{"id":13993509,"url":"https://github.com/pangafu/LeelaMasterWeight","last_synced_at":"2025-07-22T17:32:46.237Z","repository":{"id":88317062,"uuid":"126411503","full_name":"pangafu/LeelaMasterWeight","owner":"pangafu","description":"Leela Master weight is training from leela zero self-play sgf and human sgf file","archived":false,"fork":false,"pushed_at":"2019-08-28T08:29:26.000Z","size":481,"stargazers_count":49,"open_issues_count":17,"forks_count":6,"subscribers_count":10,"default_branch":"master","last_synced_at":"2024-08-10T14:13:56.856Z","etag":null,"topics":["human","leela","master","sgf","zero"],"latest_commit_sha":null,"homepage":"https://drive.google.com/drive/folders/1bB8ee1wFuRWL9nPhsl4_BPUhcWSBuxO0?usp=sharing","language":null,"has_issues":true,"has_wiki":null,"has_pages":null,"mirror_url":null,"source_name":null,"license":"gpl-3.0","status":null,"scm":"git","pull_requests_enabled":true,"icon_url":"https://github.com/pangafu.png","metadata":{"files":{"readme":"README.md","changelog":null,"contributing":null,"funding":null,"license":"LICENSE","code_of_conduct":null,"threat_model":null,"audit":null,"citation":null,"codeowners":null,"security":null,"support":null,"governance":null,"roadmap":null,"authors":null}},"created_at":"2018-03-23T00:37:50.000Z","updated_at":"2024-04-09T07:42:13.000Z","dependencies_parsed_at":null,"dependency_job_id":"b8d197c3-2241-4e95-b7c3-f07fc7845538","html_url":"https://github.com/pangafu/LeelaMasterWeight","commit_stats":null,"previous_names":[],"tags_count":0,"template":false,"template_full_name":null,"repository_url":"https://repos.ecosyste.ms/api/v1/hosts/GitHub/repositories/pangafu%2FLeelaMasterWeight","tags_url":"https://repos.ecosyste.ms/api/v1/hosts/GitHub/repositories/pangafu%2FLeelaMasterWeight/tags","releases_url":"https://repos.ecosyste.ms/api/v1/hosts/GitHub/repositories/pangafu%2FLeelaMasterWeight/releases","manifests_url":"https://repos.ecosyste.ms/api/v1/hosts/GitHub/repositories/pangafu%2FLeelaMasterWeight/manifests","owner_url":"https://repos.ecosyste.ms/api/v1/hosts/GitHub/owners/pangafu","download_url":"https://codeload.github.com/pangafu/LeelaMasterWeight/tar.gz/refs/heads/master","host":{"name":"GitHub","url":"https://github.com","kind":"github","repositories_count":227143720,"owners_count":17737216,"icon_url":"https://github.com/github.png","version":null,"created_at":"2022-05-30T11:31:42.601Z","updated_at":"2022-07-04T15:15:14.044Z","host_url":"https://repos.ecosyste.ms/api/v1/hosts/GitHub","repositories_url":"https://repos.ecosyste.ms/api/v1/hosts/GitHub/repositories","repository_names_url":"https://repos.ecosyste.ms/api/v1/hosts/GitHub/repository_names","owners_url":"https://repos.ecosyste.ms/api/v1/hosts/GitHub/owners"}},"keywords":["human","leela","master","sgf","zero"],"created_at":"2024-08-09T14:02:24.781Z","updated_at":"2024-11-29T14:33:10.715Z","avatar_url":"https://github.com/pangafu.png","language":null,"funding_links":[],"categories":["Others"],"sub_categories":[],"readme":"# Leela Master Weight\nLeela Master weight is training from leela zero self-play sgf and human sgf file, since it too big, I will upload weight file to \n\nhttps://drive.google.com/drive/folders/1bB8ee1wFuRWL9nPhsl4_BPUhcWSBuxO0?usp=sharing\n\n\nLeela Master OX is 30*256 network with Reinforcement learning to make OZ winrate more accurate\n\nLeela Master OZ is 30*256 network(last is OZ38)\n\n    OZ18 has make some handicap test, please see OZ18说明.txt\n\nLeela Master E is 10*128 network(complete)\n\nLeela Master W is 20*128 network(complete)\n\nLeela Master G/GX is 15*192 network(complete)\n\nLeela Zero Z is 30*256 network(complete)\n\nLeela Master B is 30*256 network(complete)\n\nWelcome to our discuss qq group: 693862763\n\n# About LeelaMaster strength (ELO)\n\nGo AI Ratings (by @breakwa11)\n\nhttps://github.com/breakwa11/GoAIRatings\n\nHome-made Elo ratings for some engines (by xela@lifein19x19.com)\n\nhttps://lifein19x19.com/viewtopic.php?f=18\u0026t=16086\n\nStrength of the Leela Master G13:\n\nhttp://zero.sjeng.org/match-games/5b26c77d48e3e5462acf0c4a\n\n\n# About leelazero sgf \nLeelaZero Homepage : http://zero.sjeng.org/\n\nLeelaZero Github: https://github.com/gcp/leela-zero\n\nLeelaZero Training set: https://leela.online-go.com/training/\n\nLeelaZero sgf pack: http://sjeng.org/zero\n\n\n# About human sgf \nThe human sgf is mainly download from https://github.com/yenw/computer-go-dataset\n\nI used AI, Professional, TYGEM, Tom and CGOS game, mix with leelazero sgf.