Smoke24 Agentic Benchmarks
Public, reproducible Terminal-Bench 2.0 Smoke24 benchmark artifacts for local RTX 3090-class agentic model serving.
Why Smoke24
I created the Smoke24 subset because I needed a relatively quick benchmark that could run locally on an RTX 3090-class machine in a couple of hours. The goal is to get a practical read on model performance and stability under a real agentic Terminal-Bench workload, without paying the turnaround cost of a much larger benchmark every time a quantization, context, or serving setup changes.
Smoke24 is therefore a fast local signal, not a replacement for the full Terminal-Bench 2.0 leaderboard.
Canonical publication surfaces:
- HF Dataset: https://huggingface.co/datasets/XReyRobert/smoke24-agentic-benchmarks
- Interactive HF Space: https://huggingface.co/spaces/XReyRobert/gptq-pro-smoke24-agentic-3090
This repository separates three concerns:
data/public: aggregate benchmark data intended for public distribution.reports/public: human-readable reports and static images generated from public aggregate data.site: generated static dashboard output, suitable for a Hugging Face static Space or static web host.
Raw Terminal-Bench run directories, model transcripts, private endpoint details, and IBM Bob/internal comparison artifacts are intentionally not versioned here.
Quick Start
python3 -m venv .venv
. .venv/bin/activate
pip install -r requirements.txt
make validate
make build
make serve
Then open http://127.0.0.1:8765.
Current Dataset
- Benchmark: Terminal-Bench 2.0 Smoke24
- Agent: Terminus-2
- Local hardware lens: RTX 3090-class serving boundary
- Public dashboard data:
data/public/derived - External public references:
data/public/external
External public rows are success-rate references only. They are not placed on local wall-time or local serving-cost axes.
- Downloads last month
- 97