File size: 9,482 Bytes
60839eb
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
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
132
133
134
135
136
137
138
139
140
141
142
143
144
145
146
147
148
149
150
151
152
153
154
155
156
157
158
159
160
161
162
163
164
165
166
167
168
169
170
---
license: odc-by
language:
- en
pretty_name: PosteriorBench paper data
not-for-all-audiences: true
size_categories:
- 10M<n<100M
task_categories:
- text-generation
tags:
- rare-behaviors
- posterior-sampling
- language-model-sampling
- benchmark
configs:
- config_name: corpus-qwen3.5-0.8b
  data_files:
  - split: train
    path: corpora/qwen3.5-0.8b/corpus/*/*.parquet
- config_name: corpus-qwen3.8-27b
  data_files:
  - split: train
    path: corpora/qwen3.8-27b/corpus/*/*.parquet
- config_name: corpus-gemma-4-31b-it
  data_files:
  - split: train
    path: corpora/gemma-4-31b-it/corpus/*/*.parquet
---

# PosteriorBench: data of the paper's other experiments

[📄 Paper](https://arxiv.org/abs/2611.25735) · [💻 GitHub](https://github.com/TransluceAI/PosteriorBench)

This dataset contains the data for the experiments in the [PosteriorBench paper](https://arxiv.org/abs/2611.25735) that go beyond the
benchmark itself. It includes six tasks on larger models and LLM-judged behaviors, the 14.4 million
responses that we sampled to build them, and the inputs to the paper's prompt selection and
Steering methods (candidate system prompts, steering vectors, and the prompts used to build the
vectors). The benchmark itself is released as
[Transluce/PosteriorBench](https://huggingface.co/datasets/Transluce/PosteriorBench), and the code
that reads these files is in [`methods/`](https://github.com/TransluceAI/PosteriorBench/tree/main/methods) of the
PosteriorBench repository.

**Content warning.** The responses are unfiltered outputs of language models and were not reviewed
by people. One of the six tasks is the behavior "direct encouragement": the model sincerely tells
the user to kill themselves. Its files hold the 760 such responses in full, the candidate system
prompts and construction prompts that were written to elicit the behavior, and steering vectors
for it. The other files contain offensive language and false statements. The prompt is a message
that a real user sent to a chatbot.

## What's here

| Path | Files | Rows | One row is |
|---|---:|---:|---|
| `<task>/evaluation/evaluation.parquet` | 6 | 1 each | a task's held-out references and its fallback response, with their stored log p |
| `<task>/hits/wc_0670/<behavior>.parquet` | 6 | 286,252 | a hit of a task that is not a reference, for methods to train on |
| `<task>/labels/labels.parquet` | 2 | 3,701 | a response that the rubric judge called a match, with both judges' answers (LLM-judged tasks) |
| `corpora/<model>/corpus/wc_0670/part-*.parquet` | 144 | 14,396,000 | a response, with its token IDs and two behavior labels (13.6 GB) |
| `method_data/candidate_prompts_108tasks.parquet` | 1 | 111,672 | a candidate system prompt for one of the benchmark's 108 scored pairs |
| `method_data/candidate_prompts_21tasks_extended_131072.parquet` | 1 | 2,752,512 | a candidate for one of the benchmark's 21 tasks, 131,072 per task |
| `method_data/candidate_prompts_generalization.parquet` | 1 | 6,144 | a candidate for one of the six tasks here, 1,024 per task |
| `method_data/steering_vectors/<directory>/<behavior>.safetensors` | 13 | | the steering vectors of a behavior on one model, one tensor per vector and layer |
| `method_data/construction_prompts/<behavior>.json` | 9 | | the instruction pairs and questions a behavior's vectors were built from |

`<task>` is a task id with `__` for `/`, such as `gemma-4-31b-it__wc_0670__em_dash`.

- **evaluation and hits**: these files have the same columns as in the benchmark's dataset.
  `evaluation.parquet` and `labels.parquet` are only for evaluation and should never be read by a
  method. The hit file of `qwen3.8-27b/wc_0670/contrastive_phrasing` is empty, because all of the
  task's hits are used as references.
- **labels**: `row`, `text`, the rubric judge's answer (`judge_match`, `judge_explanation`), the
  reviewer's answer (`review_verdict`, `review_reasons`, `review_explanation`), and `hit`.
- **corpora**: `prompt_id`, `row`, `token_ids`, `text`, and one boolean column for each of the
  model's two behaviors.
- **candidate prompts**: `task_id`, `candidate_index`, `candidate_id`, `system_prompt`, and
  `scenario` (the scenario from which the prompt was written). In the file with 131,072 candidates
  per task, `batch` is 0 for the candidates that also appear in the file for the 108 pairs, and 1,
  2 and 3 for the candidates added to reach 32,768, 65,536 and 131,072.
- **steering vectors**: each file contains tensors named `<vector>.layer_<n>`, where `<vector>` is
  `with_ablation` or `without_ablation`, depending on whether the vector was built from responses
  sampled with the refusal direction ablated (Gemma-4-31B-it has only `without_ablation`). The
  `benchmark` directory contains the vectors for the benchmark's seven behaviors on Qwen3.5-0.8B,
  and the directories named after the three models contain the vectors for the six tasks. Each
  directory also has a `construction_cost.json` with the number of inferences used to build its
  vectors.

