MiniChat export

Structured export of MiniChat conversations for dataset use.

Contents

data/
├── archives/
│   └── all-chats/  # 12 original legacy export files, preserved verbatim
├── chats/
│   └── markdown/   # readable transcripts, one file per chat
├── parquet/        # analysis-friendly tables for browsing/querying
│   ├── chats.parquet
│   ├── messages.parquet
│   ├── blocks.parquet
│   ├── queries.parquet
│   └── transcripts.parquet
└── queries/
    └── <export-group>/
        ├── user-queries.jsonl
        └── user-queries.txt
tools/
└── export-user-chats.mjs

The original source export contained 261 chats, 3,100 user-query records, 261 JSON records, and 261 Markdown transcripts. The additional all-chats.zip contained 5,089 legacy records across 12 files. Those files represent 2,523 unique chat IDs after snapshot deduplication; the newest updated_at snapshot is used for each canonical record. The original files are preserved verbatim under data/archives/all-chats/.

The merged dataset contains 5,338 canonical chats, 5,338 embedded Markdown transcripts, and 14,791 indexed user queries. Query indexes remain separated by export group so source grouping is not lost.

For convenient analysis, the Parquet layer provides one row per chat in chats.parquet, one row per message in messages.parquet, one row per content/reasoning/tool block in blocks.parquet, and one row per user query in queries.parquet. transcripts.parquet contains the complete readable Markdown transcript in its markdown column, keyed by chat_id. The files use Zstandard compression and are tracked with Git LFS.

Local browser preview

With Python and PyArrow installed, run:

python3 tools/preview-server.py

Then open http://127.0.0.1:8765. The viewer supports chat search, query search, message inspection, source metadata, and dataset row counts.

Data format

Each file in data/chats/json/ is one chat object with these top-level fields:

• id, creatorId, createdAt, updatedAt, title
• origin, boardId, mode, status
• showNewBoardSuggestion, hasUnread
• messages

Each message retains the original message metadata. User messages contain a plain-text content; assistant messages contain blocks, including content, reasoning, and tool blocks when present. The matching Markdown file renders a readable transcript from the same record.

Each JSONL query row has this schema:

{
  "chatId": "...",
  "chatTitle": "...",
  "messageId": "...",
  "createdAt": "...",
  "query": "..."
}

Legacy records are normalized from the archive format (user_id, chat, and chat.history.messages) into the canonical tables above. Their original source filename is retained in the chat metadata, while the raw archive and per-chat JSON exports are kept out of the published Git history to avoid redundant LFS objects.

All JSON and JSONL files are tracked with Git LFS because the export is large. Credentials and local runtime files from the source directory (.cookie, .cookie-prod, and .env) are intentionally excluded.