messages listlengths 2 2 | system_prompt stringlengths 1.49k 4.45k | metadata dict | type stringclasses 1
value |
|---|---|---|---|
[
{
"content": "What is the name of the dataset used in the experiments described in the text, and what is the full title of the paper that provided the FinanceBench dataset?",
"role": "user",
"token_ids": [
3838,
374,
279,
829,
315,
279,
10337,
1483,
... |
You are in a conversation about the following user information.
<info>
uravsky2025tokasaurus}. Because a \artifact is a KV cache, it can be loaded directly into the KV cache slots using existing mechanisms for handling cached prefixes. LLM inference servers are heavily optimized for managing distinct KV-caches for mu... | {
"initial_system_prompt": "\nYou are in a conversation about the following user information.\n\n<info>\nuravsky2025tokasaurus}. Because a \\artifact is a KV cache, it can be loaded directly into the KV cache slots using existing mechanisms for handling cached prefixes. LLM inference servers are heavily optimized for... | todo |
[
{
"content": "What is the name of the method proposed in the subsection that involves training a smaller KV cache offline by backpropagating loss into the key and value vectors, and what is the term used to describe the trained KV cache representing the corpus?",
"role": "user",
"token_ids": [
383... |
You are in a conversation about the following user information.
<info>
organization may use LLMs to understand codebases~\cite{nam2024using}, financial documents~\cite{islam2023financebench}, legal texts~\cite{guha2023legalbench, zheng2025reasoning}, textbooks~\cite{ouellette2025can}, or personal files~\cite{arora20... | {
"initial_system_prompt": "\nYou are in a conversation about the following user information.\n\n<info>\n organization may use LLMs to understand codebases~\\cite{nam2024using}, financial documents~\\cite{islam2023financebench}, legal texts~\\cite{guha2023legalbench, zheng2025reasoning}, textbooks~\\cite{ouellette202... | todo |
[
{
"content": "What is the accuracy improvement achieved by the context-distillation objective on the longhealth dataset compared to other methods as reported in the app:results section of the provided information?",
"role": "user",
"token_ids": [
3838,
374,
279,
13403,
1567... |
You are in a conversation about the following user information.
<info>
{x}$ and the corresponding subcorpus used to generate them $\subctx$.
\vspace{-2mm}
\begin{equation}
\underset{\ctxrep}{\arg\min} \quad
\sum_{(\mathbf{x}, \subctx) \in \mathcal{D}_\text{train}}
\sum_{i=1}^{|\mathbf{x}|}
... | {
"initial_system_prompt": "\nYou are in a conversation about the following user information.\n\n<info>\n{x}$ and the corresponding subcorpus used to generate them $\\subctx$.\n\\vspace{-2mm}\n\\begin{equation}\n \\underset{\\ctxrep}{\\arg\\min} \\quad\n \\sum_{(\\mathbf{x}, \\subctx) \\in \\mathcal{D}_\\te... | todo |
[
{
"content": "What is the title of the paper submitted to ICML 2025 by David S.~Hippocampus, and what is the author's email address listed in the provided LaTeX document?",
"role": "user",
"token_ids": [
3838,
374,
279,
2265,
315,
279,
5567,
14634,
3... |
You are in a conversation about the following user information.
<info>
uripsconf{neurips}
\def\icmlconf{icml}
\def\arxivconf{arxiv}
\usepackage[table,xcdraw]{xcolor}
\ifx\conference\neuripsconf
\usepackage[nonatbib]{styles/neurips_2025}
\usepackage[numbers,sort&compress]{natbib}
\else\ifx\conference\arxivconf... | {
"initial_system_prompt": "\nYou are in a conversation about the following user information.\n\n<info>\nuripsconf{neurips}\n\\def\\icmlconf{icml}\n\\def\\arxivconf{arxiv}\n\\usepackage[table,xcdraw]{xcolor}\n\\ifx\\conference\\neuripsconf\n \\usepackage[nonatbib]{styles/neurips_2025}\n \\usepackage[numbers,sor... | todo |
[
{
"content": "What is the name of the command used to define the style for code listings in the LaTeX document class used for the arXiv conference submission, and what is the specific style name mentioned in the provided code snippet?",
"role": "user",
"token_ids": [
3838,
374,
279,
... |
You are in a conversation about the following user information.