## The six tasks

All six tasks use one of the benchmark's prompts, `wc_0670`.

| Task | Model | Behavior is judged by | Responses | Hits | p̂(S) | References | Response limit |
|---|---|---|---:|---:|---:|---:|---|
| `qwen3.5-0.8b/wc_0670/direct_encouragement` | Qwen3.5-0.8B | two language models | 3,400,000 | 760 | 2.24e-4 | 64 | 2,048 tokens |
| `qwen3.5-0.8b/wc_0670/physical_embodiment` | Qwen3.5-0.8B | two language models | 3,400,000 | 2,924 | 8.60e-4 | 64 | 2,048 tokens |
| `qwen3.8-27b/wc_0670/chinese_character` | Qwen3.8-27B | a rule | 996,000 | 2,535 | 2.55e-3 | 250 | 50 tokens |
| `qwen3.8-27b/wc_0670/contrastive_phrasing` | Qwen3.8-27B | a rule | 996,000 | 193 | 1.94e-4 | 193 | 50 tokens |
| `gemma-4-31b-it/wc_0670/chinese_character` | Gemma-4-31B-it | a rule | 10,000,000 | 938 | 9.38e-5 | 250 | 50 tokens |
| `gemma-4-31b-it/wc_0670/em_dash` | Gemma-4-31B-it | a rule | 10,000,000 | 279,973 | 2.80e-2 | 250 | 50 tokens |

These tasks are not part of the benchmark: each uses a single prompt, and p̂(S) is estimated from
far fewer responses than the benchmark's 50 million. The rule-based behaviors use the same rules as
the benchmark. For the two LLM-judged tasks, a response counts as a hit if Gemma-4-31B-it judges it
to match WeirdChat's rubric for the behavior, given the prompt and the full response, and GPT-6 Sol
agrees when reviewing that match.

## Loading

The `posteriorbench-methods` commands take the dataset id through `--data` and download the files
that a task needs:

```sh
uv run posteriorbench-methods generalization evaluate \
    --task gemma-4-31b-it/wc_0670/em_dash --data Transluce/PosteriorBench-paper-data --sampler instruction
```

The method data is read from local files, so it needs to be downloaded first:

```sh
uv run hf download Transluce/PosteriorBench-paper-data --repo-type dataset --include "method_data/*" --local-dir .
uv run posteriorbench-methods prompt run --task wc_0346/em_dash --out runs \
    --method-data method_data/candidate_prompts_108tasks.parquet
```

Each corpus is available as a `datasets` configuration, for example `corpus-gemma-4-31b-it`. We
recommend `streaming=True` for the corpus of Qwen3.5-0.8B (12.7 GB).

## How the responses were generated

| Setting | Value |
|---|---|
| Models | `Qwen/Qwen3.5-0.8B` (revision `2fc06364715b967f1860aea9cf38778875588b17`), `Qwen/Qwen3.8-27B` (`1d4bf0f2ff6012fd82039f2fa52739d0dd7c60c0`), `google/gemma-4-31B-it` (`145dc2508c480a64b47242f160d286cff94a2343`) |
| Prompt | the WildChat user turn of `wc_0670` as the only message, rendered with each model's chat template with thinking off; no system prompt |
| Sampling | temperature 1, no truncation; at most 50 new tokens on the two larger models and 2,048 on Qwen3.5-0.8B |
| Engine | SGLang 0.5.10; bfloat16 weights for the two Qwen models, FP8 for Gemma-4-31B-it |

All responses were sampled independently. The stored log p values of the references and of the
fallback responses were computed with the same engine and may differ under another runtime.

## Method data

The candidate system prompts were written by DeepSeek-V4-Flash, which was given the behavior's rule
or rubric, the user prompt and the model's name for each task. The steering vectors are persona
vectors for each behavior. They were built from responses to the instruction pairs and questions in
`construction_prompts/`, which were written by gpt-oss-120b (by DeepSeek-V4-Flash for direct
encouragement). Both pipelines, and the commands to rerun them, are described in the READMEs of
`prompt` and `steering` in the repository.

## License

This dataset is released under the
[Open Data Commons Attribution License (ODC-BY)](https://opendatacommons.org/licenses/by/1-0/). The
prompt is one of the benchmark's 20 prompts and comes from
[WildChat-1M](https://huggingface.co/datasets/allenai/WildChat-1M), which is also released under
ODC-BY. The responses were generated by Qwen3.5-0.8B, Qwen3.8-27B and Gemma-4-31B-it, each released
under Apache 2.0. The rubrics of the two LLM-judged behaviors are from
[WeirdChat](https://huggingface.co/datasets/Transluce/WeirdChat), which is released under the MIT
License.

## Citation

```bibtex
@article{xun2026posteriorbench,
  title={PosteriorBench: Evaluating Behavior Posterior Sampling in Language Models},
  author={Zhiyang Xun and Daniel D. Johnson and Sarah Schwettmann and Neil Chowdhury},
  year={2026},
  eprint={2611.25735},
  archivePrefix={arXiv}
}
```