Datasets:

Modalities:
Text
Formats:
parquet
ArXiv:
License:
naghamo commited on
Commit
cf7763a
Β·
verified Β·
1 Parent(s): 622b5d0

an-alt-test: anonymous alternative annotator test annotations

Browse files
Files changed (2) hide show
  1. README.md +34 -49
  2. an_alt_test_example.ipynb +0 -0
README.md CHANGED
@@ -55,6 +55,19 @@ identities, no shared item set required), uniformly across text, image, audio an
55
  subset that all judges were run on identically (`select_items`: first 300 prompt-order items
56
  with human labels). The `N` column below is the *source* dataset size; the release uses the 300.
57
 
 
 
 
 
 
 
 
 
 
 
 
 
 
58
  ## Columns
59
 
60
  - `dataset`, `item_id`, `modality`, `label_spec`, `output_key`
@@ -64,7 +77,7 @@ with human labels). The `N` column below is the *source* dataset size; the relea
64
  - `human_labels` β€” JSON string of the human annotations
65
  - `llm` β€” JSON string mapping `"<Model>_<method>"` β†’ label. Methods: `zs300` (zero-shot),
66
  `rf5`/`rf3`/`rf7`/`rf10`/`rf15` (random few-shot), `cot`/`cotmid` (reasoning), `gepa*`, `*_ord`,
67
- `ens`. Old non-random few-shot and legacy runs are kept aside, not in this release.
68
 
