diff --git a/.gitattributes b/.gitattributes
index 58359f40a7abbd0dd7f4cfc52d55e3091232a787..3ab5a0e980292e711e4f31795c1b62a8c50f5ff6 100644
--- a/.gitattributes
+++ b/.gitattributes
@@ -907,3 +907,52 @@ benchmark/NYU_CTF_Bench/test/2022/CSAW-Quals/rev/AnyaGacha/src/client/Library/Sc
 benchmark/NYU_CTF_Bench/test/2022/CSAW-Quals/rev/AnyaGacha/src/client/Library/ScriptAssemblies/UnityEngine.TestRunner.dll filter=lfs diff=lfs merge=lfs -text
 benchmark/NYU_CTF_Bench/test/2022/CSAW-Quals/rev/AnyaGacha/src/client/Library/ScriptAssemblies/UnityEngine.UI.dll filter=lfs diff=lfs merge=lfs -text
 benchmark/NYU_CTF_Bench/test/2022/CSAW-Quals/rev/AnyaGacha/src/client/Library/ScriptAssemblies/UnityEngine.UI.pdb filter=lfs diff=lfs merge=lfs -text
+benchmark/PUMaC_Power/2019Power-SOLS.pdf filter=lfs diff=lfs merge=lfs -text
+benchmark/PUMaC_Power/2019Power.pdf filter=lfs diff=lfs merge=lfs -text
+benchmark/PUMaC_Power/Power-Round.pdf filter=lfs diff=lfs merge=lfs -text
+benchmark/PUMaC_Power/Power-Solutions.pdf filter=lfs diff=lfs merge=lfs -text
+benchmark/PUMaC_Power/Power2024_Updated_1117.pdf filter=lfs diff=lfs merge=lfs -text
+benchmark/PUMaC_Power/Power_2025.pdf filter=lfs diff=lfs merge=lfs -text
+benchmark/PUMaC_Power/PUMaC2010_Power.pdf filter=lfs diff=lfs merge=lfs -text
+benchmark/PUMaC_Power/PUMaC2010_PowerSol.pdf filter=lfs diff=lfs merge=lfs -text
+benchmark/PUMaC_Power/PUMaC2011_Power.pdf filter=lfs diff=lfs merge=lfs -text
+benchmark/PUMaC_Power/PUMaC2011_PowerSol.pdf filter=lfs diff=lfs merge=lfs -text
+benchmark/PUMaC_Power/PUMaC2012_Power.pdf filter=lfs diff=lfs merge=lfs -text
+benchmark/PUMaC_Power/PUMaC2012_PowerSol.pdf filter=lfs diff=lfs merge=lfs -text
+benchmark/PUMaC_Power/PUMaC2013_Power.pdf filter=lfs diff=lfs merge=lfs -text
+benchmark/PUMaC_Power/PUMaC2013_PowerSol.pdf filter=lfs diff=lfs merge=lfs -text
+benchmark/PUMaC_Power/PUMaC2014_Power.pdf filter=lfs diff=lfs merge=lfs -text
+benchmark/PUMaC_Power/PUMaC2015_Power.pdf filter=lfs diff=lfs merge=lfs -text
+benchmark/PUMaC_Power/PUMaC2015_PowerwithSol.pdf filter=lfs diff=lfs merge=lfs -text
+benchmark/PUMaC_Power/PUMaC2016_Power.pdf filter=lfs diff=lfs merge=lfs -text
+benchmark/PUMaC_Power/PUMaC2017_Power.pdf filter=lfs diff=lfs merge=lfs -text
+benchmark/PUMaC_Power/PUMaC2017_Power_Solutions.pdf filter=lfs diff=lfs merge=lfs -text
+benchmark/PUMaC_Power/pumac2021powersols.pdf filter=lfs diff=lfs merge=lfs -text
+benchmark/PUMaC_Power/pumac2021v20.pdf filter=lfs diff=lfs merge=lfs -text
+benchmark/PUMaC_Power/PUMaC2023PowerRoundMarch28Update.pdf filter=lfs diff=lfs merge=lfs -text
+benchmark/PUMaC_Power/PUMACPowerRound-6.pdf filter=lfs diff=lfs merge=lfs -text
