
131
---
pretty_name: MemFactory
license: cc-by-sa-4.0
language:
- en
task_categories:
- question-answering
tags:
- long-context
- evaluation
- question-answering
- multi-hop
- hotpotqa
- synthetic
configs:
- config_name: eval_50
  data_files:
  - split: train
    path: eval_50.json
- config_name: eval_100
  data_files:
  - split: train
    path: eval_100.json
- config_name: converted_hotpotqa_2000
  data_files:
  - split: train
    path: converted_hotpotqa_2000.json
- config_name: eval_fwe_16384
  data_files:
  - split: train
    path: eval_fwe_16384.json
---


# MemFactory

## Overview

This repository provides a lightweight, derivative release of data used in **MemFactory**.

To evaluate the effectiveness of MemFactory, we reuse and adapt data from the upstream dataset:

- [`BytedTsinghua-SIA/hotpotqa`](https://huggingface.co/datasets/BytedTsinghua-SIA/hotpotqa)

This repository includes four JSON files:

- `eval_50.json`
- `eval_100.json`
- `eval_fwe_16384.json`
- `converted_hotpotqa_2000.json`

### Data Sources

- The three evaluation files are directly derived from the upstream HotpotQA-based release.
- The training file `converted_hotpotqa_2000.json` is a **locally adapted version** of the upstream training data, modified for MemFactory experiments.

For full dataset context, please refer to the upstream release:
- https://huggingface.co/datasets/BytedTsinghua-SIA/hotpotqa

---


## Limitations

- This is a **derivative redistribution**, not the original dataset.
- The data may inherit:
  - annotation noise  
  - biases  
  - structural limitations  
  from the upstream sources.
- `eval_fwe_16384.json` follows a **different schema** from the QA-style files.
- For full documentation and broader coverage, users should consult the upstream dataset.

---

## License

This repository is released under **CC BY-SA 4.0**.

**Reason:**

- The data is derived from the upstream HotpotQA-based dataset, which uses the same license.
- `converted_hotpotqa_2000.json` is an adapted derivative and must preserve share-alike terms.

If you use or redistribute this repository:
- Please retain attribution to the upstream source  
- Preserve the same license  

---

## Loading with 🤗 datasets

```python
from datasets import load_dataset

eval_50 = load_dataset("nworats/MemFactory", "eval_50", split="train")
eval_100 = load_dataset("nworats/MemFactory", "eval_100", split="train")
train_converted = load_dataset("nworats/MemFactory", "converted_hotpotqa_2000", split="train")
eval_fwe_16384 = load_dataset("nworats/MemFactory", "eval_fwe_16384", split="train")
````


## Citation

If you use this dataset, please cite:

### MemFactory (this work)

*(Placeholder – replace with your paper when available)*

```bibtex
@article{memfactory2025,
  title={MemFactory: [Your Subtitle Here]},
  author={Your Name et al.},
  journal={arXiv preprint arXiv:XXXX.XXXXX},
  year={2025}
}
```

### Upstream MemAgent work

```bibtex
@article{yu2025memagent,
  title={MemAgent: Reshaping Long-Context LLM with Multi-Conv RL-based Memory Agent},
  author={Yu, Hongli and others},
  journal={arXiv preprint arXiv:2507.02259},
  year={2025}
}
```