Datasets:
Commit ·
d975d30
0
Parent(s):
PosteriorBench dataset v1.0.0
Browse filesThis view is limited to 50 files because it contains too many changes. See raw diff
- .gitattributes +60 -0
- README.md +214 -0
- corpus/wc_0055/part-00000.parquet +3 -0
- corpus/wc_0055/part-00001.parquet +3 -0
- corpus/wc_0055/part-00002.parquet +3 -0
- corpus/wc_0055/part-00003.parquet +3 -0
- corpus/wc_0055/part-00004.parquet +3 -0
- corpus/wc_0055/part-00005.parquet +3 -0
- corpus/wc_0055/part-00006.parquet +3 -0
- corpus/wc_0055/part-00007.parquet +3 -0
- corpus/wc_0055/part-00008.parquet +3 -0
- corpus/wc_0055/part-00009.parquet +3 -0
- corpus/wc_0055/part-00010.parquet +3 -0
- corpus/wc_0055/part-00011.parquet +3 -0
- corpus/wc_0055/part-00012.parquet +3 -0
- corpus/wc_0055/part-00013.parquet +3 -0
- corpus/wc_0055/part-00014.parquet +3 -0
- corpus/wc_0055/part-00015.parquet +3 -0
- corpus/wc_0055/part-00016.parquet +3 -0
- corpus/wc_0055/part-00017.parquet +3 -0
- corpus/wc_0055/part-00018.parquet +3 -0
- corpus/wc_0055/part-00019.parquet +3 -0
- corpus/wc_0055/part-00020.parquet +3 -0
- corpus/wc_0055/part-00021.parquet +3 -0
- corpus/wc_0055/part-00022.parquet +3 -0
- corpus/wc_0055/part-00023.parquet +3 -0
- corpus/wc_0055/part-00024.parquet +3 -0
- corpus/wc_0055/part-00025.parquet +3 -0
- corpus/wc_0055/part-00026.parquet +3 -0
- corpus/wc_0055/part-00027.parquet +3 -0
- corpus/wc_0055/part-00028.parquet +3 -0
- corpus/wc_0055/part-00029.parquet +3 -0
- corpus/wc_0055/part-00030.parquet +3 -0
- corpus/wc_0055/part-00031.parquet +3 -0
- corpus/wc_0055/part-00032.parquet +3 -0
- corpus/wc_0055/part-00033.parquet +3 -0
- corpus/wc_0055/part-00034.parquet +3 -0
- corpus/wc_0055/part-00035.parquet +3 -0
- corpus/wc_0055/part-00036.parquet +3 -0
- corpus/wc_0055/part-00037.parquet +3 -0
- corpus/wc_0055/part-00038.parquet +3 -0
- corpus/wc_0055/part-00039.parquet +3 -0
- corpus/wc_0055/part-00040.parquet +3 -0
- corpus/wc_0055/part-00041.parquet +3 -0
- corpus/wc_0055/part-00042.parquet +3 -0
- corpus/wc_0055/part-00043.parquet +3 -0
- corpus/wc_0055/part-00044.parquet +3 -0
- corpus/wc_0055/part-00045.parquet +3 -0
- corpus/wc_0055/part-00046.parquet +3 -0
- corpus/wc_0055/part-00047.parquet +3 -0
.gitattributes
ADDED
|
@@ -0,0 +1,60 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
*.7z filter=lfs diff=lfs merge=lfs -text
|
| 2 |
+
*.arrow filter=lfs diff=lfs merge=lfs -text
|
| 3 |
+
*.avro filter=lfs diff=lfs merge=lfs -text
|
| 4 |
+
*.bin filter=lfs diff=lfs merge=lfs -text
|
| 5 |
+
*.bz2 filter=lfs diff=lfs merge=lfs -text
|
| 6 |
+
*.ckpt filter=lfs diff=lfs merge=lfs -text
|
| 7 |
+
*.ftz filter=lfs diff=lfs merge=lfs -text
|
| 8 |
+
*.gz filter=lfs diff=lfs merge=lfs -text
|
| 9 |
+
*.h5 filter=lfs diff=lfs merge=lfs -text
|
| 10 |
+
*.joblib filter=lfs diff=lfs merge=lfs -text
|
| 11 |
+
*.lfs.* filter=lfs diff=lfs merge=lfs -text
|
| 12 |
+
*.lz4 filter=lfs diff=lfs merge=lfs -text
|
| 13 |
+
*.mds filter=lfs diff=lfs merge=lfs -text
|
| 14 |
+
*.mlmodel filter=lfs diff=lfs merge=lfs -text
|
| 15 |
+
*.model filter=lfs diff=lfs merge=lfs -text
|
| 16 |
+
*.msgpack filter=lfs diff=lfs merge=lfs -text
|
| 17 |
+
*.npy filter=lfs diff=lfs merge=lfs -text
|
| 18 |
+
*.npz filter=lfs diff=lfs merge=lfs -text
|
| 19 |
+
*.onnx filter=lfs diff=lfs merge=lfs -text
|
| 20 |
+
*.ot filter=lfs diff=lfs merge=lfs -text
|
| 21 |
+
*.parquet filter=lfs diff=lfs merge=lfs -text
|
| 22 |
+
*.pb filter=lfs diff=lfs merge=lfs -text
|
| 23 |
+
*.pickle filter=lfs diff=lfs merge=lfs -text
|
| 24 |
+
*.pkl filter=lfs diff=lfs merge=lfs -text
|
| 25 |
+
*.pt filter=lfs diff=lfs merge=lfs -text
|
| 26 |
+
*.pth filter=lfs diff=lfs merge=lfs -text
