PaperLab.space core

pip install paperlab-space

Evaluate LLMs on your own spreadsheet data and get paper-ready results.

Give it an Excel or CSV file, say which columns are inputs and answers, pick models and metrics, and get back an Excel workbook of outputs and scores, a LaTeX table for your paper, and a reproducibility config. paperlab-space is the open-source engine behind PaperLab.space; use it from the command line or from Python.

• Spreadsheets in: .xlsx (any sheet), .csv, .tsv, with encoding detection and Unicode-safe handling for non-Latin scripts. Validation catches empty cells, duplicate IDs, and garbled text before you spend API calls.
• Any model: OpenRouter (including free models), Gemini, OpenAI, Anthropic, Ollama, and everything else LiteLLM supports, run side by side. You bring your own API keys.
• Robust runs: per-model rate limits tuned for free tiers, retries with backoff, and resume after interruption. A cost and time estimate is shown before anything is sent.
• Metrics: exact match, accuracy (with label extraction from chatty answers), precision/recall/F1, token F1. Tokenisation handles Chinese, Japanese, Thai, and Burmese.
• Paper-ready output: Excel, a booktabs LaTeX table, and a YAML config recording model versions, prompt, seed, metric settings, and a hash of your data.

Status: early release (0.1). See PLAN.md for the roadmap: BLEU/chrF/ROUGE, LLM-as-judge, and significance tests are next.

Install

pip install paperlab-space

This installs both the Python library and the paperlab-space command.

Quickstart (command line)

export OPENROUTER_API_KEY=sk-or-...   # free key from https://openrouter.ai
export GEMINI_API_KEY=...             # free key from https://aistudio.google.com

paperlab-space init my_data.xlsx            # writes run.yaml, guessing the column roles
paperlab-space validate my_data.xlsx -i question -e answer   # check the data, preview JSONL
paperlab-space estimate run.yaml            # tokens, cost, and time; no requests sent
paperlab-space run run.yaml                 # writes results-run/

run.yaml looks like this (see examples/sentiment):

dataset:
  path: data.csv
mapping:
  input: text            # column(s) available to the prompt
  context: [language]    # extra columns available to the prompt
  expected: label        # the answer key (never shown to the model)
  id: id
prompt: |-
  Classify the sentiment of this {language} text as "positive" or "negative".
  Text: {text}
models:
  - openrouter/nvidia/nemotron-3-super-120b-a12b:free
  - id: gemini/gemini-2.5-flash
    temperature: 0
metrics:
  - name: accuracy
    labels: [positive, negative]
  - f1
seed: 42

The output folder contains:

Free models change often; paperlab-space run stops a model early with a clear message if its ID no longer exists. Browse current free models at openrouter.ai/models.

Quickstart (Python)

from paperlab_space import ColumnMapping, RunConfig, estimate, export, load_dataset, run

data = load_dataset("my_data.xlsx", sheet="test")
config = RunConfig(
    mapping=ColumnMapping(input="question", expected="answer"),
    prompt="Answer briefly: {question}",
    models=["openrouter/nvidia/nemotron-3-super-120b-a12b:free", "gemini/gemini-2.5-flash"],
    metrics=["exact_match", "token_f1"],
)
print(estimate(config, data))
results = run(config, data)
print(results.scores["gemini-2.5-flash"]["exact_match"].value)
export.write_all(results, "results/")

run works in scripts and Jupyter notebooks. Use await arun(...) inside your own async code.

Metrics

Run paperlab-space metrics for the list. Every metric takes normalisation options in the config:

metrics:
  - exact_match                         # case-insensitive, trims punctuation like "Paris."
  - name: exact_match
    label: exact_match_strict           # same metric again with different options
    ignore_case: false
    trim_punctuation: false
  - name: f1
    average: macro                      # macro | micro | weighted

Custom metrics subclass paperlab_space.metrics.Metric and use the @register decorator.

Citing

If paperlab-space helps your research, please cite it. See CITATION.cff, or use "Cite this repository" on GitHub.

Development

uv sync
uv run pytest                   # unit tests, no network
uv run pytest -m live           # real-provider smoke tests (needs API keys)
uv run ruff check . && uv run ruff format --check .
uv run mypy src tests

License

Apache 2.0