{"name":"hangzhou-mahjong","repo":"MaxwellCCC/hangzhou-mahjong","url":"https://midorreal.com/project/maxwellccc-hangzhou-mahjong","repository":"https://github.com/MaxwellCCC/hangzhou-mahjong","description":"Hangzhou Mahjong training and evaluation framework with public-information Transformers, paired branch reinforcement learning, and the r2600 checkpoint.","facts":{"language":{"value":"Python","source":"https://github.com/MaxwellCCC/hangzhou-mahjong","checked":"22 Nov 2026"},"releases":{"value":"0","source":"https://github.com/MaxwellCCC/hangzhou-mahjong/releases","checked":"19 Nov 2026"},"contributors":{"value":"1","source":"https://github.com/MaxwellCCC/hangzhou-mahjong/graphs/contributors","checked":"19 Nov 2026"},"archived":{"value":"no","source":"https://github.com/MaxwellCCC/hangzhou-mahjong","checked":"22 Nov 2026"}},"stars":1.0,"writeup":{"what_it_is":"A Hangzhou Mahjong training and evaluation framework with a 52.01-million-parameter Transformer-based policy agent. It includes a game simulator, reinforcement learning training system using paired branch exploration, and a released r2600 checkpoint that ranked first in a 36-participant tournament.","audience":"Researchers studying imperfect-information games and reinforcement learning.","claims":[{"kind":"specific","claim":"The policy model has 52.01 million parameters","excerpt":"The 52.01-million-parameter policy combines a tile-relation Transformer, a causal event encoder, legal-action queries, public tactical successor features, and compact opponent memory.","status":"not_checked"},{"kind":"specific","claim":"The r2600 checkpoint ranked first in a 36-participant tournament with 1,081,008 games","excerpt":"In a completed 36-participant tournament comprising 1,081,008 games, with 120,112 seat-games per participant, the released r2600 checkpoint ranked first under the predefined weighted score, at +0.452171 per seat-game.","status":"not_checked"},{"kind":"specific","claim":"The tournament result shows a narrow margin between first and second place with overlapping confidence intervals","excerpt":"The next participant scored +0.430258; their difference was +0.021914, with an exploratory 50% bootstrap interval of [−0.024367, +0.069381]. The interval does not establish a reliable separation between the two.","status":"not_checked"},{"kind":"specific","claim":"Uses paired branch reinforcement learning with exploratory continuations","excerpt":"Each eligible decision on a reference trajectory produces five exploratory continuations. Their realized scores are compared with a reference over the same game window, and an empirical clipped objective updates the exploration distribution using its recorded temperature and tactical bias.","status":"not_checked"},{"kind":"specific","claim":"The framework does not include historical rule opponents, other neural weights, or the demonstration corpus","excerpt":"Historical rule opponents, other neural weights, and the demonstration corpus are not included.","status":"not_checked"}],"alternatives":[{"name":"verl","url":"https://midorreal.com/project/xiaomimimo-verl"},{"name":"agent-sandbox","url":"https://midorreal.com/project/kubernetes-sigs-agent-sandbox"}],"written_by":"AI, from the project's README","date":"2026-10-15"},"why_now":null,"owner_supplied":null,"score":{"verdict":"mid","hype":2,"reality":12,"gap":-9,"momentum":null,"confidence":60,"version":"score-2.0","rule":"verdict-2.0","calculated_at":"2026-11-22T08:00:00+00:00","verdict_source":"rule"},"early_read":null,"changes":[{"date":"2026-10-10T14:37:11+00:00","text":"Repository created","source":"https://github.com/MaxwellCCC/hangzhou-mahjong"}],"cite":"hangzhou-mahjong is a GitHub project at https://github.com/MaxwellCCC/hangzhou-mahjong written mainly in Python, described there as \"Hangzhou Mahjong training and evaluation framework with public-information Transformers, paired branch reinforcement learning, and the r2600 checkpoint\".","attribution":"Data from Mid or Real, https://midorreal.com/project/maxwellccc-hangzhou-mahjong, checked 22 November 2026. Attribution with a link is required.","still_real":null}