|
| 27 |
+
*.rar filter=lfs diff=lfs merge=lfs -text
|
| 28 |
+
*.safetensors filter=lfs diff=lfs merge=lfs -text
|
| 29 |
+
saved_model/**/* filter=lfs diff=lfs merge=lfs -text
|
| 30 |
+
*.tar.* filter=lfs diff=lfs merge=lfs -text
|
| 31 |
+
*.tar filter=lfs diff=lfs merge=lfs -text
|
| 32 |
+
*.tflite filter=lfs diff=lfs merge=lfs -text
|
| 33 |
+
*.tgz filter=lfs diff=lfs merge=lfs -text
|
| 34 |
+
*.wasm filter=lfs diff=lfs merge=lfs -text
|
| 35 |
+
*.xz filter=lfs diff=lfs merge=lfs -text
|
| 36 |
+
*.zip filter=lfs diff=lfs merge=lfs -text
|
| 37 |
+
*.zst filter=lfs diff=lfs merge=lfs -text
|
| 38 |
+
*tfevents* filter=lfs diff=lfs merge=lfs -text
|
| 39 |
+
# Audio files - uncompressed
|
| 40 |
+
*.pcm filter=lfs diff=lfs merge=lfs -text
|
| 41 |
+
*.sam filter=lfs diff=lfs merge=lfs -text
|
| 42 |
+
*.raw filter=lfs diff=lfs merge=lfs -text
|
| 43 |
+
# Audio files - compressed
|
| 44 |
+
*.aac filter=lfs diff=lfs merge=lfs -text
|
| 45 |
+
*.flac filter=lfs diff=lfs merge=lfs -text
|
| 46 |
+
*.mp3 filter=lfs diff=lfs merge=lfs -text
|
| 47 |
+
*.ogg filter=lfs diff=lfs merge=lfs -text
|
| 48 |
+
*.wav filter=lfs diff=lfs merge=lfs -text
|
| 49 |
+
# Image files - uncompressed
|
| 50 |
+
*.bmp filter=lfs diff=lfs merge=lfs -text
|
| 51 |
+
*.gif filter=lfs diff=lfs merge=lfs -text
|
| 52 |
+
*.png filter=lfs diff=lfs merge=lfs -text
|
| 53 |
+
*.tiff filter=lfs diff=lfs merge=lfs -text
|
| 54 |
+
# Image files - compressed
|
| 55 |
+
*.jpg filter=lfs diff=lfs merge=lfs -text
|
| 56 |
+
*.jpeg filter=lfs diff=lfs merge=lfs -text
|
| 57 |
+
*.webp filter=lfs diff=lfs merge=lfs -text
|
| 58 |
+
# Video files - compressed
|
| 59 |
+
*.mp4 filter=lfs diff=lfs merge=lfs -text
|
| 60 |
+
*.webm filter=lfs diff=lfs merge=lfs -text
|
README.md
ADDED
|
@@ -0,0 +1,214 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
---
|
| 2 |
+
license: odc-by
|
| 3 |
+
language:
|
| 4 |
+
- en
|
| 5 |
+
pretty_name: PosteriorBench
|
| 6 |
+
not-for-all-audiences: true
|
| 7 |
+
size_categories:
|
| 8 |
+
- 1B<n<10B
|
| 9 |
+
task_categories:
|
| 10 |
+
- text-generation
|
| 11 |
+
tags:
|
| 12 |
+
- rare-behaviors
|
| 13 |
+
- posterior-sampling
|
| 14 |
+
- language-model-sampling
|
| 15 |
+
- benchmark
|
| 16 |
+
configs:
|
| 17 |
+
- config_name: tasks
|
| 18 |
+
default: true
|
| 19 |
+
data_files:
|
| 20 |
+
- split: train
|
| 21 |
+
path: tasks/tasks.parquet
|
| 22 |
+
- config_name: corpus
|
| 23 |
+
data_files:
|
| 24 |
+
- split: train
|
| 25 |
+
path: corpus/*/*.parquet
|
| 26 |
+
---
|
| 27 |
+
|
| 28 |
+
# PosteriorBench
|
| 29 |
+
|
| 30 |
+
[📄 Paper](https://arxiv.org/abs/2611.25735) · [💻 GitHub](https://github.com/TransluceAI/PosteriorBench)
|
| 31 |
+
|
| 32 |
+
PosteriorBench measures how well a method can sample the responses of a language model that show a
|
| 33 |
+
rare behavior. To build it, we sampled one billion responses from
|
| 34 |
+
[Qwen3.5-0.8B](https://huggingface.co/Qwen/Qwen3.5-0.8B) to 20 WildChat prompts (50 million per
|
| 35 |
+
prompt) and labeled each response for seven behaviors, such as using an em dash or mentioning a
|
| 36 |
+
goblin. The responses that show a behavior are exact samples from the model conditioned on that
|
| 37 |
+
behavior. Some of them are given to methods as training data, and others are held out as references
|
| 38 |
+
for evaluation. We include all one billion responses with their token IDs and labels, together with
|
| 39 |
+
the files used by the `posteriorbench` Python package. See the [paper](https://arxiv.org/abs/2611.25735) for details and the
|
| 40 |
+
[GitHub repository](https://github.com/TransluceAI/PosteriorBench) for the package.