69
  ## Datasets / splits
70
 
@@ -74,29 +87,29 @@ $m_i$: human annotations per item. $R_c$: human control agreement (Krippendorff'
74
  | Split | Dataset | Task | Type | $Y$ | $N$ | $m_i$ | $R_c$ | rows | #models | #methods | Paper |
75
  |---|---|---|---|--:|--:|--|--:|--:|--:|--:|---|
76
  | **πŸ“„ Text** | | | | | | | | | | | |
77
- | `epic` | EPIC | Irony detection | ord. | 2 | 1500 | 5--8 | 0.20 | 300 | 9 | 12 | Frenda et al., 2023 |
78
- | `dices350` | [DICES-350](https://arxiv.org/abs/2306.11247) | Conversational safety | ord. | 3 | 350 | 123 | 0.20 | 300 | 9 | 7 | [Aroyo et al., 2023](https://arxiv.org/abs/2306.11247) |
79
- | `chaosnli` | [ChaosNLI](https://arxiv.org/abs/2010.03532) | NLI entailment | nom. | 3 | 1500 | 100 | 0.28 | 300 | 9 | 10 | [Nie et al., 2020](https://arxiv.org/abs/2010.03532) |
80
- | `measuring_hate_speech` | Measuring-Hate-Speech | Hate-speech severity | ord. | 3 | 1500 | 4--815 | 0.53 | 300 | 9 | 7 | Sachdeva et al., 2022 |
81
- | `loneliness` | [Loneliness](https://huggingface.co/datasets/yael-katsman/Loneliness-Causes-and-Intensity) | Loneliness intensity | ord. | 5 | 497 | 3 | 0.90 | 300 | 9 | 10 | [Katsman et al., 2024](https://huggingface.co/datasets/yael-katsman/Loneliness-Causes-and-Intensity) |
82
- | `cebab_stars` | CEBaB | Restaurant star rating | ord. | 5 | 711 | 3--4 | 0.72 | 300 | 13 | 15 | Abraham et al., 2022 |
83
- | `multipref` | [MultiPref](https://arxiv.org/abs/2410.19133) | Preference strength | ord. | 5 | 1500 | 4 | 0.28 | 300 | 9 | 7 | [Miranda et al., 2025](https://arxiv.org/abs/2410.19133) |
84
- | `scruples` | Scruples | Moral judgment | nom. | 5 | 1500 | 30 | 0.38 | 300 | 9 | 5 | Lourie et al., 2021 |
85
- | `summeval` | SummEval | Summary quality | ord. | 5 | 1600 | 3 | 0.60 | 300 | 9 | 7 | Fabbri et al., 2021 |
86
- | `d10k_prompts` | [10k-prompts](https://huggingface.co/datasets/data-is-better-together/10k_prompts_ranked) | Prompt quality rank | ord. | 5 | 391 | 3--5 | 0.35 | 300 | 13 | 9 | [Data is Better Together, 2024](https://huggingface.co/datasets/data-is-better-together/10k_prompts_ranked) |
87
- | `convabuse` | [ConvAbuse](https://arxiv.org/abs/2109.09483) | Abuse severity | ord. | 5 | 2225 | 3--8 | 0.72 | 300 | 9 | 10 | [Cercas Curry et al., 2021](https://arxiv.org/abs/2109.09483) |
88
- | `persuasion` | [Persuasion](https://www.anthropic.com/news/measuring-model-persuasiveness) | Argument persuasiveness | ord. | 7 | 1313 | 3 | 0.31 | 300 | 9 | 7 | [Durmus et al., 2024](https://www.anthropic.com/news/measuring-model-persuasiveness) |
89
  | **πŸ–ΌοΈ Image** | | | | | | | | | | | |
90
- | `t2i_dpo` | T2I-DPO | Text-to-image preference | nom. | 2 | 400 | 8 | 0.11 | 300 | 8 | 10 | Cai et al., 2025 |
91
- | `vibe_arena` | [Vibe-Arena](https://huggingface.co/datasets/datapointai/vibe-landing-page-arena) | Landing-page preference | nom. | 2 | 400 | 15 | 0.12 | 300 | 8 | 5 | [DatapointAI, 2026](https://huggingface.co/datasets/datapointai/vibe-landing-page-arena) |
92
- | `flickr8k` | Flickr8k-Expert | Caption quality | ord. | 4 | 400 | 3 | 0.87 | 300 | 8 | 11 | Hodosh et al., 2013 |
93
- | `fer_plus` | FER+ | Facial expression | nom. | 8 | 600 | 10 | 0.51 | 300 | 8 | 5 | Barsoum et al., 2016 |
94
- | `cifar10h` | CIFAR-10H | Image classification | nom. | 10 | 600 | 30 | 0.88 | 300 | 8 | 5 | Peterson et al., 2019 |
95
  | **πŸ”Š Audio** | | | | | | | | | | | |
96
- | `tts_pref` | [TTS-Preference](https://huggingface.co/datasets/datapointai/tts-human-preferences-medium) | TTS naturalness pref. | nom. | 2 | 1000 | 15 | 0.02 | 300 | 6 | 9 | [DatapointAI, 2026](https://huggingface.co/datasets/datapointai/tts-human-preferences-medium) |
97
- | `somos` | [SOMOS](https://arxiv.org/abs/2204.03040) | Speech naturalness (MOS) | ord. | 5 | 20100 | 16--27 | 0.05 | 300 | 6 | 12 | [Maniati et al., 2022](https://arxiv.org/abs/2204.03040) |
98
  | **🎬 Video** | | | | | | | | | | | |
99
- | `t2v_pref` | [T2V-Preference](https://arxiv.org/abs/2509.22799) | Text-to-video preference | nom. | 2 | 300 | 74--90 | 0.04 | 300 | 5 | 4 | [He et al., 2025](https://arxiv.org/abs/2509.22799) |
100
 
