davidquarel/connect4-pons-eval

Connect-4 Pons-labelled policy eval

A frozen evaluation set of 6705 decisive Connect-4 positions, each labelled by Pascal Pons' perfect solver with the game-theoretic per-column scores. Used by the ARENA [2.5] MCTS & AlphaZero material to measure how close a trained policy/value net is to perfect play (net-independent ground truth), via pons/acc, pons/ce, pons/val_signacc.

File

• pons_eval_dataset.json — JSON, {"meta": {...}, "positions": [...]}.

Schema (per position)

• moves: string of 0-indexed columns (0..6) played from the empty board (red moves on even plies).
• scores: length-7 list, Pons WEAK analyze per column from the mover's perspective — sign = outcome class (>0 win / 0 draw / <0 loss); -1000 = illegal/full column.
• depth: number of moves played (ply count) to reach the position.

Every position is decisive: the mover can reach ≥2 distinct outcome classes, so a non-optimal move genuinely throws away the achievable result. Plies range 2–36.

Load

import json
from huggingface_hub import hf_hub_download
p = hf_hub_download("davidquarel/connect4-pons-eval", "pons_eval_dataset.json", repo_type="dataset")
data = json.load(open(p))

Regenerate with the solver + build_dataset.py in the ARENA repo (pascal_pons/).