|
| 41 |
+
|
| 42 |
+
**Content warning.** The responses are unfiltered outputs of a small language model and were not
|
| 43 |
+
reviewed by people. They contain offensive language (one of the behaviors is profanity) and racial
|
| 44 |
+
slurs; responses that discuss suicide or self-harm, a few of which urge the reader to kill
|
| 45 |
+
themselves; and false statements. The prompts are messages that real users sent to a chatbot.
|
| 46 |
+
|
| 47 |
+
## What's here
|
| 48 |
+
|
| 49 |
+
| Path | Files | Rows | One row is |
|
| 50 |
+
|---|---:|---:|---|
|
| 51 |
+
| `tasks/tasks.parquet` | 1 | 140 | a prompt–behavior pair: prompt, rule, hit count, p̂(S) |
|
| 52 |
+
| `hits/<prompt_id>/<behavior>.parquet` | 108 | 45,646,759 | a hit of a scored pair, for methods to train on (3.42 GB in all) |
|
| 53 |
+
| `evaluation/evaluation.parquet` | 1 | 108 | a scored pair's 250 held-out references and its fallback response (5.5 MB) |
|
| 54 |
+
| `corpus/<prompt_id>/part-*.parquet` | 1,000 | 1,000,000,000 | a response, with its token IDs and the seven labels (146 GB) |
|
| 55 |
+
|
| 56 |
+
Pairing every prompt with every behavior gives 140 prompt–behavior pairs. A pair's hits are the
|
| 57 |
+
responses to the prompt that show the behavior. The fraction of hits among the prompt's 50 million
|
| 58 |
+
responses is p̂(S), our estimate of the behavior's probability. We score methods on the 108 pairs
|
| 59 |
+
that have at least 500 hits; their p̂(S) ranges from 1e-5 to 0.1. Each scored pair has 250
|
| 60 |
+
references (hits held out for evaluation) and a fallback response. The benchmark's 21 tasks are the
|
| 61 |
+
scored pairs with the lowest, median and highest p̂(S) for each behavior.
|
| 62 |
+
|
| 63 |
+
- **tasks**: `task_id` (`<prompt_id>/<behavior>`), the prompt (`prompt_id`, `prompt`, its rendered
|
| 64 |
+
`prompt_token_ids`, and the `wildchat_conversation_hash` of the WildChat-1M conversation it comes
|
| 65 |
+
from), the `behavior` and its `rule`, and `probability`, which is p̂(S) =
|
| 66 |
+
`hit_count / num_responses`. `scored` marks the 108 scored pairs and `benchmark_task` marks the
|
| 67 |
+
benchmark's 21 tasks.
|
| 68 |
+
- **hits**: one file per scored pair, containing all of the pair's hits except its references. Each
|
| 69 |
+
row has the hit's `row` in the prompt's corpus, its `token_ids`, and a `reference_duplicate` flag
|
| 70 |
+
(see [Held-out references](#held-out-references)).
|
| 71 |
+
- **evaluation**: for each scored pair (`task_id`), the `references` (`row`, `token_ids`, `text`)
|
| 72 |
+
and their `reference_logp`, and a fallback response (`fallback_text`, `fallback_token_ids`,
|
| 73 |
+
`fallback_logp`). The fallback is a fixed response that shows the behavior, written in advance by
|
| 74 |
+
a language model; a sampler whose hit rate is below 10% receives the fallback's scores. This file
|
| 75 |
+
is only for evaluation and should never be read by a method.
|
| 76 |
+
- **corpus**: `prompt_id`, `row` (0 to 49,999,999 within the prompt), `token_ids`, `text` (the
|
| 77 |
+
decoded response) and one boolean column per behavior. Within each prompt the rows are in a
|
| 78 |
+
random order, so any range of rows is a uniform random sample of that prompt's responses.
|
| 79 |
+
|
| 80 |
+
## Behaviors
|
| 81 |
+
|
| 82 |
+
Each behavior is defined by an exact rule on the decoded text `y` of a response. The rule is stored
|
| 83 |
+
as a Python expression in the `rule` column of the task table. Each rule is a substring or a
|
| 84 |
+
regular expression that stands in for a broader behavior; for example, `profanity` is a single
|
| 85 |
+
substring.