101
  ## Load
102
 
@@ -108,31 +121,3 @@ row = ds["chaosnli"][0]
108
  human = json.loads(row["human_labels"]) # [labels] or {annotator: label}
109
  llm = json.loads(row["llm"]) # {"<Model>_<method>": label}
110
  ```
111
-
112
- ## Citing the source datasets
113
-
114
- Each split is a fixed 300-item subset of an existing dataset. **Please cite the original
115
- dataset authors** when you use a split. Full BibTeX for all sources is in
116
-
117
- ### References
118
-
119
- - **EPIC** β€” Frenda et al. EPIC: Multi-Perspective Annotation of a Corpus of Irony. ACL 2023.
120
- - **DICES-350** β€” Aroyo et al. DICES Dataset: Diversity in Conversational AI Evaluation for Safety. arXiv:2306.11247, 2023. [[link](https://arxiv.org/abs/2306.11247)]
121
- - **ChaosNLI** β€” Nie, Zhou, Bansal. What Can We Learn from Collective Human Opinions on NLI Data? (ChaosNLI) arXiv:2010.03532, 2020. [[link](https://arxiv.org/abs/2010.03532)]
122
- - **Measuring-Hate-Speech** β€” Sachdeva et al. The Measuring Hate Speech Corpus. NLPerspectives@LREC 2022.
123
- - **Loneliness** β€” Katsman, Segal, Kamienney. Loneliness Causes and Intensity. Hugging Face dataset, 2024. [[link](https://huggingface.co/datasets/yael-katsman/Loneliness-Causes-and-Intensity)]
124
- - **CEBaB** β€” Abraham et al. CEBaB: Estimating the Causal Effects of Real-World Concepts on NLP Model Behavior. NeurIPS 2022.
125
- - **MultiPref** β€” Miranda et al. Hybrid Preferences: Learning to Route Instances for Human vs. AI Feedback. (MultiPref) arXiv:2410.19133, 2025. [[link](https://arxiv.org/abs/2410.19133)]
126
- - **Scruples** β€” Lourie, Le Bras, Choi. Scruples: A Corpus of Community Ethical Judgments. AAAI 2021.
127
- - **SummEval** β€” Fabbri et al. SummEval: Re-evaluating Summarization Evaluation. TACL 2021.
128
- - **10k-prompts** β€” Data is Better Together community. 10k_prompts_ranked. Hugging Face dataset, 2024. [[link](https://huggingface.co/datasets/data-is-better-together/10k_prompts_ranked)]
129
- - **ConvAbuse** β€” Cercas Curry, Abercrombie, Rieser. ConvAbuse: Data, Analysis, and Benchmarks for Nuanced Abuse Detection. arXiv:2109.09483, 2021. [[link](https://arxiv.org/abs/2109.09483)]
130
- - **Persuasion** β€” Durmus et al. Measuring the Persuasiveness of Language Models. Anthropic, 2024. [[link](https://www.anthropic.com/news/measuring-model-persuasiveness)]
131
- - **T2I-DPO** β€” Cai et al. Diffusion Self-Distillation for Zero-Shot Customized Image Generation. (T2I-DPO) CVPR 2025.
132
- - **Vibe-Arena** β€” DatapointAI. Vibe Landing Page Arena. Hugging Face dataset, 2026. [[link](https://huggingface.co/datasets/datapointai/vibe-landing-page-arena)]
133
- - **Flickr8k-Expert** β€” Hodosh, Young, Hockenmaier. Framing Image Description as a Ranking Task. (Flickr8k) JAIR 2013.
134
- - **FER+** β€” Barsoum et al. Training Deep Networks for FER with Crowd-Sourced Label Distribution. (FER+) ICMI 2016.
135
- - **CIFAR-10H** β€” Peterson et al. Human Uncertainty Makes Classification More Robust. (CIFAR-10H) ICCV 2019.
136
- - **TTS-Preference** β€” DatapointAI. TTS Human Preferences (Medium). Hugging Face dataset, 2026. [[link](https://huggingface.co/datasets/datapointai/tts-human-preferences-medium)]
137
