Guides

Output files

What each file in the results folder contains, and how to customise the LaTeX table.

paperlab-space run run.yaml writes to results-run/ next to the config. Change the folder with --out. In Python, export.write_all(results, "results/") writes the same four files.

results.xlsx

Sheet	Contents
Summary	One row per model: label, model ID, each metric's score, rows, failed requests, cost (USD), and input and output tokens.
Outputs	One row per data row: the ID, the input and context columns, the expected answer, and for each model its output, its row-level scores, and its error message if any request failed.
Config	The reproducibility config, as in `run.yaml`.

Several valid answers are joined with | in the expected column. Outputs that start with = are stored as text, never as formulas, and outputs longer than Excel's cell limit (32,767 characters) are truncated.

table.tex

A booktabs table with one row per model and one column per metric:

LaTeX

% Generated by paperlab-space 0.1.5
% Requires \usepackage{booktabs}
\begin{table}[t]
\centering
\begin{tabular}{lrr}
\toprule
Model & Accuracy & F1 \\
\midrule
nvidia/nemotron-3-super-120b-a12b:free & 85.0 & 84.2 \\
Gemini 3.5 Flash & \textbf{92.5} & \textbf{92.1} \\
\bottomrule
\end{tabular}
\caption{Results on 40 examples (\%).}
\label{tab:results}
\end{table}

Add \usepackage{booktabs} to your preamble and \input{table.tex} where the table goes. Scores are shown as percentages with one decimal. With more than one model, the best score in each column is bold, every tied one included. Model names and metric titles are escaped for LaTeX.

To change the table, call latex_table from Python:

Python

from paperlab_space import export

tex = export.latex_table(
    results,
    metrics=["accuracy"],  # which columns, in order
    caption="Sentiment accuracy on the test set.",
    label="tab:sentiment",
    decimals=2,
    percent=False,  # raw scores in [0, 1]
    bold_best=True,
    floating=False,  # just the tabular, without table/caption/label
)
export.write_latex(results, "table.tex", decimals=2)  # the same options

run.yaml

The reproducibility config: your full run config plus the package versions, model versions, and a hash of the data. Run it again with paperlab-space run run.yaml. See Reproducibility and citing.

results.json

Everything about the run: the config, the environment, the dataset rows, every output with its tokens, cost, latency, and error, and every score. Use it to resume a run, or load it in Python:

Python

from paperlab_space import RunResults

results = RunResults.load_json("results-run/results.json")

results.json contains your full dataset. Keep that in mind before you share it. Local folder paths are not included; the dataset is recorded by file name only.

Edit this page on GitHub