Wild-Trajectory

A filtered, normalized subset of open-source agent trajectories for behavioral analysis. 27,803 trajectories across 19 sources, 80 models and 50 harnesses, one row per trajectory.

Built by (1) collecting raw trajectories from public releases, (2) normalizing each source's native log format into a common typed event stream, (3) resolving model/harness/task strings to canonical identifiers, (4) rendering every trajectory as a Harbor ATIF-v1.8 document, (5) keeping trajectories placeable in a Model × Harness × Task design that are linear and have ≥5 annotatable agent steps, (6) forming model × task-type × harness cells (≥10 trajectories each) and keeping only models whose cells span ≥2 task types, so the subset forms a connected core for cross-context trait analysis, and (7) drawing at most 30 trajectories per cell, outcome-balanced. funnel.json records every drop; _summary.json the per-source / per-task-type row counts.

Columns

• trajectory_id, source, benchmark, task_type, task_subtype, native_task_id
• model — canonical checkpoint id (model_raw = the upstream string); harness — canonical scaffold id (harness_raw, harness_version)
• instruction; score_norm — normalized outcome in [0,1] (null when unknown); score_type, exit_status
• reasoning_tier, parse_tier, has_user_sim, has_multi_agent, license, fidelity_level, l0_run, l0_logical_path
• counts: n_events, n_llm_calls, n_assistant_turns, tokens_input, tokens_output, cost_usd, duration_s, and per-ATIF-document n_steps, n_agent_steps, n_annotatable_steps, n_user_steps, n_system_steps, n_tool_calls, n_observations, n_reasoning_visible_steps, n_reasoning_hidden_steps, has_branching, has_image_obs
• atif — the trajectory as one Harbor ATIF-v1.8 JSON document (validated with harbor==0.23.0). One source=agent step per model call; observations sit on the step that issued the call. Everything this corpus knows beyond the ATIF vocabulary is under extra["wild"] at the root (task, outcome, flags, provenance), agent (raw model/harness strings, config), step (kind for system steps, scripted for runner-delivered user text, harness_injected, summary_only / redacted reasoning, l1_event_idxs), tool-call and observation-result (modality — screenshots are placeholders, l0_ref, error) levels. Consumers that only know ATIF can ignore extra.

Task types present: swe, science, terminal, ml-engineering, security, tool-api, research-qa, other, dialogue, web, office, gui.

Changes since wild-trajectory-v0

• New sources (all frozen at a pinned upstream revision): OSWorld 2.0 (osworld2, task type gui; screenshots not inlined — image observations carry content=null and obs.modality="image"; checkpoint evaluations are score_event rows; the runner's canned replies are user_msg rows with scripted=true), the official Terminal-Bench 2.0 leaderboard (terminal-bench-leaderboard, the 26 submissions shipping ATIF trajectories; overlaps the community scrape on some agent/model pairs — dedupe by native_task_id + trial when needed, official wins), Long-Horizon Terminal-Bench (lhtb-leaderboard; only tasks with a complete trajectory), LiveClawBench (liveclawbench, OpenClaw over mock web/email/ calendar/shell services), NatureBench-traces (naturebench-traces, Claude Code / Codex / Gemini CLI reproducing Nature-paper analyses, task type science) and the Open Agent Leaderboard v2 traces (exgentic-traces-v2; 6 benchmarks × 5 harnesses).
• Retired sources: nebius, swe-gym, claude-code-traces (synthetic / training data or unclear provenance).
• Harness ids follow <scaffold family>/<variant> (e.g. openhands/codeact, swe-agent/v1, terminus/2, claude-code/cli); scaffold versions live in the harness registry, not the id.
• Task types: data-ml split into ml-engineering and science; new gui.
• reasoning_tier: one rule for every source — full (verbatim reasoning on most turns), mixed, encrypted (only API summaries / redacted bodies — OpenAI reasoning models, Codex, Gemini CLI; the summaries are kept in reasoning events flagged summary_only), narrated, none.
• Adapter fixes in HAL (simulated-user turns, smolagents observations, langchain tool calls), terminus-2 observations, Cybench subtask scores.
• Sampling: cell floor 25 → 10 and cap 50 → 30 (the one-trial-per-task sources such as LHTB never reached 25; the largest sources dominated at 50); connectivity threshold 3 → 2 task types, measured on the floor-passing cells; the extra lower-floor draw over the reasoning-visible / interpersonal strata is retired; the linearity and ≥5-agent-steps checks moved into the eligibility gate (no write-time drops).
• Format: the trajectory event list is replaced by atif (ATIF-v1.8).

Per-source upstream licenses vary; see the source column and the original releases.