On the Monitorability of Natural Language Autoencoders: Identification vs Discrimination
Paper • 2610.18727 • Published
You need to agree to share your Hugging Face information to access this dataset
This repository is publicly accessible, but you have to accept the conditions to access its files and content.
These files contain excerpts of, and text derived from, WildChat-1M conversations. They are released under ODC-BY, like WildChat-1M, and you must also follow WildChat-1M's terms of use.
Log in or Sign Up to review the conditions and access this dataset content.
OnMonitorabilityOfNLAs: artefacts
The released intermediate files needed to re-score the experiments of On the
Monitorability of Natural Language Autoencoders: Identification vs
Discrimination (arXiv:2610.18727), with
the code at
github.com/Near32/OnMonitorabilityOfNLAs.
setup.sh in that repository downloads this dataset into artefacts/.
Derived from WildChat
These files contain excerpts of, and text derived from, conversations in
WildChat-1M (shard
data/train-00000-of-00014.parquet, English, non-toxic). The explanations
written by the NLA quote fragments of those conversations, and the rivals and
paraphrases are rewrites of the explanations. The activations are hidden states
of Qwen2.5-7B-Instruct computed on the conversations. They are released under
the same licence as WildChat-1M, the Open Data Commons Attribution License
(ODC-BY), and anyone using them must also follow WildChat's terms of use.
@inproceedings{zhao2024wildchat,
title={WildChat: 1M Chat{GPT} Interaction Logs in the Wild},
author={Wenting Zhao and Xiang Ren and Jack Hessel and Claire Cardie and Yejin Choi and Yuntian Deng},
booktitle={The Twelfth International Conference on Learning Representations},
year={2024},
url={https://openreview.net/forum?id=Bl8u7ZRlbM}
}
Files
| file | what it is | read by |
|---|---|---|
traces/hardneg_donor_pos_traces.jsonl |
the AV's explanation of each conversation, plus donor + POS corruptions at doses 1–16 | rivals, paraphrase |
traces/hardneg_traces.jsonl |
the AV's own top-k substitutions at doses 1, 4, 16 (30 conversations) | paraphrase |
gold_activations.npz |
H: the layer-20 residual activation of each conversation (conv: its index) |
rivals, paraphrase |
rival_texts.jsonl |
8 rivals per conversation for 1, 2 and 4 changed claims, written by Qwen2.5-7B-Instruct | rivals, word cloud |
rival_judge.json |
per rival text: [meaning changed, fluent], as judged by Qwen2.5-7B-Instruct | rivals |
paraphrase_texts.jsonl |
light and heavy meaning-preserving rewrites of each explanation | paraphrase |
probe_cache.npz |
H, AR(explanation) T and AR(corruption) D per conversation and dose |
angular span |
hardneg/*.csv |
one row per conversation and dose: rank of the true explanation among its corruptions, for six mechanical constructions | mechanical table |
All files were produced with seed 0 by the pipeline the paper describes.
Citation
@misc{denamganaï2026monitorabilitynaturallanguageautoencoders,
title={On the Monitorability of Natural Language Autoencoders: Identification vs Discrimination},
author={Kevin Yandoka Denamganaï},
year={2026},
eprint={2610.18727},
archivePrefix={arXiv},
primaryClass={cs.LG},
url={https://arxiv.org/abs/2610.18727},
}
- Downloads last month
- -