textcraft-synth-single-agent

TextCraft (synthetic) rollouts from a single (linear) Qwen3-4B agent. The agent was trained with a 40K context window; these rollouts were run without a context limit.

There is one row per rollout. The full event log for that rollout is in the events column.

import json
from datasets import load_dataset

ds = load_dataset("apurvaga/textcraft-synth-single-agent", split="train")
row = ds[0]
for e in row["events"]:
    if e["type"] == "trajectory_step_added":
        step = json.loads(e["step"])  # thought, code, output, error, reward, misc

Columns

column	description
task_id, rollout_index	which task and rollout this row is
status, error, wall_time_seconds, created_at	from the rollout's metadata.json
collection_id	ID of the trajectory collection (root agent and its sub-agents)
reward	the root agent's final reward
num_trajectories	number of agents in the rollout (1 plus the number of sub-agents)
num_steps, num_events	step and event counts across all agents
events	the ordered event log (see below)

Each entry in events has type, ts (Unix timestamp), process_id, collection_id, trajectory_id, step_index, reward, finish_message, error_message, trajectory, task and step. Fields that don't apply to an event type are null. trajectory, task, step and finish_message are JSON-encoded strings; decode them with json.loads.

event type	fields set
trajectory_created	trajectory (initial trajectory; parent_info is null for the root agent and set for sub-agents)
trajectory_task_set	task (goal, id, max_steps, misc)
trajectory_step_added	step_index, step (thought, code, output, error, reward, misc)
trajectory_finished	reward, finish_message, error_message

Within a row, events interleaves events from the root agent and all of its sub-agents in time order. To get one agent's trajectory, filter by trajectory_id; each sub-agent's parent_info (in its trajectory_created event) links it to the agent that spawned it. Single-agent datasets have one trajectory per row. The task in trajectory_task_set is the state at the start of the run (for TextCraft, task.misc.initial_inventory is the true starting inventory).

Viewing with platoon's visualization tool

The platoon viewer reads JSONL event logs. hf_to_events.py (in this repo) writes each row back out as events_<task_id>_<collection_id>.jsonl, byte-for-byte identical to the original log:

pip install datasets
python hf_to_events.py apurvaga/textcraft-synth-single-agent ./events
uv run -m platoon.visualization.cli tail --dir ./events

tail loads every event at once, so you see the full trajectory trees immediately. To watch a single run play out step by step, use replay <file.jsonl> --delay 0.25 instead.