ISETrace

English · 简体中文

Execution-grounded, multi-turn agent trajectories synthesized by the ISE (Intent → Simulate → Execute) paradigm.

ISETrace is a corpus of 23,132 multi-turn OS-agent trajectories in which every tool call is executed against a live, isolated operating-system workspace — yielding authentic failure–recovery dynamics (real exit codes, error strings, written files) rather than simulated tool responses. The trajectories are grown from 43,955 structured user intents sampled over a Persona × Domain × Task × Complexity space, then driven through a role-locked user simulator that sustains long, coherent multi-turn exchanges.

• 📄 Paper: arXiv:2606.11520
• 🛠️ Pipeline (umbrella): https://github.com/Valiere01/ISE-Trace
• 🧩 Stage 1 — intent construction: https://github.com/NairongZheng/intent_creator
• ⚙️ Stage 2+3 — simulation + execution: https://github.com/NairongZheng/openclaw_gen_data

---

What's in here

This repository exposes two configs (select with name= in load_dataset):

Config	Rows	Description
trajectories	23,132	Full multi-turn agent trajectories with real tool execution. The main SFT corpus.
intents	43,955	The unique structured 4D intents that seed the trajectories (Stage-1 output).

Repository layout:

ISETrace/
├── README.md
├── LICENSE
├── trajectories/
│   ├── trajectories-00000.jsonl
│   ├── ...
│   └── trajectories-00007.jsonl   # 8 shards, 3,000 trajectories each (last shard smaller)
└── intents/
    └── intents.jsonl

Note on nested fields. To keep a uniform column schema (so the HF Dataset Viewer renders cleanly), messages[].tool_calls[].function.arguments and tools[].function.parameters are stored as JSON strings — call json.loads(...) on them downstream. This also matches the native OpenAI tool-call wire format, where arguments is a string. All other message structure is preserved as native objects.

from datasets import load_dataset

# Main trajectory corpus
traj = load_dataset("valiere/ISETrace", name="trajectories", split="train")

# The structured intent pool that seeds them
intents = load_dataset("valiere/ISETrace", name="intents", split="train")

The two configs are linked by intent id: every intents row id appears as an intent_id / source_intent_ids entry on the trajectory it seeds. All 43,955 released intents are referenced by at least one trajectory.

---

Corpus statistics

Statistic	Value
Trajectories	23,132
Unique structured intents	43,955
Avg. user turns / trajectory	8.12 (med. 8, max 23)
Trajectories with 6–10 user turns	91.1%
Avg. total dialogue turns / trajectory	68.24 (max 565)
Avg. tool calls / trajectory	29.26
Avg. unique tools / trajectory	4.69
Distinct tools	16
Total tool calls (released set)	676,901
Domains × Tasks × Tools (avg. per intent)	2.35 × 4.40 × 3.18
Complexity (complex / medium / simple)	50% / 40% / 10%
Full-pool Vendi Score (mpnet-base-v2, cosine, q=1)	61.57

Trajectories span 10 functional domains (Intelligence-Core, Code-Runtime, File-IO, Source-Chain, Automation-Flow, Web-Extraction, Comms-Social, Multimedia, Agent-Memory, System-Infrastructure) and 965 distinct personas across 47 industries.

---

Schema

trajectories

Each row is one complete agent session.

Field	Type	Description
status	string	Session terminal status (e.g. completed).
session_id	string	Anonymized stable surrogate id (traj_<hash>).
intent_id	string	Primary intent that finalized the session.
source_intent_count	int	Number of source intents fused into this session.
source_intent_ids	list[string]	Intent ids that seed this trajectory (join key to intents).
source_intents	list[object]	The source intents' intent_id + natural_language_intent + task_type.
session_finalized_by_intent_id	string	Which source intent closed the session.
total_steps	int	Number of agent steps.
enable_thinking	bool	Whether reasoning traces were enabled.
messages	list[object]	The full OpenAI-style message log (see below).
tools	list[object]	The 16 tool schemas available to the agent. Each is {type, function:{name, description, parameters}}; parameters is a JSON string (see note above).
final_output	string	The agent's final assistant message to the user.
metadata	object	{persona, domains} — the persona profile and the sampled domain/task definitions.

messages entries follow the standard chat format:

• role: system | user | assistant | tool
• content: string
• reasoning_content: string (assistant turns, when thinking enabled)
• tool_calls: list of {id, type, function:{name, arguments}} (assistant turns); arguments is a JSON string — json.loads(...) it to recover the call's parameter object.
• name, tool_call_id, success: (tool turns — which tool ran, for which call, and whether it succeeded)

Role distribution across the corpus: 23,132 system · 193,894 user · 777,916 assistant · 677,017 tool.

intents

Each row is one structured 4D intent.

Field	Type	Description
id	string	Intent id (join key to trajectories).
natural_language_intent	string	The rendered natural-language user request.
domains	list[string]	Sampled domain UUIDs.
tasks	list[string]	Sampled task UUIDs.
persona_id	string	Sampled persona UUID.
tools_used	list[string]	Tool UUIDs associated with the intent.
skills_used	list[string]	Skill UUIDs associated with the intent.
complexity	string	simple | medium | complex.
metadata	object	Human-readable expansion: {persona, domains} (name/role/industry, domain & task definitions).

---

Provenance, privacy & processing

• Execution grounding. Trajectories were produced by driving a local agent through real multi-turn user simulation and executing every tool call against an isolated OS workspace. Shell paths such as /root/.openclaw/workspaces/... appearing in exec/read/ write calls are authentic execution artifacts and are preserved deliberately — they are the evidence of grounding.
• De-identification. The release strips internal runtime telemetry that is irrelevant to training: the generation model/provider identity, per-run cost and token accounting, the runtime skill snapshot, and internal worker indices have been removed or replaced with neutral placeholders (<GEN_MODEL>, <GEN_PROVIDER>, <GEN_VENDOR>, <GEN_ENDPOINT>, worker). Trajectory logic, dialogue, tool calls and their real outputs are unchanged.
• Personas are synthetic. All persona names, roles, and contexts are LLM-synthesized and do not correspond to real individuals.
• Language. English.

---

Intended use & limitations

ISETrace targets supervised fine-tuning of OS/tool-use agents. Fine-tuning Qwen3-8B on ISETrace lifts ClawEval pass@1 from 19.3 → 37.7 (+18.4 pts, 1.95×), surpassing both a GPT-4o zero-shot reference and a 4×-larger Qwen3-32B base.

The corpus targets macOS/Linux OS-terminal interaction; it does not cover Windows, GUI-based interaction, or browser automation. It is a fixed-size checkpoint (23,132 trajectories), not a continuously updated dataset.

---

License & citation

The dataset is released under CC BY 4.0. The pipeline code is released separately under the MIT License (see the linked repositories).

@misc{isetrace2026,
  title        = {From Intent to Trajectory: Execution-Grounded Multi-Turn Data Synthesis for OS Agents},
  author       = {Valiere01},
  year         = {2026},
  howpublished = {\url{https://github.com/Valiere01/ISE-Trace}},
  note         = {Paper link forthcoming}
}