Accessing Benchmark Leaderboard Data

Benchmark datasets on the Hub contain leaderboards ranking models by their evaluation scores. You can access this data programmatically to analyse, build dashboards or tools on top of it.

Discovering official benchmarks

Use huggingface_hub to find all official benchmark datasets:

from huggingface_hub import HfApi

api = HfApi()
for ds in api.list_datasets(benchmark=True):
    print(ds.id)

Or via the REST API directly (useful for agents and scripting):

GET https://huggingface.co/api/datasets?filter=benchmark:official

Getting leaderboard rankings

The leaderboard API returns ranked model scores for a benchmark dataset:

GET https://huggingface.co/api/datasets/{dataset_id}/leaderboard

Use get_dataset_leaderboard to fetch ranked model scores as typed DatasetLeaderboardEntry objects:

from huggingface_hub import HfApi

api = HfApi()
leaderboard = api.get_dataset_leaderboard("SWE-bench/SWE-bench_Verified")

for entry in leaderboard[:5]:
    print(f"#{entry.rank} {entry.model_id}: {entry.value}")

huggingface_hub uses your cached token by default. For gated benchmark datasets, make sure you are logged in (huggingface-cli login) or pass a token explicitly:

Copied

leaderboard = api.get_dataset_leaderboard("gated/benchmark", token="hf_...")

Curl one-liner for quick access (useful for agents and scripting):

Copied

curl https://huggingface.co/api/datasets/cais/hle/leaderboard \
  --header "Authorization: Bearer $(cat ~/.cache/huggingface/token)" | jq .

Response fields

Each DatasetLeaderboardEntry contains:

Pre-aggregated multi-benchmark dataset

If you want scores from multiple benchmarks in a single file, the OpenEvals/leaderboard-data dataset aggregates scores across official benchmarks into one Parquet file:

You can load it directly with pandas using the hf:// path:

import pandas as pd

df = pd.read_parquet(
    "hf://datasets/OpenEvals/leaderboard-data/data/train-00000-of-00001.parquet"
)
print(df[["model_name", "provider", "aime2026_score", "mmluPro_score"]].head())

This is the fastest way to get a cross-benchmark view without calling multiple API endpoints.

Enriching with model metadata

Use huggingface_hub to enrich leaderboard data with release dates, parameter counts, and other metadata:

from huggingface_hub import HfApi

api = HfApi()
info = api.model_info("Qwen/Qwen3.5-397B-A17B")

print(f"Released: {info.created_at}")
print(f"Parameters: {info.safetensors.total / 1e9:.1f}B" if info.safetensors else "")

Model-centric view: eval results per model

The leaderboard API gives a dataset-centric view (all models on one benchmark). For the reverse — all benchmark scores for a single model — use model_info with expand=["evalResults"]:

from huggingface_hub import HfApi

api = HfApi()
info = api.model_info("Qwen/Qwen3.5-397B-A17B", expand=["evalResults"])

for result in info.eval_results:
    print(f"{result.dataset_id}: {result.value}")

This returns EvalResultEntry objects parsed from the model’s .eval_results/ files.

Example: building on leaderboard data

The Benchmark Leaderboard Race Space combines these data sources to create an animated visualization of how model rankings evolve over time. You can build your own analyses and visualizations on top of this data — see the source code for a complete example.

Related

• Eval Results — how to submit evaluation results and register benchmarks
• Official Benchmark Datasets — browse all official benchmarks