- - **SOMOS** β€” Maniati et al. SOMOS: The Samsung Open MOS Dataset for TTS Evaluation. arXiv:2204.03040, 2022. [[link](https://arxiv.org/abs/2204.03040)]
138
- - **T2V-Preference** β€” He et al. VideoScore2: Think Before You Score in Generative Video Evaluation. (T2V) arXiv:2509.22799, 2025. [[link](https://arxiv.org/abs/2509.22799)]
 
55
  subset that all judges were run on identically (`select_items`: first 300 prompt-order items
56
  with human labels). The `N` column below is the *source* dataset size; the release uses the 300.
57
 
58
+ ## Citation
59
+
60
+ If you use this dataset or the an-alt-test, please cite our paper (anonymized for review):
61
+
62
+ ```bibtex
63
+ @misc{anonymous2026analttest,
64
+ title = {The Anonymous Alternative Annotator Test for LLM-as-a-Judge},
65
+ author = {Anonymous},
66
+ year = {2026},
67
+ note = {Under review}
68
+ }
69
+ ```
70
+
71
  ## Columns
72
 
73
  - `dataset`, `item_id`, `modality`, `label_spec`, `output_key`
 
77
  - `human_labels` β€” JSON string of the human annotations
78
  - `llm` β€” JSON string mapping `"<Model>_<method>"` β†’ label. Methods: `zs300` (zero-shot),
79
  `rf5`/`rf3`/`rf7`/`rf10`/`rf15` (random few-shot), `cot`/`cotmid` (reasoning), `gepa*`, `*_ord`,
80
+ `ens`.
81
 
82
  ## Datasets / splits
83
 
 
87
  | Split | Dataset | Task | Type | $Y$ | $N$ | $m_i$ | $R_c$ | rows | #models | #methods | Paper |
88
  |---|---|---|---|--:|--:|--|--:|--:|--:|--:|---|
89
  | **πŸ“„ Text** | | | | | | | | | | | |
90
+ | `epic` | EPIC | Irony detection | ord. | 2 | 1500 | 5--8 | 0.20 | 300 | 9 | 12 | Frenda et al. EPIC: Multi-Perspective Annotation of a Corpus of Irony. ACL 2023. |
91
+ | `dices350` | DICES-350 | Conversational safety | ord. | 3 | 350 | 123 | 0.20 | 300 | 9 | 7 | Aroyo et al. DICES Dataset: Diversity in Conversational AI Evaluation for Safety. 2023. |
92
+ | `chaosnli` | ChaosNLI | NLI entailment | nom. | 3 | 1500 | 100 | 0.28 | 300 | 9 | 10 | Nie, Zhou, Bansal. What Can We Learn from Collective Human Opinions on NLI Data? (ChaosNLI) 2020. |
93
+ | `measuring_hate_speech` | Measuring-Hate-Speech | Hate-speech severity | ord. | 3 | 1500 | 4--815 | 0.53 | 300 | 9 | 7 | Sachdeva et al. The Measuring Hate Speech Corpus. NLPerspectives@LREC 2022. |
94
+ | `loneliness` | [Loneliness](https://huggingface.co/datasets/yael-katsman/Loneliness-Causes-and-Intensity) | Loneliness intensity | ord. | 5 | 497 | 3 | 0.90 | 300 | 9 | 10 | [Katsman, Segal, Kamienney. Loneliness Causes and Intensity. Hugging Face dataset, 2024.](https://huggingface.co/datasets/yael-katsman/Loneliness-Causes-and-Intensity) |
95
+ | `cebab_stars` | CEBaB | Restaurant star rating | ord. | 5 | 711 | 3--4 | 0.72 | 300 | 13 | 15 | Abraham et al. CEBaB: Estimating the Causal Effects of Real-World Concepts on NLP Model Behavior. NeurIPS 2022. |
96
+ | `multipref` | MultiPref | Preference strength | ord. | 5 | 1500 | 4 | 0.28 | 300 | 9 | 7 | Miranda et al. Hybrid Preferences: Learning to Route Instances for Human vs. AI Feedback. (MultiPref) 2025. |
97
+ | `scruples` | Scruples | Moral judgment | nom. | 5 | 1500 | 30 | 0.38 | 300 | 9 | 5 | Lourie, Le Bras, Choi. Scruples: A Corpus of Community Ethical Judgments. AAAI 2021. |
98
