Document Match and Ranked tabs

README.md

@@ -21,8 +21,12 @@ tags:
 
 Can a language model that only ever read text (e.g. FineWeb-edu) play Tetris **without any training on the game**?
 
-Pick 2+ decoder-only models (50K–250M parameters, custom architectures welcome), watch them play the **same piece
-sequence** side by side on CPU, and see their **Elo** change on a public leaderboard.
+Decoder-only models (50K–250M parameters, custom architectures welcome) play the **same piece sequence** side by
+side on CPU. Two ways to play:
+
+- **Match** (friendly): pick any 2+ models and the seed. Nothing is recorded.
+- **Ranked**: press Play and the arena picks up to 4 models **at random** from its pool, all within 20M parameters of
+  each other. The result updates a public **Elo** leaderboard.
 
 ## How a model plays
 
@@ -54,11 +58,14 @@ Each protocol has its own leaderboard.
 
 ## Elo
 
-Ranked matches use a random seed. Placement is decided by score (100/300/500/800 for 1–4 lines), then lines, then
-pieces survived (cap: 500 pieces). Multiplayer Elo, K = 32: each pair of players is a game, scaled by 1/(N−1).
-Only language models whose sizes are within 20M parameters of each other can play ranked (so a big model can't
-farm Elo from tiny ones); a ranked match outside that range is blocked (it can still be played unranked).
-Baselines are never rated.
+Only the Ranked tab changes Elo. The arena picks the players at random from the suggested models (models with fewer
+ranked games are more likely to be picked), all within 20M parameters of each other, and uses a random seed. Nobody
+chooses who plays ranked, so Elo can't be farmed by pairing a model with weak opponents. The match runs on the server
+in the background: it finishes and counts even if the viewer leaves, and only one ranked match runs at a time
+(pressing Play while one is running lets you watch it).
+
+Placement is decided by score (100/300/500/800 for 1–4 lines), then lines, then pieces survived (cap: 500 pieces).
+Multiplayer Elo, K = 32: each pair of players is a game, scaled by 1/(N−1). Baselines are never rated.
 Every ranked match (seed, model commit SHAs, scores, Elo before/after) is stored in the
 [results dataset](https://huggingface.co/datasets/DedeProGames/lm-tetris-arena-results).
 
@@ -72,7 +79,7 @@ Every ranked match (seed, model commit SHAs, scores, Elo before/after) is stored
 | `MAX_PARAMS` | `250000000` | Parameter limit |
 | `MIN_PARAMS` | `50000` | Smallest model allowed (smaller ones are rejected and removed from the leaderboard) |
 | `MAX_PIECES` | `500` | Piece cap per game |
-| `MAX_PARAM_GAP` | `20000000` | Largest size difference allowed in a ranked match |
+| `MAX_PARAM_GAP` | `20000000` | Largest size difference between models picked for a ranked match |
 | `ALLOW_REMOTE_CODE` | `1` | Allow models with custom code (`trust_remote_code`) |
 | `TORCH_THREADS` | `2` | CPU threads for inference |