|
| 86 |
+
|
| 87 |
+
| Label | Rule | Scored pairs |
|
| 88 |
+
|---|---|---:|
|
| 89 |
+
| `em_dash` | `"—" in y` | 20 |
|
| 90 |
+
| `chinese_character` | `re.search(r"[㐀-䶿一-鿿]", y)` | 20 |
|
| 91 |
+
| `chatgpt_mention` | `"chatgpt" in y.lower()` | 12 |
|
| 92 |
+
| `profanity` | `"fuck" in y.lower()` | 9 |
|
| 93 |
+
| `goblin_mention` | `"goblin" in y.lower()` | 8 |
|
| 94 |
+
| `contrastive_phrasing` | `re.search(r"\b(?:it\|this\|that)(?:['’]s not\| is not\| isn['���]t) [^.!?\n]{2,60}?[,;:—–-]+\s*(?:it\|this\|that)(?:['’]s\| is)\b", y, re.IGNORECASE)` | 19 |
|
| 95 |
+
| `word_repetition` | `re.search(r"\b([^\W\d_]{1,20})(?:\s+\1\b){2,}", y, re.IGNORECASE)` | 20 |
|
| 96 |
+
|
| 97 |
+
## Loading
|
| 98 |
+
|
| 99 |
+
The task table is the default configuration, and the corpus is available as the `corpus`
|
| 100 |
+
configuration. The hit files and the evaluation file can be read directly as Parquet files.
|
| 101 |
+
|
| 102 |
+
```python
|
| 103 |
+
import pandas as pd
|
| 104 |
+
from datasets import load_dataset
|
| 105 |
+
|
| 106 |
+
tasks = load_dataset("Transluce/PosteriorBench", "tasks", revision="v1.0.0", split="train")
|
| 107 |
+
hits = pd.read_parquet(
|
| 108 |
+
"hf://datasets/Transluce/PosteriorBench@v1.0.0/hits/wc_1853/goblin_mention.parquet"
|
| 109 |
+
)
|
| 110 |
+
```
|
| 111 |
+
|
| 112 |
+
To download everything except the corpus (3.4 GB):
|
| 113 |
+
|
| 114 |
+
```sh
|
| 115 |
+
hf download Transluce/PosteriorBench --repo-type dataset --revision v1.0.0 \
|
| 116 |
+
--include "tasks/*" --include "evaluation/*" --include "hits/*" --local-dir PosteriorBench
|
| 117 |
+
```
|
| 118 |
+
|
| 119 |
+
We recommend loading the `corpus` configuration with `streaming=True`; otherwise `datasets`
|
| 120 |
+
downloads all 146 GB.
|
| 121 |
+
|
| 122 |
+
To run the benchmark, use the `posteriorbench` package. It downloads only the hit files that are
|
| 123 |
+
requested, and it charges a method 1/p̂(S) inferences for each distinct hit it uses, which is the
|
| 124 |
+
expected number of samples from the model needed to find one hit by rejection sampling.
|
| 125 |
+
|
| 126 |
+
```python
|
| 127 |
+
import posteriorbench as pb
|
| 128 |
+
|
| 129 |
+
bench = pb.Benchmark()
|
| 130 |
+
task = bench.task("wc_1853/goblin_mention") # prompt, behavior, rule, p̂(S)
|
| 131 |
+
ledger = pb.Ledger()
|
| 132 |
+
hits = bench.corpus(task, ledger).take_first(64) # 64 training hits, charged 64 / p̂(S)
|
| 133 |
+
```
|
| 134 |
+
|
| 135 |
+
## Held-out references
|
| 136 |
+
|
| 137 |
+
The 250 references of each pair are a uniform random sample of its hits, and they are excluded from
|
| 138 |
+
the pair's hit file. A few hits have exactly the same token IDs as a reference. These are kept in
|
| 139 |
+
the hit file with `reference_duplicate` set to true, and the package excludes them by default.
|
| 140 |
+
|
| 141 |
+
Note that references are held out per pair: a reference of one pair may appear in the hit file of
|
| 142 |
+
another pair on the same prompt, so a sampler for a pair should be built only from that pair's own
|
| 143 |
+
hits. If you read the corpus directly, you need to exclude the pair's reference rows yourself.
|
| 144 |
+
|
| 145 |
+
## How the responses were generated
|
| 146 |
+
|
| 147 |
+
| Setting | Value |
|
| 148 |
+
|---|---|
|
| 149 |
+
| Model | `Qwen/Qwen3.5-0.8B` (revision `2fc06364715b967f1860aea9cf38778875588b17`), bfloat16 weights |
|
| 150 |
+
| Prompts | 20 drawn at random from the first user turns in [WildChat-1M](https://huggingface.co/datasets/allenai/WildChat-1M) (revision `7d6490e462285cf85d91eabea0f9a954fbddcd1f`) that have 10–100 characters, whose letters are all ASCII, that WildChat labels as English and that it does not mark as redacted |
|
| 151 |
+
| Prompt format | the user turn as the only message, rendered with the model's chat template with thinking off (`enable_thinking=False`); no system prompt |
|
| 152 |
+
| Sampling | temperature 1, top-p 1, no top-k or min-p truncation; at most 50 new tokens; generation stops at `<\|im_end\|>` or `<\|endoftext\|>` |
|
| 153 |
+
| Engine | SGLang 0.5.10 |
|
| 154 |
+
|
| 155 |
+
All responses were sampled independently, and they cannot be regenerated from a seed. Responses are
|
| 156 |
+
truncated at 50 tokens, and 98.85% of them reach this limit. A small number of responses are
|
| 157 |
+
irregular: a response that was truncated inside a multi-byte character ends in U+FFFD, some
|
| 158 |
+
responses decode to an empty string, and some contain `<think>` or `</think>` even though thinking
|
| 159 |
+
was off. Each of these cases affects fewer than 1% of a prompt's responses. The log p values in the
|
| 160 |
+
evaluation file were computed with the same engine and may differ under another runtime.
