NIKA Trajectories
Agent diagnosis trajectories and session artifacts for official NIKA benchmark runs.
Numerical leaderboard scores are submitted as GitHub PRs to
sands-lab/nika-leaderboard.
This dataset stores the paired trajectories (messages, system logs, answers,
ground truth, and per-trial metrics).
Full operator instructions: leaderboard-submission.md.
Layout
trajectories/<release_version>/{YYYYMMDD}_{slug}/
README.md
metadata.yaml
identity.yaml
trials/{trial_id}/
run.json
messages.jsonl
nika.jsonl
ground_truth.json
eval_metrics.json
submission.json # required when outcome=success
The {YYYYMMDD}_{slug} dirname matches the GitHub scores package under
submissions/<release_version>/{YYYYMMDD}_{slug}/.
Per-trial files
| File | Role |
|---|---|
run.json |
Session meta, task text, outcome |
messages.jsonl |
Agent diagnosis / submission transcript |
nika.jsonl |
NIKA system / workflow events |
ground_truth.json |
Expected RCA labels (published with the submission) |
eval_metrics.json |
Rule-based scores + token/step counters |
submission.json |
Agent final RCA answer |
Do not include pcaps, validation dumps, llm_judge.json, or sandbox extras.
How to submit
# After an official release run and `nika leaderboard pack`:
nika leaderboard submit results/my-run/YYYYMMDD_slug
Requires HF_TOKEN or HUGGINGFACE_ACCESS_TOKEN with write access to this
dataset (or the ability to open a Hub PR). The CLI opens a dataset Discussion/PR;
maintainers merge accepted entries.
Optional Hub CLI for admins:
hf discussions list Zhihao98/nika-trajectories --type dataset
hf discussions merge Zhihao98/nika-trajectories <N> --yes --type dataset
Review checklist
- Paired scores package opened / linked on GitHub leaderboard
- Trial set matches the release (
case_count × n_trials) - No secrets or absolute machine paths in package text
- No pcaps or forbidden dump artifacts
- Downloads last month
- 393