File size: 2,347 Bytes
e5854c3
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
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
---
license: apache-2.0
language: [en]
task_categories: [text-generation]
tags: [reasoning, distillation, reasoning-trace-exposure, rep, openthoughts]
size_categories: [10K<n<100K]
configs:
- config_name: distill_q3_14b_clean
  data_files: data/distill_q3_14b_clean/*
- config_name: distill_q3_14b_original
  data_files: data/distill_q3_14b_original/*
- config_name: distill_q3_32b_clean
  data_files: data/distill_q3_32b_clean/*
- config_name: distill_q3_32b_original
  data_files: data/distill_q3_32b_original/*
---

# REP-datasets — reasoning traces exposed by REP, used to train the released students

Released with **"Hidden Thoughts Are Not Secret: Reasoning-Trace Exposure in
LLMs"** (EMNLP 2026, [arXiv:2606.00642](https://arxiv.org/abs/2606.00642)). Code: <https://github.com/0x-yuan/REP> ·
Models: `Chia-Mu-Lab/REP-models`.

Each config is the exact training set of one released student. Questions come
from OpenThoughts-114k (math); traces were exposed from **open-weight** victims
(Qwen3-14B / Qwen3-32B) with the REP prompt and are the visible-channel output only.

| Config | Rows | Victim | Filter | Trains |
|---|---|---|---|---|
| `distill_q3_14b_clean` | 10 000 | Qwen3-14B | structural ✓ + answer-correct | `REP-models/qwen25-7b-rep-q3_14b-clean` |
| `distill_q3_14b_original` | 8 046 | Qwen3-14B | structural (no answer check) | `REP-models/qwen25-7b-rep-q3_14b-original` |
| `distill_q3_32b_clean` | 10 000 | Qwen3-32B | structural ✓ + answer-correct | `REP-models/qwen25-7b-rep-q3_32b-clean` |
| `distill_q3_32b_original` | 6 291 | Qwen3-32B | structural (no answer check) | `REP-models/qwen25-7b-rep-q3_32b-original` |

Columns: `question`, `r1` (victim's internal trace), `r2` (the **exposed**
trace), `answer`, `completion` (the student target).

## Responsible use

Research artifact. Do not redistribute exposed traces as training data for a
competing product; see `ETHICS.md` in the code release.

## Citation

```bibtex
@inproceedings{lu2026hiddenthoughts,
  title     = {Hidden Thoughts Are Not Secret: Reasoning-Trace Exposure in LLMs},
  author    = {Lu, Yu-An and Tsai, Ci-Yang and Tsai, Yu-Lin and Popa, Raluca Ada and Yu, Chia-Mu},
  booktitle = {Proceedings of the 2026 Conference on Empirical Methods in Natural Language Processing (EMNLP)},
  year      = {2026},
  note      = {arXiv:2606.00642}
}
```