|
| 161 |
+
|
| 162 |
+
## Adding a behavior
|
| 163 |
+
|
| 164 |
+
A new behavior can be added without sampling from the model again. Any rule on the text defines a
|
| 165 |
+
new pair: its probability is the fraction of rows that match, and the matching rows are samples
|
| 166 |
+
from the model conditioned on the rule.
|
| 167 |
+
|
| 168 |
+
```python
|
| 169 |
+
import re
|
| 170 |
+
from datasets import load_dataset
|
| 171 |
+
|
| 172 |
+
rows = load_dataset(
|
| 173 |
+
"Transluce/PosteriorBench", data_files="corpus/wc_0670/*.parquet", split="train",
|
| 174 |
+
revision="v1.0.0", streaming=True, columns=["text"],
|
| 175 |
+
)
|
| 176 |
+
hits = sum(re.search(r"\bdragons?\b", row["text"], re.IGNORECASE) is not None for row in rows)
|
| 177 |
+
probability = hits / 50_000_000
|
| 178 |
+
```
|
| 179 |
+
|
| 180 |
+
With 50 million responses per prompt, a behavior with probability 1e-5 has about 500 hits. The
|
| 181 |
+
package's repository includes a tool, `tools/add_behavior.py`, that builds the files for a new
|
| 182 |
+
pair.
|
| 183 |
+
|
| 184 |
+
## Limitations
|
| 185 |
+
|
| 186 |
+
- The dataset covers one model and one sampling setting: Qwen3.5-0.8B, a small model, with thinking
|
| 187 |
+
off, at temperature 1.
|
| 188 |
+
- There are 20 prompts, all short English first user turns from WildChat, and responses are at most
|
| 189 |
+
50 tokens long.
|
| 190 |
+
- p̂(S) is an estimate from 50 million responses. Its estimated relative standard error is below 5%
|
| 191 |
+
on the scored pairs. The 32 pairs with fewer than 500 hits are not scored and have no references.
|
| 192 |
+
|
| 193 |
+
## License
|
| 194 |
+
|
| 195 |
+
This dataset is released under the
|
| 196 |
+
[Open Data Commons Attribution License (ODC-BY)](https://opendatacommons.org/licenses/by/1-0/). The
|
| 197 |
+
prompts come from WildChat-1M, which is also released under ODC-BY; please cite WildChat when you
|
| 198 |
+
use them. The responses were generated by Qwen3.5-0.8B, which is released by Qwen under Apache 2.0.
|
| 199 |
+
|
| 200 |
+
## Changelog
|
| 201 |
+
|
| 202 |
+
- **v1.0.0**: Initial release, read by `posteriorbench` 1.0.0.
|
| 203 |
+
|
| 204 |
+
## Citation
|
| 205 |
+
|
| 206 |
+
```bibtex
|
| 207 |
+
@article{xun2026posteriorbench,
|
| 208 |
+
title={PosteriorBench: Evaluating Behavior Posterior Sampling in Language Models},
|
| 209 |
+
author={Zhiyang Xun and Daniel D. Johnson and Sarah Schwettmann and Neil Chowdhury},
|
| 210 |
+
year={2026},
|
| 211 |
+
eprint={2611.25735},
|
| 212 |
+
archivePrefix={arXiv}
|
| 213 |
+
}
|
| 214 |
+
```
|
corpus/wc_0055/part-00000.parquet
ADDED
|
@@ -0,0 +1,3 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
version https://git-lfs.github.com/spec/v1
|
| 2 |
+
oid sha256:de24daf3be4dba038515051f45ce791ddb5cecfc904462fb515bd83683b7ee74
|
| 3 |
+
size 150923626
|
corpus/wc_0055/part-00001.parquet
ADDED
|
@@ -0,0 +1,3 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
version https://git-lfs.github.com/spec/v1
|
| 2 |
+
oid sha256:0dc5109f49865e7024116ff85baf531118b39d4632d722b8e7edf9b9c7deae4a
|
| 3 |
+
size 151012538
|
corpus/wc_0055/part-00002.parquet
ADDED
|
@@ -0,0 +1,3 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
version https://git-lfs.github.com/spec/v1
|
| 2 |
+
oid sha256:271301550f3658c60e316af08d75f84ffbceee3487a89b902e8c94a6a02dbe33
|
| 3 |
+
size 150904683
|
corpus/wc_0055/part-00003.parquet
ADDED
|
@@ -0,0 +1,3 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
version https://git-lfs.github.com/spec/v1
|
| 2 |
+
oid sha256:ddcd66b02722417532fbb2664cd943bb173149091cb62fa1869eff0c97f9a736
|
| 3 |
+
size 151003134
|
corpus/wc_0055/part-00004.parquet
ADDED
|
@@ -0,0 +1,3 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
version https://git-lfs.github.com/spec/v1
|
| 2 |
+
