chataiopensource

Datasets

What the arena records

Anonymous comparisons are only useful as data if you know exactly what was kept. Here is the full list, including the two fields that can contain something you wrote or something about you.

This deployment right now

204 Battles started
25 Votes recorded
212 Conversation turns
51 Models competing

20 feedback reasons recorded across 10 prompt categories.

Where the battles went

  • Overall 149
  • Coding 9
  • Math 1
  • Knowledge 1
  • Long Context 1
  • Reasoning 0
  • Creative Writing 0
  • Roleplay 0
  • Hard Prompts 0
  • Multilingual 0

Every field, one by one

Field What it holds Can identify you?
Field What it holds Contains anything identifying you?
prompt The text you sent, cleaned and truncated. No
response_a / response_b Both model replies, in full. No
model_a_id / model_b_id Which models answered. Hidden from you until after you vote. No
winner_id Which side you chose, or null for a tie. No
both_bad 1 when you flagged both responses as unusable. No
category The prompt category, from a fixed list. No
temperature / max_tokens The sampling settings the battle ran under. No
style_preset / style_prompt The style you chose. Your own words, so treat it as yours. Yes, if you typed it
client_ip Your IP address at the time of the vote. Truncated to 45 characters. Yes
vote_feedback Why you picked that side, from a fixed list of reasons. No

What is never collected

  • No accounts, no login, no email address.
  • The application sets no cookies of its own and runs no analytics. The reference pages (this one, blog, FAQ, About, leaderboard) load a third-party ad tag; the battle, comparison, direct chat, shared battle and admin pages load no third-party JavaScript at all.
  • No prompt is shared with anyone for any purpose other than calling the two models.
  • The model provider does receive your prompt and the conversation history, because that is what answering it requires.

Getting the data out

Everything lives in one SQLite file, so an export is a file copy. Stop the web server first, or copy the file and the -wal and -shm files beside it.

# a consistent snapshot
sqlite3 database.db ".backup 'arena-snapshot.db'

# or the schema and rows as SQL
sqlite3 database.db .dump > arena.sql

# just the votes
sqlite3 -header -csv database.db \
  "SELECT id, model_a_id, model_b_id, winner_id, both_bad, category, temperature, voted_at FROM votes"

To drop everything and start over, delete database.db and reload schema.sql. To upgrade an existing database instead, run php migrate.php.