Skip to content

troubleshooting.md

Latest commit

 

History

History
49 lines (31 loc) · 1.6 KB

troubleshooting.md

File metadata and controls

49 lines (31 loc) · 1.6 KB
Raw

Troubleshooting

"Run directory already exists"

The orchestrator refuses to overwrite an existing run. Either:

  • Use a different --run_id, or
  • Delete the existing run: rm -rf runs/run_1

Target agent fails during execution

Check the logs in the generation directory:

cat runs/run_1/gen_1/agent_execution.json

Common causes:

  • Dataset paths are wrong — make sure absolute paths are used in your task
  • Required Python packages are missing from the per-run venv
  • The api_key_env for your profile's provider is not set (e.g. ANTHROPIC_API_KEY for the default profiles, NEBIUS_API_KEY for kimi-nebius) — the orchestrator warns at startup

ImportError: No module named 'anthropic'

SIA creates a fresh venv per run. If packages are missing:

  1. Check the venv creation step in the orchestrator logs
  2. Install manually into that venv: 
    runs/run_1/venv/bin/pip install anthropic

"No GEMINI_API_KEY environment variable set"

This only affects prepare_mlebench_dataset.py when it tries to generate similar task descriptions. Either:

  • Set the variable: export GEMINI_API_KEY="..."
  • Or skip that step: pass --skip-gemini

PermissionError: Kaggle authentication failed!

mle-bench downloads competitions via the Kaggle API and needs credentials. Two ways to provide them:

  • Env vars: export KAGGLE_USERNAME="..." KAGGLE_KEY="..."
  • File: drop the API token from Kaggle (Account → Create New Token) at ~/.kaggle/kaggle.json

You also need to accept the competition's rules on Kaggle's website before mlebench prepare -c <competition> can fetch it.