Jurgen1161/synthetic-code-agent-trajectories

For buyers

To purchase the full dataset, contact me via telegram. @HecklerKoch416C

Synthetic Code Agent Trajectories (Sample)

125 synthetic agent trajectories for training and evaluating code agents in a Python sandbox.

Each trajectory is a complete, multi-step interaction where an AI agent solves a programming task using real tools: writing files, reading them, executing Python code in an isolated sandbox, and observing the actual output.

What's inside

• 125 complete trajectories (2–6 steps each)
• Real tool calls — not hallucinated; every run_python call was executed in a real E2B sandbox and the observation contains actual stdout/stderr
• 8 task categories: math_and_stats, data_processing, algorithms, file_operations, regex_and_text, debugging, testing, utilities
• ATIF-like format (Agent Trajectory Interchange Format)
• 100% synthetic — no PII, no copyright issues

Why this dataset is different

Most existing trajectory datasets focus on SWE-bench or OpenHands. This one is different:

• Python sandbox environment — every code execution is real, not simulated
• Real observations — stdout, stderr, and tracebacks come from E2B, not from the model's imagination
• Multi-step reasoning — agents write files, run code, fix errors, and iterate
• Tool calling — the agent uses run_python, read_file, write_file, list_files

Dataset statistics

Metric	Value
Total trajectories	125
Average steps per trajectory	3.6
Average tool calls per trajectory	3.0
Trajectories with 0 tool calls	0
Categories	8

Category distribution:

Category	Count
file_operations	30
debugging	27
data_processing	20
algorithms	14
testing	11
regex_and_text	10
math_and_stats	9
utilities	4

Schema

Each record is a JSON object with the following structure:

{
  "schema_version": "ATIF-v0.1",
  "session_id": "code-0001-1789651957",
  "task": {
    "id": "code-0001",
    "category": "math_and_stats",
    "difficulty": "easy",
    "description": "Calculate the sum, mean, minimum, and maximum of the numbers [12, 7, 19, 4, 23, 15] and return the four results."
  },
  "agent": {
    "name": "code-agent",
    "model": "qwen3.6-flash",
    "temperature": 0.3
  },
  "environment": {
    "type": "e2b-sandbox",
    "tools": ["run_python", "read_file", "write_file", "list_files"]
  },
  "steps": [
    {
      "step_id": 1,
      "source": "agent",
      "message": "",
      "tool_calls": [
        {
          "id": "call_...",
          "name": "run_python",
          "arguments": {"code": "..."}
        }
      ],
      "observation": [
        {
          "tool_call_id": "call_...",
          "name": "run_python",
          "result": {
            "stdout": "...",
            "stderr": "",
            "error": null
          }
        }
      ]
    }
  ],
  "final_answer": "...",
  "metrics": {
    "num_steps": 2,
    "num_tool_calls": 1,
    "success": true,
    "error": null,
    "started_at": "...",
    "finished_at": "..."
  }
}

Field reference

Field	Type	Description
schema_version	string	ATIF schema version
session_id	string	Unique session identifier
task.id	string	Task identifier
task.category	string	Task category
task.difficulty	string	easy, medium, hard
task.description	string	Full task instruction
agent.name	string	Agent name
agent.model	string	Model used
agent.temperature	float	Sampling temperature
environment.type	string	Sandbox type (e2b-sandbox)
environment.tools	array	Available tools
steps	array	Sequence of agent actions
steps[].step_id	int	Step index
steps[].source	string	user or agent
steps[].message	string	Agent reasoning (may be empty)
steps[].tool_calls	array	Tool invocations
steps[].observation	array	Tool results
final_answer	string	Final response from the agent
metrics	object	Run metrics

Example

{"schema_version":"ATIF-v0.1","session_id":"code-0001-1789651957","task":{"id":"code-0001","category":"math_and_stats","difficulty":"easy","description":"Calculate the sum, mean, minimum, and maximum of the numbers [12, 7, 19, 4, 23, 15] and return the four results."},"agent":{"name":"code-agent","model":"qwen3.6-flash","temperature":0.3},"environment":{"type":"e2b-sandbox","tools":["run_python","read_file","write_file","list_files"]},"steps":[{"step_id":1,"source":"agent","message":"","tool_calls":[{"id":"call_abc123","name":"run_python","arguments":{"code":"numbers = [12, 7, 19, 4, 23, 15]\nprint(sum(numbers))\nprint(sum(numbers)/len(numbers))\nprint(min(numbers))\nprint(max(numbers))"}}],"observation":[{"tool_call_id":"call_abc123","name":"run_python","result":{"stdout":"80\n13.333333333333334\n4\n23\n","stderr":"","error":null}}]},{"step_id":2,"source":"agent","message":"Here are the calculated results:\n- Sum: 80\n- Mean: 13.33\n- Minimum: 4\n- Maximum: 23","tool_calls":[],"observation":null}],"final_answer":"Here are the calculated results:\n- Sum: 80\n- Mean: 13.33\n- Minimum: 4\n- Maximum: 23","metrics":{"num_steps":2,"num_tool_calls":1,"success":true,"error":null,"started_at":"2026-09-17T13:32:37Z","finished_at":"2026-09-17T13:33:07Z"}}

Usage

from datasets import load_dataset

ds = load_dataset("jurgen1161/synthetic-code-agent-trajectories")
print(ds["train"][0])

# Filter debugging trajectories
debug = [x for x in ds["train"] if x["task"]["category"] == "debugging"]
print(f"Debugging trajectories: {len(debug)}")

# Count multi-step trajectories
multi = [x for x in ds["train"] if x["metrics"]["num_steps"] >= 4]
print(f"Multi-step (>=4 steps): {len(multi)}")

Use cases

• Fine-tuning code agents — teach models to use tools, iterate on errors, and produce verifiable solutions
• RLHF / DPO — trajectories with success=true as positive examples
• Agent evaluation — a held-out benchmark for tool-use correctness
• Reasoning distillation — extract step-by-step reasoning patterns
• Sandbox debugging — study how agents handle real Python errors and tracebacks

Generation details

• Agent model: qwen3.6-flash via MixRoute API
• Sandbox: E2B isolated Python environment
• Temperature: 0.3
• Tools available: run_python, read_file, write_file, list_files
• Max steps per trajectory: 15

License

CC BY-NC 4.0 — free for research and evaluation. Commercial use is prohibited.

For the full commercial license with 500+ trajectories, contact the author.

Citation

@misc{synthetic-code-agent-trajectories,
  title={Synthetic Code Agent Trajectories},
  author={jurgen1161},
  year={2026},
  url={https://huggingface.co/datasets/jurgen1161/synthetic-code-agent-trajectories}
}