Results from model evaluations on GSO (Global Software Optimization).
gso-experiments/
├── models.json # Model configurations
├── results/
│ ├── manifest.json # Evaluation summaries and Docent URLs to agent trajectories
│ └── reports/ # Full evaluation reports from GSO
│ └── <model>.json
└── scripts/ # Internal tooling
Agent trajectories are viewable via Docent, a tool by Transluce for exploring and analyzing agent behavior.
- Trajectories: See Docent URLs in
results/manifest.jsonto browse agent runs - Reports: See
results/reports/<model>.jsonfor evaluation reports - Raw data:
gsutil -m cp -r gs://gso-experiments/<model> ./
We recommend using Docent for most analysis cases. However, if you need to download raw data, you can use the option below.
Raw trajectories and evaluation logs are stored in a public Google Cloud Storage bucket. To download them, you'll need to:
- Create a Google Cloud account
- Install the Google Cloud CLI
- Run
gcloud initto configure your credentials
Then download data with:
# List available models
gsutil ls gs://gso-experiments/
# Download a specific model's data
gsutil -m cp -r gs://gso-experiments/<model> ./Each model's data is stored in the following structure:
gs://gso-experiments/<model>/
├── trajs/ # Agent trajectories
│ ├── output.jsonl
│ └── metadata.json
├── logs/ # GSO evaluation logs
└── report.json # Aggregated results
[!TODO] Add instructions for submitting results for a new model
MIT