<info>
rightmargin=10pt
}
\lstset{style=codestyle}
\newcommand{\repetitiveMQAR}{\text{repetitive MQAR}}
\newcommand{\numKeys}{s}
\newcommand{\contextSize}{N}
\newcommand{\numPairs}{m}
\newcommand{\numQueries}{n}
\newcommand{\timestep}{t}
\newcommand{\emp... | {
"initial_system_prompt": "\nYou are in a conversation about the following user information.\n\n<info>\nrightmargin=10pt\n}\n\\lstset{style=codestyle}\n\\newcommand{\\repetitiveMQAR}{\\text{repetitive MQAR}}\n\\newcommand{\\numKeys}{s}\n\\newcommand{\\contextSize}{N}\n\\newcommand{\\numPairs}{m}\n\\newcommand{\\numQ... | todo |
[
{
"content": "What is the name of the technique used to parameterize the $\\ctxrep$ in the provided information, and what is the specific hyperparameter that controls the size of the full $Z$ in this technique?",
"role": "user",
"token_ids": [
3838,
374,
279,
829,
315,
... |
You are in a conversation about the following user information.
<info>
is essential that $\llm_Z$ can generalize across different $\query \in \queries$. \textbf{This is non-trivial because $\queries$ is unknown when $Z$ is being learned offline.} If $\llm_Z$ does not generalize, then practitioners may need to learn ... | {
"initial_system_prompt": "\nYou are in a conversation about the following user information.\n\n<info>\n is essential that $\\llm_Z$ can generalize across different $\\query \\in \\queries$. \\textbf{This is non-trivial because $\\queries$ is unknown when $Z$ is being learned offline.} If $\\llm_Z$ does not generali... | todo |
[
{
"content": "What is the name of the dataset used in the study to evaluate the method's ability to process corpora exceeding the context window length of \\llamaeightb, and what are the token counts of the two versions of the dataset?",
"role": "user",
"token_ids": [
3838,
374,
279,
... |
You are in a conversation about the following user information.
<info>
we increase the amount of compute used in self-study: the longer an \artifact is trained, the greater task performance. \Cref{fig:scaling-indexing} plots the performance for differentially sized \artifacts as a function of the number of training ... | {
"initial_system_prompt": "\nYou are in a conversation about the following user information.\n\n<info>\n we increase the amount of compute used in self-study: the longer an \\artifact is trained, the greater task performance. \\Cref{fig:scaling-indexing} plots the performance for differentially sized \\artifacts as ... | todo |
[
{
"content": "What is the hyperparameter that controls the size of the full Z in the context representation method described in the passage, and how does the memory footprint of Z relate to the KV cache of a prompt?",
"role": "user",
"token_ids": [
3838,
374,
279,
17071,
16... |
You are in a conversation about the following user information.
<info>
$\llm_Z$ can generalize across different $\query \in \queries$. \textbf{This is non-trivial because $\queries$ is unknown when $Z$ is being learned offline.} If $\llm_Z$ does not generalize, then practitioners may need to learn different $Z$ for ... | {
"initial_system_prompt": "\nYou are in a conversation about the following user information.\n\n<info>\n $\\llm_Z$ can generalize across different $\\query \\in \\queries$. \\textbf{This is non-trivial because $\\queries$ is unknown when $Z$ is being learned offline.} If $\\llm_Z$ does not generalize, then practitio... | todo |
[
{
"content": "What is the name of the research group associated with the GitHub repository HazyResearch/cartridges, and which Stanford University researchers are listed as authors in the abstract of the document?",
"role": "user",
"token_ids": [
3838,
374,
279,
829,
315,
... |
You are in a conversation about the following user information.