oid sha256:57d547a5019afaad3460fab3b34e42a14bac008613ac3bd92e36153cfbf7bf58
|
| 3 |
+
size 150912902
|
corpus/wc_0055/part-00005.parquet
ADDED
|
@@ -0,0 +1,3 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
version https://git-lfs.github.com/spec/v1
|
| 2 |
+
oid sha256:cd8486fb304719ab78192edb5ef6a7e0fba280cafb6a04e5545033048cebf4d6
|
| 3 |
+
size 150956836
|
corpus/wc_0055/part-00006.parquet
ADDED
|
@@ -0,0 +1,3 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
version https://git-lfs.github.com/spec/v1
|
| 2 |
+
oid sha256:3a429841c62c9db1760ef5bb800152089fe0a9daf33587ff88eb5f3ba29110dc
|
| 3 |
+
size 150887173
|
corpus/wc_0055/part-00007.parquet
ADDED
|
@@ -0,0 +1,3 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
version https://git-lfs.github.com/spec/v1
|
| 2 |
+
oid sha256:177e8380c0b3a98880efb77f7380854daabb6d7c71b887e938a265b7f729f81b
|
| 3 |
+
size 150906648
|
corpus/wc_0055/part-00008.parquet
ADDED
|
@@ -0,0 +1,3 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
version https://git-lfs.github.com/spec/v1
|
| 2 |
+
oid sha256:b7699dcf2835cc0a72ca8dd13ca9a046975cb739bd9bb535228b590871fadb95
|
| 3 |
+
size 150898902
|
corpus/wc_0055/part-00009.parquet
ADDED
|
@@ -0,0 +1,3 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
version https://git-lfs.github.com/spec/v1
|
| 2 |
+
oid sha256:0108a3446f49c80b3bcf9b0fa6acac24bf0b2dd40385099b8ceba19badd05ce4
|
| 3 |
+
size 151054065
|
corpus/wc_0055/part-00010.parquet
ADDED
|
@@ -0,0 +1,3 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
version https://git-lfs.github.com/spec/v1
|
| 2 |
+
oid sha256:5b21052dd97ffea8604e2d613c926b1fa75906d1e55691d7c77b7d51bbdf56ae
|
| 3 |
+
size 150924251
|
corpus/wc_0055/part-00011.parquet
ADDED
|
@@ -0,0 +1,3 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
version https://git-lfs.github.com/spec/v1
|
| 2 |
+
oid sha256:f57cedbd8789d8de4afdb5ae273752cdd17caa17c6480d3e41005e1ec8724ada
|
| 3 |
+
size 150881199
|
corpus/wc_0055/part-00012.parquet
ADDED
|
@@ -0,0 +1,3 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
version https://git-lfs.github.com/spec/v1
|
| 2 |
+
oid sha256:9a651b957313a0fa211d0055c02dcd1af668271ca9882b6c07431de362990ce9
|
| 3 |
+
size 150913448
|
corpus/wc_0055/part-00013.parquet
ADDED
|
@@ -0,0 +1,3 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
version https://git-lfs.github.com/spec/v1
|
| 2 |
+
oid sha256:ced1589e234b3c642b8b6a22799e626004daaa03d6b377a1bbe94815b8d16c2f
|
| 3 |
+
size 151085751
|
corpus/wc_0055/part-00014.parquet
ADDED
|
@@ -0,0 +1,3 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
version https://git-lfs.github.com/spec/v1
|
| 2 |
+
oid sha256:7c665526129a6a03d7085ea778b4a29639afcacb3d4a35dda47dc051be198317
|
| 3 |
+
size 150973464
|
corpus/wc_0055/part-00015.parquet
ADDED
|
@@ -0,0 +1,3 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
version https://git-lfs.github.com/spec/v1
|
| 2 |
+
oid sha256:f12f97e7d1fadd5940626b3b194fc14c578cd48d5556105c65b3e73188fe9c80
|
| 3 |
+
size 151085550
|
corpus/wc_0055/part-00016.parquet
ADDED
|
@@ -0,0 +1,3 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
version https://git-lfs.github.com/spec/v1
|
| 2 |
+
oid sha256:efc9a964c21c444f5eda24735322b65a56e062b55959a4410b4be20cb2f7a33b
|
| 3 |
+
size 150786811
|
corpus/wc_0055/part-00017.parquet
ADDED
|
@@ -0,0 +1,3 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
version https://git-lfs.github.com/spec/v1
|
| 2 |
+
oid sha256:c2fab237ef231a239966d8befb1dbbbd2f45136c2221d84793e0f34d4afd1502
|
| 3 |
+
size 150958284
|
corpus/wc_0055/part-00018.parquet
ADDED
|
@@ -0,0 +1,3 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
version https://git-lfs.github.com/spec/v1
|
| 2 |
+
oid sha256:f05355a180fa79c96a5cecd9923987db36ac306ac3d53ddb2a6ee4a58604d37f
|
| 3 |
+
size 151019604
|
corpus/wc_0055/part-00019.parquet
