benchflow-ai

Skillsbench Trajectories — Data skill for Claude Code

Data community

Result snapshots and supporting data for the SkillsBench leaderboard.

How to install Skillsbench Trajectories

This entry records only its repository, not the path inside it, so there is no exact command to give. Open benchflow-ai/skillsbench-trajectories and copy the folder into ~/.claude/skills/, or the file into ~/.claude/agents/.

What Skillsbench Trajectories does

Result snapshots and supporting data for the [SkillsBench](https://skillsbench.ai) leaderboard.

Alternatives in Data

Bitbucket Data Center 71.9k ★
Financial Services — Reference agents, skills, and data connectors for the financial-services workflows we see most — investment ba 34.5k ★
Data Contract — This document defines which files belong to the system (auto-updatable) and which belong to the user (never to 32.5k ★

README

SkillsBench — results & website data

Result snapshots and supporting data for the [SkillsBench](https://skillsbench.ai) leaderboard.

Current snapshot

The latest experimental results are the HuggingFace dataset [`benchflow/skillsbench-leaderboard`](https://huggingface.co/datasets/benchflow/skillsbench-leaderboard), **main branch + [PR #11](https://huggingface.co/datasets/benchflow/skillsbench-leaderboard/discussions/11)** (v0.1 + v1.1). It covers **18 model–harness configurations × 87 tasks × 2 conditions (with / without curated Skills) × 3 trials**, run at each model's maximum reasoning effort and scored pass/fail. That dataset is the source of truth for both trajectories and results.

`website-data/`

Build-time data consumed by skillsbench.ai (fetched from this repo's `main`):

results-registry.json — per (task, model–harness, condition) aggregate (score, pass/perfect counts) backing each task page's Results board. Regenerated from the HuggingFace snapshot above.

The website no longer hosts agent trajectory files; trajectories are browsed and downloaded on the HuggingFace dataset.

Other directories

The contributor-named directories (`yimin/`, `xiangyi-*/`, `shenghan/`, `wenbo/`, `minimax-m2.1+claude-code/`) are an **archive of earlier-generation runs**. Current trajectories live on the HuggingFace dataset, not in this repo.