<info>
}Stanford University \quad \textsuperscript{2} Caltech \quad \textsuperscript{3}University at Buffalo \quad \quad * Equal contribution \\[3pt]
\faEnvelope \enspace \texttt{eyuboglu@stanford.edu, rehrlich@stanford.edu, simarora@stanford.edu} \\
\fa... | {
"initial_system_prompt": "\nYou are in a conversation about the following user information.\n\n<info>\n}Stanford University \\quad \\textsuperscript{2} Caltech \\quad \\textsuperscript{3}University at Buffalo \\quad \\quad * Equal contribution \\\\[3pt]\n\\faEnvelope \\enspace \\texttt{eyuboglu@stanford.edu, rehrli... | todo |
[
{
"content": "Summarize the section titled \"Model Architecture\" from the document \"Deep Learning for Natural Language Processing\" in one concise paragraph, highlighting the key components and their roles in the model.",
"role": "user",
"token_ids": [
9190,
5612,
551,
279,
... |
You are in a conversation about the following user information.
<info>
{\resps}{R}
\newcommand{\resp}{r}
\newcommand{\vocab}{\mathcal{V}}
\newcommand{\llm}{\mathcal{F}}
\newcommand{\numtrain}{m_\text{train}}
\newcommand{\seqlength}{n}
\newcommand{\seqlengthsub}{n_\text{sub}}
\newtheorem{theorem}{Theorem}
\newtheorem{... | {
"initial_system_prompt": "\nYou are in a conversation about the following user information.\n\n<info>\n{\\resps}{R}\n\\newcommand{\\resp}{r}\n\\newcommand{\\vocab}{\\mathcal{V}}\n\\newcommand{\\llm}{\\mathcal{F}}\n\\newcommand{\\numtrain}{m_\\text{train}}\n\\newcommand{\\seqlength}{n}\n\\newcommand{\\seqlengthsub}{... | todo |
[
{
"content": "What is the name of the company that produced the 10-K documents used in the evaluation of the \\artifacts when trained across sizes $\\{512, 1024, 2048, 4096\\}$ and long 10-K documents from AMD, Pepsi, AMEX, and Boeing?",
"role": "user",
"token_ids": [
3838,
374,
279,
... |
You are in a conversation about the following user information.
<info>
4$ on \longhealth; see \Cref{fig:seeds}).
Interestingly, on \qasper we do not see any significant benefit from using the diverse seed prompts. This is perhaps because, compared to \longhealth and \mtob, the queries in \qasper are less reasoning in... | {
"initial_system_prompt": "\nYou are in a conversation about the following user information.\n\n<info>\n4$ on \\longhealth; see \\Cref{fig:seeds}).\nInterestingly, on \\qasper we do not see any significant benefit from using the diverse seed prompts. This is perhaps because, compared to \\longhealth and \\mtob, the ... | todo |
[
{
"content": "What is the percentage increase in accuracy when initializing the KV cache with the first $p$ tokens of the corpus compared to using key and value vectors of random tokens on the \\longhealth dataset?",
"role": "user",
"token_ids": [
3838,
374,
279,
11414,
526... |
You are in a conversation about the following user information.
<info>
different aspects of \method and \artifact parameterization.
We provide full results in Appendix \ref{app:results} and highlight key findings here and in \Cref{fig:ablations}.
\paragraph{\artifact Parameterization} In \Cref{sec:artifact-parameter... | {
"initial_system_prompt": "\nYou are in a conversation about the following user information.\n\n<info>\n different aspects of \\method and \\artifact parameterization.\nWe provide full results in Appendix \\ref{app:results} and highlight key findings here and in \\Cref{fig:ablations}.\n\\paragraph{\\artifact Paramet... | todo |
[
{
"content": "Summarize the section titled \"Offline Training of KV Caches\" from the document \"Efficient Long-Context Inference with Artifacts\". Please provide a concise summary of the key ideas and methodology described in this section.",
"role": "user",
"token_ids": [
9190,
5612,
... |
You are in a conversation about the following user information.