ADDED
|
@@ -0,0 +1,3 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
version https://git-lfs.github.com/spec/v1
|
| 2 |
+
oid sha256:79c116238e3808d5794464bf8ed1d6fa04a923d8858f9392e05e436b833a21d7
|
| 3 |
+
size 150959701
|
corpus/wc_0055/part-00020.parquet
ADDED
|
@@ -0,0 +1,3 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
version https://git-lfs.github.com/spec/v1
|
| 2 |
+
oid sha256:a8dfcfb666b220a360110882b7aeb8bbd6069a69c57c93f1f140cf51d5f78f3c
|
| 3 |
+
size 151063019
|
corpus/wc_0055/part-00021.parquet
ADDED
|
@@ -0,0 +1,3 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
version https://git-lfs.github.com/spec/v1
|
| 2 |
+
oid sha256:adc350f9cabef688b74d53009c3adb4d0d1aa6dcad8c1ca33612c729d23fcd91
|
| 3 |
+
size 150979854
|
corpus/wc_0055/part-00022.parquet
ADDED
|
@@ -0,0 +1,3 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
version https://git-lfs.github.com/spec/v1
|
| 2 |
+
oid sha256:1f5e36deb92eb463ef4ce11e08530a264ce5b05e512f3eac0e5f84ac2df8c137
|
| 3 |
+
size 150927777
|
corpus/wc_0055/part-00023.parquet
ADDED
|
@@ -0,0 +1,3 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
version https://git-lfs.github.com/spec/v1
|
| 2 |
+
oid sha256:654273fbf553a0f63808d5c4f8e0752c309f5fce6a0bec7be393b92c1c2810f5
|
| 3 |
+
size 150872981
|
corpus/wc_0055/part-00024.parquet
ADDED
|
@@ -0,0 +1,3 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
version https://git-lfs.github.com/spec/v1
|
| 2 |
+
oid sha256:ea113b20fe208356862f719d53a5d872ddf378c64b3e299393c2abe89a696ba5
|
| 3 |
+
size 150929099
|
corpus/wc_0055/part-00025.parquet
ADDED
|
@@ -0,0 +1,3 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
version https://git-lfs.github.com/spec/v1
|
| 2 |
+
oid sha256:b0fb6989db1f5c84059d47ca0f31fedd1fd98cc56cc0bb76aa3090c0c191d261
|
| 3 |
+
size 150924678
|
corpus/wc_0055/part-00026.parquet
ADDED
|
@@ -0,0 +1,3 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
version https://git-lfs.github.com/spec/v1
|
| 2 |
+
oid sha256:261ad46ed3b3da533fed2551b4195a29299ee0b72ef2f1519b7108de1fb0374d
|
| 3 |
+
size 150967689
|
corpus/wc_0055/part-00027.parquet
ADDED
|
@@ -0,0 +1,3 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
version https://git-lfs.github.com/spec/v1
|
| 2 |
+
oid sha256:f2e53ec09a4755eb556b2aff03d5a8c988f89a380c22c060c663795098cc1e94
|
| 3 |
+
size 151058244
|
corpus/wc_0055/part-00028.parquet
ADDED
|
@@ -0,0 +1,3 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
version https://git-lfs.github.com/spec/v1
|
| 2 |
+
oid sha256:a27c63c67065882df2e313065813a2159c92a8eeb0fff641e47d93b486323da7
|
| 3 |
+
size 150965398
|
corpus/wc_0055/part-00029.parquet
ADDED
|
@@ -0,0 +1,3 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
version https://git-lfs.github.com/spec/v1
|
| 2 |
+
oid sha256:765a4dfb891cff60aa0f82c3e868214d17dda02beb4c4acc5b7a74d99579db95
|
| 3 |
+
size 150899147
|
corpus/wc_0055/part-00030.parquet
ADDED
|
@@ -0,0 +1,3 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
version https://git-lfs.github.com/spec/v1
|
| 2 |
+
oid sha256:eedd65288d87dc1c3f14c72eed84541e0096394f1c7b4d5f19590edb654aff70
|
| 3 |
+
size 150988230
|
corpus/wc_0055/part-00031.parquet
ADDED
|
@@ -0,0 +1,3 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
version https://git-lfs.github.com/spec/v1
|
| 2 |
+
oid sha256:04d057b88d86fe28086b63c06b95137417565e05b66fd5522e99944367f24d93
|
| 3 |
+
size 150827802
|
corpus/wc_0055/part-00032.parquet
ADDED
|
@@ -0,0 +1,3 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
version https://git-lfs.github.com/spec/v1
|
| 2 |
+
oid sha256:f38a66905fbd132c0f266b66bda307822c764afd10cbdbe20bb8e60faacadc63
|
| 3 |
+
size 150933303
|
corpus/wc_0055/part-00033.parquet
ADDED
|
@@ -0,0 +1,3 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
version https://git-lfs.github.com/spec/v1
