Datasets:

openzot
/

arcade

license: mit
pretty_name: zot arcade sessions
language:
  - en
tags:
  - agent-trajectories
  - tool-use
  - code-generation
  - games
  - zot
task_categories:
  - text-generation
size_categories:
  - n<1K
configs:
  - config_name: default
    data_files: trajectories/*/*.jsonl

zot arcade sessions

Every conversation behind every game in the zot arcade - a software factory where an agent takes the same standing order every half hour, reads the catalogue of what already exists, and designs, writes, playtests and ships one brand-new browser game.

Each row is one shift: the full agent trajectory from the order to the finished game (or to where the shift was cut short), in the chat shape the rest of the ecosystem reads, plus what the arcade knows about the result.

The rows are appended by the arcade's own workflow as shifts happen. Nobody reviews them; they are the record as it was written.

Layout

trajectories/<session-id>/<session-id>.jsonl   one row
trajectories/<session-id>/images/<digest>.png  the screenshots the model was shown

Row schema

field
`id`	the zot session the row was exported from
`chain`	the sessions behind it, oldest first - a shift cut short is continued by the next, and exports as one row
`task`	the order as the model received it
`model`, `provider`, `driver`	what ran it
`started`, `ended`	wall clock, UTC
`outcome`	how the run ended: `reason` (`success`, `failed`, `error`, ...), `iterations`, `calls`, ... - absent when the run was cut short
`complete`	whether an outcome was recorded
`messages`	the conversation as it stood at the end - see below
`snapshots`	earlier states of the conversation that compaction or a resume superseded, oldest first
`images`	the image files this row refers to, relative to the row's directory
`events`	counts by kind: iterations, nudges, retries
`arcade`	the arcade's side: `game` (the catalogue entry: slug, name, genre, mechanic, theme, tagline, controls, created), `files` (the game's `index.html`, `game.css`, `game.js`), `outcome` (the shift's verdict: `settled` / `failed` / `error`), `catalogue_check`, `committed`, `commit`, `run`

Messages

OpenAI chat convention, with additive fields:

{"role": "user",      "type": "user",        "content": "Begin working on your task..."}
{"role": "assistant", "type": "bot",         "content": "", "reasoning": "...", "tool_calls": [{"id": "call_1", "type": "function", "function": {"name": "read", "arguments": "{\"path\":\"site/games.json\"}"}}]}
{"role": "tool",      "type": "activity",    "tool_call_id": "call_1", "name": "read", "content": "[...]"}
{"role": "user",      "type": "attachment",  "content": [{"type": "text", "text": "screenshot of the game"}, {"type": "image", "image": "images/3f2a....png"}]}
{"role": "system",    "type": "checkpoint",  "content": "summary of the conversation so far"}

type is zot's own message type: user, bot, activity, attachment, checkpoint (a compaction summary), instructions.
reasoning is the model's scratchpad for the turn, when the provider surfaced it.
Tool arguments are the JSON string the model sent, verbatim.
The system prompt is not recorded; task is the brief.

messages is what a resume of the run would replay. After compaction that is a checkpoint plus the recent turns - snapshots holds the earlier states, so every turn that happened is in the row, at the price of the recent ones being repeated.

Filtering

Finished games: complete and arcade.outcome == "settled" and arcade.catalogue_check == "success".
A shift that was cut short ships as a partial row; the shift that continued it ships the whole chain under its own id, with the earlier id in chain. Drop rows whose id appears in another row's chain to keep only tips.

Provenance

Factory: https://github.com/openzot/arcade - the standing order is orders/new-game.yaml, the conventions the agent reads are AGENTS.md.
Harness: zot; rows are produced by zot sessions export and shipped by scripts/ship.py after each shift.
Model: whatever the workflow names at the time; each row says which.

The games are model output, published as-is under the arcade's license. Check the terms of the model named in a row before training on its reasoning.