Skip to content

Leaderboard

Every evaluation run aggregates its results into a leaderboard row and updates a persistent leaderboard dataset. This lets you compare models and agent types by success rate, efficiency, and environmental cost.

Leaderboard Dataset

Each run updates {username}/smoltrace-leaderboard (on the HuggingFace Hub) or the smoltrace-leaderboard index (in OpenSearch). Community rankings are published at huggingface.co/datasets/huggingface/smolagents-leaderboard.

Metrics Tracked

The leaderboard aggregates the following per model/agent-type:

Column Description
Model Model ID evaluated
Agent Type tool, code, or both
Success Rate Percentage of tasks passed
Avg Steps Average agent steps per task
Avg Duration (ms) Average per-task duration
Total Duration (ms) Total run duration
Total Tokens Prompt + completion tokens
CO2 (g) Estimated CO2 emissions
Total Cost (USD) Estimated cost

Example rows:

Model Agent Type Success Rate Avg Steps Avg Duration (ms) Total Duration (ms) Total Tokens CO2 (g) Total Cost (USD)
mistral/mistral-large both 92.5% 2.5 500.0 15000 15k 0.22 0.005
meta-llama/Llama-3.1-8B tool 88.0% 2.1 450.0 12000 12k 0.18 0.004

How Rows Are Computed

The leaderboard row is produced by compute_leaderboard_row(...) and appended by update_leaderboard(...). When running with the default Hub output, this happens automatically. For manual control, see the Manual Dataset Management example and the Python API reference.

In OpenSearch, the leaderboard index uses run_id as the document ID, so re-running the same run_id upserts the row rather than duplicating it. See Output Formats.

Contributing Runs

Run an evaluation with the comprehensive benchmark and push to the Hub to contribute your results:

smoltrace-eval \
  --model openai/gpt-4.1-nano \
  --provider litellm \
  --dataset-name kshitijthakkar/smoltrace-benchmark-v1 \
  --agent-type both \
  --enable-otel