\n\n\n# About handicap game（关于让子棋）\nIt report that GX37/GX38 - GX6x serial suit for handicap game with komi version leelazero by @alreadydone ~ So you can try handicap games by GX37 or newer.\n\nThread: https://github.com/gcp/leela-zero/issues/1599 \n\nkomi version leelazero(by @alreadydone): https://github.com/alreadydone/lz/tree/komi\n\nrelease version: https://github.com/alreadydone/lz/releases\n\n[0308 updated] OZ-serial with komi engine(white) vs zen7(black), can handicap 4 stone and komi 0\n\n[0327 updated] upload LeelaZero_DKomi+Filter to google share folder, if you can read chinese, you can setup a very powerful hadicape engine, maybe the world most powerful handicap engine untile 2019/03/27.\n\n# About LeelaZero_DKomi+Filter\n\nLeelaZero_DKomi+Filter is LeelaZero with dyminate komi, handicape filter and color-plan komi engine, maybe the world most powerful handicap engine untile 2019/03/27.\n\nHow to use:\n\n1. There is a chinese doc in the folder.\n\n2. Simple guide:\n\n   a) Run parameter:\n   \n      leelaz -m 12 -g -t 8 -r 1 --km-player=1 --km-startmovenum=7 --km-filterstep=40 --km-s1 --km-s1-bias=82.5 --km-s1-step=3.75 --km-s1-maxwr=0.35 --km-s1-minwr=0.05 --km-s2 --km-s2-target=15 --km-s2-step=0.5 --km-s2-maxwr=0.80 --km-s2-minwr=0.55 --km-s3 --km-s3-step=0.5 --km-s3-maxwr=0.80 --km-s3-minwr=0.55 -w OZ.txt.gz\n     \n   b) Adjust parameter:\n   \n   km-startmovenum=hadicape * 2-1\n   \n   km-filterstep=hadicape * 10\n   \n   km-s1-bias Adjust to winrate 15% when OZ(white) play first move\n\n\n# About OX-serial\nOX-serial use Reinforcement learning from OZ38 to make winrate more accurate, last is OX24 \n\n\n# About OZ-serial\nOZ-serial is awesome in handicap mode , in my test , OZ13 with LeelaZero_DKomi+Filter engine(white) vs zen7(black), can handicap 4 stone and komi 0\n\nOZ serial is start from all-zero network, has some special supervisor-training parameter:\n\n1. Traning color plan:(thanks @alreadydone @Hersmunch) \n\n   Color plan can be used for handicape game, please refer to \n   \n   a. https://github.com/gcp/leela-zero/issues/1599 \n   \n   b. https://github.com/gcp/leela-zero/pull/1825\n   \n   c. https://github.com/alreadydone/lz/tree/stm4komi\n      https://github.com/Hersmunch/leela-zero/tree/komi\n   \n2. Split traning. (thanks @icee)\n\n   After compare GX-Serial, it seems that zero sgf is good at opening and human is good at ending game, so I split the traning data, and use more zero data at opening game, more human data at ending game.\n   \n3. Balance training.\n\n   The 80%  exist go sgf is end before 200 step, so the traning data is poor after 200 step. I sampling more at ending game to balance traning.\n   \nIt highly recommend use OZ-serial network with @alreadydone @Hersmunch komi version leelazero branch:\n   \n   https://github.com/Hersmunch/leela-zero/tree/komi\n   \n   https://github.com/alreadydone/lz\n   branch: komi+batch, komi+next, komi+tensorbatch, komi ....