|
| 2 |
+
oid sha256:89e95710f1094d691e4447ec623d4967bdcc2b8282d96863d92f346beac98634
|
| 3 |
+
size 150891729
|
corpus/wc_0055/part-00034.parquet
ADDED
|
@@ -0,0 +1,3 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
version https://git-lfs.github.com/spec/v1
|
| 2 |
+
oid sha256:2fb780c83d68ac2fb3e391e02d5ee343db8b09d767b3bebc8ced43151e66eaaf
|
| 3 |
+
size 151096860
|
corpus/wc_0055/part-00035.parquet
ADDED
|
@@ -0,0 +1,3 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
version https://git-lfs.github.com/spec/v1
|
| 2 |
+
oid sha256:8dd44ff328dfc5642dd81dff5baf831f317ee63b77debf992754fc165248d9f0
|
| 3 |
+
size 151022778
|
corpus/wc_0055/part-00036.parquet
ADDED
|
@@ -0,0 +1,3 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
version https://git-lfs.github.com/spec/v1
|
| 2 |
+
oid sha256:b58b4198ed5c93fd954ca3b41cc166b9d6148bfe799ab1dcb671d0e520fade29
|
| 3 |
+
size 150964253
|
corpus/wc_0055/part-00037.parquet
ADDED
|
@@ -0,0 +1,3 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
version https://git-lfs.github.com/spec/v1
|
| 2 |
+
oid sha256:e45b4959b0e7d59c3666c7c31335920b20f9cf444c9fafcfae6a2ddd728b6722
|
| 3 |
+
size 151007467
|
corpus/wc_0055/part-00038.parquet
ADDED
|
@@ -0,0 +1,3 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
version https://git-lfs.github.com/spec/v1
|
| 2 |
+
oid sha256:621dcce70429b9ceac29bbd1dd1c99184a471d426630274df15c11fb458de6f8
|
| 3 |
+
size 150957707
|
corpus/wc_0055/part-00039.parquet
ADDED
|
@@ -0,0 +1,3 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
version https://git-lfs.github.com/spec/v1
|
| 2 |
+
oid sha256:b0bb527c2b139e919f3fea1b6d03f4b1d959d14414ec1274cf807ffc0880a96d
|
| 3 |
+
size 150922925
|
corpus/wc_0055/part-00040.parquet
ADDED
|
@@ -0,0 +1,3 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
version https://git-lfs.github.com/spec/v1
|
| 2 |
+
oid sha256:533e6da40e915809c7d541993723b4caa0f41eaf05d427e7306c2ab8e370b4f4
|
| 3 |
+
size 150951273
|
corpus/wc_0055/part-00041.parquet
ADDED
|
@@ -0,0 +1,3 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
version https://git-lfs.github.com/spec/v1
|
| 2 |
+
oid sha256:b3569fe9d5f4f4ec6c25a64a57ec64c96a78749068004a7cdd04cc71c213dcdb
|
| 3 |
+
size 151026939
|
corpus/wc_0055/part-00042.parquet
ADDED
|
@@ -0,0 +1,3 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
version https://git-lfs.github.com/spec/v1
|
| 2 |
+
oid sha256:0bdeadd5c2e284f37248228a15edda4461061afd65f8753a77a4d1b303aefe4f
|
| 3 |
+
size 150832792
|
corpus/wc_0055/part-00043.parquet
ADDED
|
@@ -0,0 +1,3 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
version https://git-lfs.github.com/spec/v1
|
| 2 |
+
oid sha256:c3b181e657c41e0a540614735d9181002161550d4856f4167ed70f68800df945
|
| 3 |
+
size 151056326
|
corpus/wc_0055/part-00044.parquet
ADDED
|
@@ -0,0 +1,3 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
version https://git-lfs.github.com/spec/v1
|
| 2 |
+
oid sha256:ef6f265bf11af1f3142863bf90e32bebaafdb109be63553e5c3e29fe219c988a
|
| 3 |
+
size 150718691
|
corpus/wc_0055/part-00045.parquet
ADDED
|
@@ -0,0 +1,3 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
version https://git-lfs.github.com/spec/v1
|
| 2 |
+
oid sha256:a9533a741194ae3c3a358a469fa85037ff0d2144482b19b4e49c7c7c4a4b0390
|
| 3 |
+
size 150912245
|
corpus/wc_0055/part-00046.parquet
ADDED
|
@@ -0,0 +1,3 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
version https://git-lfs.github.com/spec/v1
|
| 2 |
+
oid sha256:217351ed4fe620ff0d4438b9e3b0da9e6185990bcbe1c584b98c4675a8bb677d
|
| 3 |
+
size 150970618
|
corpus/wc_0055/part-00047.parquet
ADDED
|
@@ -0,0 +1,3 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
version https://git-lfs.github.com/spec/v1
|
| 2 |
+
oid sha256:6e3223a972e87e292935b8b0cda3a859c074cdff3e1bce553062db9580ee6421
|
| 3 |
+
size 150859467
|