Datasets:

BrachioLab
/

dist-defense-traces-multicampaign-main100-augbenign-anchors-eval-v3-artemis-turnlike

You need to agree to share your contact information to access this dataset

This repository is publicly accessible, but you have to accept the conditions to access its files and content.

Log in or Sign Up to review the conditions and access this dataset content.

BrachioLab/dist-defense-traces-multicampaign-main100-augbenign-anchors-eval-v3-artemis-turnlike

Combined dataset built from the tool-action split baseline plus ARTEMIS decomposition traces.

Base Dataset

Source repo: BrachioLab/dist-defense-traces-multicampaign-main100-augbenign-anchors-eval-v2-toolactions-v1
Source SHA: f743c4d7175906e5ae038add2ef7996f29ed1514

Split Sizes

dist_train: 85585 -> 86923
dist_test: 264296 -> 272612

Turn-Like Normalization

dist_train conversations with multi-tool output rows split: 225
dist_train rows added from splitting: 1338
dist_test conversations with multi-tool output rows split: 591
dist_test rows added from splitting: 2634

ARTEMIS Append (dist_test only)

Trace sources: 10
Runs scanned: 120
Subagent conversations extracted: 284
Rows added: 5682
- Input rows: 79
- Output rows: 5603
- Exec tool-action rows: 3944
- Assistant narrative rows: 1659
Duplicates skipped: 0

ARTEMIS Outcomes (From Trace Runs)

Traces with matched eval rows: 10/10
Total tasks: 120
Successes: 100
Failures: 20
Errors: 0
Overall success rate: 0.8333
minimax/minimax-m2.5: 49/60 successes (0.8167), failures=11, errors=0, traces=5/5
openai/gpt-oss-120b:exacto: 51/60 successes (0.8500), failures=9, errors=0, traces=5/5

Harmful Row Counts

dist_train (is_misuse=True): 0
dist_test (is_misuse=True): 40065

Build Metadata

Metadata JSON: artemis_turnlike_build_meta.json