+ | `summeval` | SummEval | Summary quality | ord. | 5 | 1600 | 3 | 0.60 | 300 | 9 | 7 | Fabbri et al. SummEval: Re-evaluating Summarization Evaluation. TACL 2021. |
99
+ | `d10k_prompts` | [10k-prompts](https://huggingface.co/datasets/data-is-better-together/10k_prompts_ranked) | Prompt quality rank | ord. | 5 | 391 | 3--5 | 0.35 | 300 | 13 | 9 | [Data is Better Together community. 10k_prompts_ranked. Hugging Face dataset, 2024.](https://huggingface.co/datasets/data-is-better-together/10k_prompts_ranked) |
100
+ | `convabuse` | ConvAbuse | Abuse severity | ord. | 5 | 2225 | 3--8 | 0.72 | 300 | 9 | 10 | Cercas Curry, Abercrombie, Rieser. ConvAbuse: Data, Analysis, and Benchmarks for Nuanced Abuse Detection. 2021. |
101
+ | `persuasion` | [Persuasion](https://www.anthropic.com/news/measuring-model-persuasiveness) | Argument persuasiveness | ord. | 7 | 1313 | 3 | 0.31 | 300 | 9 | 7 | [Durmus et al. Measuring the Persuasiveness of Language Models. Anthropic, 2024.](https://www.anthropic.com/news/measuring-model-persuasiveness) |
102
  | **πŸ–ΌοΈ Image** | | | | | | | | | | | |
103
+ | `t2i_dpo` | T2I-DPO | Text-to-image preference | nom. | 2 | 400 | 8 | 0.11 | 300 | 8 | 10 | Cai et al. Diffusion Self-Distillation for Zero-Shot Customized Image Generation. (T2I-DPO) CVPR 2025. |
104
+ | `vibe_arena` | [Vibe-Arena](https://huggingface.co/datasets/datapointai/vibe-landing-page-arena) | Landing-page preference | nom. | 2 | 400 | 15 | 0.12 | 300 | 8 | 5 | [DatapointAI. Vibe Landing Page Arena. Hugging Face dataset, 2026.](https://huggingface.co/datasets/datapointai/vibe-landing-page-arena) |
105
+ | `flickr8k` | Flickr8k-Expert | Caption quality | ord. | 4 | 400 | 3 | 0.87 | 300 | 8 | 11 | Hodosh, Young, Hockenmaier. Framing Image Description as a Ranking Task. (Flickr8k) JAIR 2013. |
106
+ | `fer_plus` | FER+ | Facial expression | nom. | 8 | 600 | 10 | 0.51 | 300 | 8 | 5 | Barsoum et al. Training Deep Networks for FER with Crowd-Sourced Label Distribution. (FER+) ICMI 2016. |
107
+ | `cifar10h` | CIFAR-10H | Image classification | nom. | 10 | 600 | 30 | 0.88 | 300 | 8 | 5 | Peterson et al. Human Uncertainty Makes Classification More Robust. (CIFAR-10H) ICCV 2019. |
108
  | **πŸ”Š Audio** | | | | | | | | | | | |
109
+ | `tts_pref` | [TTS-Preference](https://huggingface.co/datasets/datapointai/tts-human-preferences-medium) | TTS naturalness pref. | nom. | 2 | 1000 | 15 | 0.02 | 300 | 6 | 9 | [DatapointAI. TTS Human Preferences (Medium). Hugging Face dataset, 2026.](https://huggingface.co/datasets/datapointai/tts-human-preferences-medium) |
110
+ | `somos` | SOMOS | Speech naturalness (MOS) | ord. | 5 | 20100 | 16--27 | 0.05 | 300 | 6 | 12 | Maniati et al. SOMOS: The Samsung Open MOS Dataset for TTS Evaluation. 2022. |
111
  | **🎬 Video** | | | | | | | | | | | |
112
+ | `t2v_pref` | T2V-Preference | Text-to-video preference | nom. | 2 | 300 | 74--90 | 0.04 | 300 | 5 | 4 | He et al. VideoScore2: Think Before You Score in Generative Video Evaluation. (T2V) 2025. |
113
 
114
  ## Load
115
 
 
121
  human = json.loads(row["human_labels"]) # [labels] or {annotator: label}
122
  llm = json.loads(row["llm"]) # {"<Model>_<method>": label}
123
  ```
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
an_alt_test_example.ipynb CHANGED
The diff for this file is too large to render. See raw diff