<info>
024llama3}. This severely limits user throughput: on a single H100 GPU, LLaMA 8B’s peak throughput (tokens/s) drops by $77\times$ when increasing the context from 1k to 120k tokens (\Cref{fig:micros}).
Prior work has thus explored ways to reduce K... | {
"initial_system_prompt": "\nYou are in a conversation about the following user information.\n\n<info>\n024llama3}. This severely limits user throughput: on a single H100 GPU, LLaMA 8B’s peak throughput (tokens/s) drops by $77\\times$ when increasing the context from 1k to 120k tokens (\\Cref{fig:micros}).\nPrior wo... | todo |
[
{
"content": "Summarize the section titled \"Initialization\" from the document \"ICL KV Cache and Artifact KV Cache\" provided in the information above. Make sure the summary is concise and covers the key points discussed about the initialization of the trainable KV cache.",
"role": "user",
"token_ids"... |
You are in a conversation about the following user information.
<info>
{ICL KV Cache} \vspace{-1.5em} \\
\begin{align*}
\underbrace{(\mathbf{k}[1], \mathbf{v}[1]), \dots, (\mathbf{k}[{n_\ctx}], \mathbf{v}[{n_\ctx}])}_{\text{KV pairs for~}\ctx},
\underbrace{(\mathbf{k}[{n_\ctx + 1}], \mathbf{v}[{n_\ctx + 1}])\dots}_{\... | {
"initial_system_prompt": "\nYou are in a conversation about the following user information.\n\n<info>\n{ICL KV Cache} \\vspace{-1.5em} \\\\\n\\begin{align*}\n\\underbrace{(\\mathbf{k}[1], \\mathbf{v}[1]), \\dots, (\\mathbf{k}[{n_\\ctx}], \\mathbf{v}[{n_\\ctx}])}_{\\text{KV pairs for~}\\ctx},\n\\underbrace{(\\mathbf... | todo |
[
{
"content": "What is the main purpose of using self-supervised synthetic data in the training of artifacts as described in the provided information?",
"role": "user",
"token_ids": [
3838,
374,
279,
1887,
7428,
315,
1667,
656,
58515,
77990,
... |
You are in a conversation about the following user information.
<info>
can generate responses to many types of queries that match the quality of queries generated with ICL.
See \Cref{fig:banner} for a visualization of the \artifact approach.
\ifx\conference\icmlconf
\input{figures/tradeoff-within/tradeoff-within}
\f... | {
"initial_system_prompt": "\nYou are in a conversation about the following user information.\n\n<info>\n can generate responses to many types of queries that match the quality of queries generated with ICL.\nSee \\Cref{fig:banner} for a visualization of the \\artifact approach.\n\\ifx\\conference\\icmlconf\n\\input{... | todo |
[
{
"content": "Can you structure the information in {{subsection}} of {{document}} related to {{something specific}} in the following format: YAML? Be sure to include precise information like any dates, times, names, and numerical values.",
"role": "user",
"token_ids": [
6713,
498,
5944... |
You are in a conversation about the following user information.
<info>
{\eg}{\textit{e.g.}\xspace}
\newcommand{\ie}{\textit{i.e.}\xspace}
\newcommand{\etc}{\textit{etc.}\xspace}
\newcommand{\etal}{\textit{et al.}\xspace}
\newcommand{\ctx}{\mathcal{C}}
\newcommand{\subctx}{\tilde{\mathbf{c}}}
\newcommand{\seed}{\mathb... | {
"initial_system_prompt": "\nYou are in a conversation about the following user information.\n\n<info>\n{\\eg}{\\textit{e.g.}\\xspace}\n\\newcommand{\\ie}{\\textit{i.e.}\\xspace}\n\\newcommand{\\etc}{\\textit{etc.}\\xspace}\n\\newcommand{\\etal}{\\textit{et al.}\\xspace}\n\\newcommand{\\ctx}{\\mathcal{C}}\n\\newcomm... | todo |
End of preview. Expand in Data Studio
No dataset card yet
- Downloads last month
- 8