What is Wrong with Perplexity for Long-context Language Modeling?
Paper • 2410.23771 • Published
Datasets:
text stringlengths 29.4k 173k | key_token_spans listlengths 1 351 |
|---|---|
In our prior work, we have found that technological innovation involves not only creating new ideas but also translating those ideas into a new product or service. Innovation, and the research driving it, is inherently risky because the likelihood that research can be translated into a product or service and the ultima... | [
[
28607,
28610
],
[
29326,
29330
],
[
43424,
43435
],
[
45478,
45489
],
[
45512,
45515
],
[
50726,
50734
],
[
50750,
50757
],
[
50782,
50808
],
[
54778,
54781
],
[
57088,
57094
],
[
57541,
... |
A variety of federal laws, regulations, and policies establish requirements and guidance for EPA to follow when appointing members to serve on advisory committees. For example, one purpose of FACA is to ensure that uniform procedures govern the establishment and operation of advisory committees. Also under FACA, an age... | [
[
36457,
36467
],
[
36536,
36542
],
[
36557,
36561
],
[
36596,
36597
],
[
36600,
36612
],
[
36666,
36670
],
[
37462,
37465
],
[
38904,
38907
],
[
41865,
41872
],
[
43135,
43137
],
[
44885,
... |
"Foreign students interested in studying in the United States must first be admitted to an SEVP-cert(...TRUNCATED) | [[29905,29908],[29938,29941],[29980,29983],[30010,30027],[30062,30089],[30106,30109],[30130,30137],[(...TRUNCATED) |
"Zika is spread to people primarily through the bite of an infected mosquito but can also be transmi(...TRUNCATED) | [[38356,38364],[38368,38372],[38652,38660],[38674,38678],[38725,38733],[39405,39407],[39583,39587],[(...TRUNCATED) |
"Prevalence of Forced Labor in Thailand A 2020 report by the United Nation’s International Labor O(...TRUNCATED) | [[34406,34416],[35074,35081],[37970,37977],[37980,37988],[37995,37998],[38032,38053],[38058,38066],[(...TRUNCATED) |
"SOCOM has a unique structure and responsibilities in that it has both combatant command responsibil(...TRUNCATED) | [[28578,28581],[39574,39575],[39584,39585],[39619,39630],[39811,39820],[39864,39868],[39882,39896],[(...TRUNCATED) |
"Two State entities play key roles in education assistance in the West Bank and Gaza—State’s Bur(...TRUNCATED) | [[26231,26234],[28206,28218],[31625,31629],[31636,31637],[32229,32235],[32263,32271],[32308,32320],[(...TRUNCATED) |
"TSA is the primary federal agency responsible for implementing and overseeing the security of the n(...TRUNCATED) | [[44490,44498],[44503,44505],[45350,45351],[46670,46675],[57728,57730],[57735,57743],[58120,58133],[(...TRUNCATED) |
"State and local governments rely on a range of revenue sources to support their activities, includi(...TRUNCATED) | [[41133,41146],[41576,41577],[42636,42637],[42656,42657],[44603,44612],[45464,45471],[46475,46484],[(...TRUNCATED) |
"The number of AI/AN veterans eligible for both VA and IHS services is unknown. The U.S. Census Bure(...TRUNCATED) | [[25869,25880],[25918,25927],[32374,32384],[39266,39267],[39268,39269],[39270,39271],[39274,39277],[(...TRUNCATED) |
End of preview. Expand in Data Studio
GovReport LongPPL Key Tokens
Precomputed key-token spans for computing LongPPL (Fang et al., ICLR 2025) on the GovReport corpus.
LongPPL measures perplexity only over key tokens — the (<10%) tokens whose prediction genuinely depends on long-range context — instead of averaging over all tokens. Identifying key tokens requires a strong evaluator model and a long-vs-short context loss contrast, which is expensive and must be done once. This dataset ships that precomputation so that evaluating a target model needs only a forward pass over the target model — no evaluator at eval time.
Contents
| Column | Type | Description |
|---|---|---|
text |
string |
A GovReport long document |
key_token_spans |
list[list[int]] |
Key-token character intervals [start, end] in text |
- 100 documents, 4,867 key-token spans total.
- Spans are character offsets, so they are tokenizer-independent — map them onto any target model's tokenization at eval time.
How it was built
- Source corpus:
ccdv/govreport-summarization(reportfield,testsplit); documents with 5,000–32,768 tokens. - Evaluator model:
Qwen/Qwen3.6-27B.Deviation from the paper: the original LongPPL uses
Qwen2-72B-Instructas the evaluator. A 27B model is used here because it fits a single GPU. Key-token selection depends on the evaluator, so absolute LongPPL values are not directly comparable to the paper's. - Hyperparameters (official LongPPL defaults):
trunc_len=4096,sliding_window=1024,alpha=2,beta=-2.
Usage
from datasets import load_dataset
ds = load_dataset("wchai/govreport-longppl-keytokens", split="train")
for row in ds:
text = row["text"]
key_spans = row["key_token_spans"] # [[start, end], ...] char intervals
# tokenize `text` with your target model (return_offsets_mapping=True),
# map the char spans onto token indices, and average the target model's
# cross-entropy over the key tokens: LongPPL = exp(mean CE over key tokens)
Citation
@inproceedings{fang2025longppl,
title={What is Wrong with Perplexity for Long-context Language Modeling?},
author={Fang, Lizhe and Wang, Yifei and Liu, Zhaoyang and Zhang, Chenheng and
Jegelka, Stefanie and Gao, Jinyang and Ding, Bolin and Wang, Yisen},
booktitle={ICLR},
year={2025}
}
Key tokens derive from GovReport (public U.S. government reports); see the source dataset for its terms.
- Downloads last month
- 29