tldraw-datasets

Snapshot history datasets from tldraw rooms. Each row is a single editing session (a "trajectory") from one room, captured as a sequence of periodic JSON snapshots with matching rendered PNGs.

Source and tooling: https://github.com/tldraw/tldraw-datasets

Schema

One row per room. Each row has three columns:

column	type	description
filename	string	e.g. 1dUMRx3oRxs33vPdsy2uY.jsonl.
jsonl	Sequence[struct]	Row 0 is the initial full snapshot; rows 1..N are per-change diffs.
images	Sequence[Image]	1:1 with jsonl. images[0] is the initial snapshot PNG; images[i>0] is the state after diff i.

Rows where the diff has no document changes (idle periods) are dropped from both jsonl and images, so each kept frame corresponds to a real edit.

jsonl item schema

Each element of jsonl is a struct. Fields that only apply to snapshots or diffs are empty on the other kind.

field	type	notes
kind	string	"snapshot" on row 0, "diff" on rows 1..
ts	string	ISO timestamp with : → - (lex sort = chronological).
prev	string | null	Previous snapshot's ts. null on row 0.
clock	int64
documentClock	int64
tombstoneHistoryStartsAtClock	int64
schemaJson	string	JSON-encoded tldraw schema block.
documents (snapshot only)	Sequence[string]	JSON-encoded {state, lastChangedClock} records.
tombstones (snapshot only)	Sequence[struct]	{id, clock}.
documentsAdded (diff only)	Sequence[string]	JSON-encoded records.
documentsModified (diff only)	Sequence[string]	JSON-encoded records.
documentsRemoved (diff only)	Sequence[string]	Record IDs.
tombstonesAdded (diff only)	Sequence[struct]	{id, clock}.
tombstonesRemoved (diff only)	Sequence[string]	Record IDs.

Why JSON strings for documents / schema? tldraw shape types have widely varying schemas (geo, draw, arrow, image, embed, …). Unioning all their props into one struct produces a combinatorial schema that breaks parquet. The record payloads are stored as JSON strings; call json.loads(...) to recover the original tldraw record. See the RoomSnapshot type for the deserialized shape.

Usage

from datasets import load_dataset

ds = load_dataset("steveruizoktldraw/tldraw-datasets", split="train")
row = ds[0]
print(row["filename"])        # 1dUMRx3oRxs33vPdsy2uY.jsonl
print(len(row["jsonl"]))      # number of keyframes in the trajectory
print(row["jsonl"][0]["kind"])  # 'snapshot'
print(row["images"][0])       # PIL.Image of the initial state

# Reconstruct the nth intermediate tldraw state:
import json
initial_docs = [json.loads(d) for d in row["jsonl"][0]["documents"]]

Reconstructing a full RoomSnapshot

Replay semantics (apply diffs in order onto the initial snapshot state):

import json

def reconstruct(jsonl, up_to: int):
    initial = jsonl[0]
    docs = {json.loads(d)["state"]["id"]: json.loads(d) for d in initial["documents"]}
    tombs = {t["id"]: t["clock"] for t in initial["tombstones"]}
    schema = json.loads(initial["schemaJson"])

    for step in jsonl[1 : up_to + 1]:
        for s in step["documentsAdded"] + step["documentsModified"]:
            r = json.loads(s)
            docs[r["state"]["id"]] = r
        for rid in step["documentsRemoved"]:
            docs.pop(rid, None)
        for t in step["tombstonesAdded"]:
            tombs[t["id"]] = t["clock"]
        for rid in step["tombstonesRemoved"]:
            tombs.pop(rid, None)
        schema = json.loads(step["schemaJson"])

    last = jsonl[up_to]
    return {
        "clock": last["clock"],
        "documentClock": last["documentClock"],
        "tombstoneHistoryStartsAtClock": last["tombstoneHistoryStartsAtClock"],
        "schema": schema,
        "tombstones": tombs,
        "documents": list(docs.values()),
    }

License

MIT — see LICENSE.