\n   \n[update 20190327] OZ with LeelaZero_DKomi+Filter engine is powerful! \n\nPlease enjoy the different go game style. \n\nMore detail please see https://github.com/pangafu/LeelaMasterWeight/blob/master/LeelaMasterOZ.md\n\n\n[update 0401]Some handicap test result(OZ18说明.txt):\n\n    b) 测试结果:\n   \n      OZ18(多卡15wpo)  vs  沙包3 (Zen6 maxsim 12000), 让4子 贴0,  W+R\n      \n      OZ18(多卡15wpo)  vs  沙包4 (Zen7 maxsim 6000), 让3子 贴0,  W+R\n      \n      OZ18(多卡15wpo)  vs  沙包4 (Zen7 maxsim 6000), 让4子 贴0,  W+R\n      \n      OZ18(多卡15wpo)  vs  沙包4 (Zen7 maxsim 6000), 让4子 贴0,  W+R\n      \n      OZ18(多卡15wpo)  vs  沙包6 激进 (GX37 2500po S3 不退让), 让3子 贴0, W+R\n      \n      OZ18(多卡15wpo)  vs  沙包6 激进 (GX37 2500po S3 不退让), 让4子 贴0, B+R\n      \n      OZ18(多卡15wpo)  vs  沙包7 激进 (GX37 12500po S3 不退让), 让3子 贴0, B+R\n      \n      OZ18(多卡15wpo)  vs  沙包7 激进 (GX37 12500po S3 不退让), 让2子 贴0, W+R\n      \n      OZ18(单卡10wpo)  vs  沙包6 激进 (GX37 2500po S3 不退让), 让3子 贴0, W+R\n      \n      OZ18(单卡10wpo)  vs  沙包6 激进 (GX37 2500po S3 不退让), 让4子 贴0, W+R\n      \n      OZ18(单卡10wpo)  vs  沙包7 激进 (GX37 12500po S3 不退让), 让3子 贴0, W+R\n      \n      OZ18(单卡10wpo)  vs  沙包7 激进 (GX37 12500po S3 不退让), 让4子 贴0, W+R\n      \n      OZ18(单卡10wpo)  vs  沙包7 保守 (GX37 12500po 原版引擎), 让4子 贴0, B+R\n      \n      OZ18(单卡10wpo)  vs  沙包7 保守 (GX37 12500po 原版引擎), 让3子 贴0, W+R\n      \n      OZ18(单卡10wpo)  vs  沙包6 保守 (GX37 12500po 原版引擎), 让4子 贴0, B+R\n      \n      OZ18(单卡13wpo)  vs  沙包8(Zen7 S 10w), 让3子 贴0, B+R\n      \n    c) 测试结论：\n   \n      1) 让子在单卡或者低batch情况下，可能棋力更高，多卡大batch可能会显著降低让子棋力，原因未知，可能是让子情况训练的少，不适合大规模乱撒点 (低batch未测试) \n      \n      2) gx或者lz目前训练权重的方法，在非komi=7.5的时候，可能会大幅降低棋力。\n      \n      3) 如果zen7 10w确实分先打不赢GX37 12500po，那么传统算法可能更适合被让子的情况，或者说现在的zero算法只适合分先\n\n\n\n[update 20190411] OZ14 komi vs winrate ( thanks @zliu1022 )\n\n    空棋盘:  OZ14 black\n\n    让1子:   OZ14 white, black move 1 stone\n    \n    让2子：  OZ14 white, handicap black 2 stone\n    \n    让3子：  OZ14 white, handicap black 3 stone\n    \n    让4子：  OZ14 white, handicap black 4 stone\n    \n    终局：   OZ14 at very close end game( W+0.5 @ komi 7.5)\n\n![OZ14](/pictures/OZ14.jpg?raw=true)\n\nCompared to untrained color plan weight(GX5B):\n\n![GX5B](/pictures/GX5B.jpg?raw=true)\n\nIt indicate color plan is suit for komi and handicap game~\n\n# About GX-serial\n\nGX1x is 10% human style game\n\nGX2x is 20% human style game\n\nGX3x is 30% human style game\n\nGX4x is 40% human style game \n\nGX5x is 50% human style game \n\nGX6x is 60% human style game \n\nGX7x is 70% human style game \n\nGX8x is 80% human style game \n\nGX9x is 90% human style game \n\nGXAx is 100% human style game \n\n\nAnd I will keep to increase human style game percent to see when the strength of weight will decent, and in our test, the style of leela master weight changed very frequently between each weight, and the strength is almost same. So you can test each weight to find the style you love.\n\n# About W/E-serial\nW/E-serial is training with 70% human style game + 30% leelazero sgf,\n\nthe strength is equal to same network size of leela-zero.\n\n\n# About G-serial\nG Serial now merged into GX Serial\n\nG01 - G03 : 90% leelazero sgf + 10% human style game\n\nG04 - G08 : 80% leelazero sgf + 20% human style game\n\nG09 - G13 : 70% leelazero sgf + 30% human style game\n\nThe training set is keeping change by random sample for each 256k step.\n\n\n# About B-serial\nB-serial is traing from leelazero + human sgf in 30 * 256 network size, it based on z-serial network.\n\nB01 - B03 : 80% leelazero sgf + 20% human style game\n\nCurrent learning rate is 0.0005\n\n\n# About Z-serial\nZ-serial is training from leelazero sgf in 30 * 256 network size.\n\nThe last learning rate is 0.0005, and the traning set is random sample from leela zero's last 3,000,000 games\n\nZ-serial's end network is used to make B-serial.\n\n\n\n# About Octopus (章鱼围棋)\nOctopus go another way to make human style AI, we will try some different techology and some new idea, it not same with leela master, and it will opensource when ready.\n\n章鱼围棋会用另外一条思路去做人谱AI， 会采用一些新的技术和尝试一些新的想法，请不要用leelamaster的棋力去衡量章鱼围棋，当整个成熟后，会考虑开源。\n","project_url":"https://awesome.ecosyste.ms/api/v1/projects/github.com%2Fpangafu%2FLeelaMasterWeight","html_url":"https://awesome.ecosyste.ms/projects/github.com%2Fpangafu%2FLeelaMasterWeight","lists_url":"https://awesome.ecosyste.ms/api/v1/projects/github.com%2Fpangafu%2FLeelaMasterWeight/lists"}