+benchmark/PUMaC_Power/PUMaC_2020__Power_Round___Problem_Statements-f5gg.pdf filter=lfs diff=lfs merge=lfs -text
+benchmark/PUMaC_Power/PUMaC_2020__Power_Round___Solutions.pdf filter=lfs diff=lfs merge=lfs -text
+benchmark/PUMaC_Power/PUMaC_2023_Power_Round_Sol.pdf filter=lfs diff=lfs merge=lfs -text
+benchmark/qanta/qanta.train.2021.12.20.json filter=lfs diff=lfs merge=lfs -text
+benchmark/vis_moot/26th-2019-problem-incl-PO2.pdf filter=lfs diff=lfs merge=lfs -text
+benchmark/vis_moot/28th-2021-problem-incl-appendix.pdf filter=lfs diff=lfs merge=lfs -text
+benchmark/vis_moot/29th-2022-problem-incl-PO2.pdf filter=lfs diff=lfs merge=lfs -text
+benchmark/vis_moot/30th-2023-problem.pdf filter=lfs diff=lfs merge=lfs -text
+benchmark/vis_moot/31st-2024-problem-incl-PO2.pdf filter=lfs diff=lfs merge=lfs -text
+benchmark/vis_moot/32nd-2025-problem-incl-PO2.pdf filter=lfs diff=lfs merge=lfs -text
+benchmark/vis_moot/33rd-2026-problem-incl-PO2.pdf filter=lfs diff=lfs merge=lfs -text
+benchmark/vis_moot/33rd-2026-rules.pdf filter=lfs diff=lfs merge=lfs -text
+benchmark/WSC_Writing/from_repo_raw/rubric/wsc_writing_rubric.pdf filter=lfs diff=lfs merge=lfs -text
+benchmark/science_bowl/MS-Sample-Questions/Sample-Set-10/10A_MS_Reg_2016.pdf filter=lfs diff=lfs merge=lfs -text
+benchmark/science_bowl/MS-Sample-Questions/Sample-Set-10/11A_MS_Reg_2016.pdf filter=lfs diff=lfs merge=lfs -text
+benchmark/science_bowl/MS-Sample-Questions/Sample-Set-10/12A_MS_Reg_2016.pdf filter=lfs diff=lfs merge=lfs -text
+benchmark/science_bowl/MS-Sample-Questions/Sample-Set-10/14A_MS_Reg_2016.pdf filter=lfs diff=lfs merge=lfs -text
+benchmark/science_bowl/MS-Sample-Questions/Sample-Set-10/15A_MS_Reg_2016.pdf filter=lfs diff=lfs merge=lfs -text
+benchmark/science_bowl/MS-Sample-Questions/Sample-Set-10/16A_MS_Reg_2016.pdf filter=lfs diff=lfs merge=lfs -text
+benchmark/science_bowl/MS-Sample-Questions/Sample-Set-10/17A_MS_Reg_2016.pdf filter=lfs diff=lfs merge=lfs -text
+benchmark/science_bowl/MS-Sample-Questions/Sample-Set-10/1A_MS_Reg_2016.pdf filter=lfs diff=lfs merge=lfs -text
+benchmark/science_bowl/MS-Sample-Questions/Sample-Set-10/2A_MS_Reg_2016.pdf filter=lfs diff=lfs merge=lfs -text
+benchmark/science_bowl/MS-Sample-Questions/Sample-Set-10/3A_MS_Reg_2016.pdf filter=lfs diff=lfs merge=lfs -text
+benchmark/science_bowl/MS-Sample-Questions/Sample-Set-10/4A_MS_Reg_2016.pdf filter=lfs diff=lfs merge=lfs -text
+benchmark/science_bowl/MS-Sample-Questions/Sample-Set-10/6A_MS_Reg_2016.pdf filter=lfs diff=lfs merge=lfs -text
diff --git a/benchmark/PUMaC_Power/2019Power-SOLS.pdf b/benchmark/PUMaC_Power/2019Power-SOLS.pdf
new file mode 100644
index 0000000000000000000000000000000000000000..749e6d29ffee98a37aca600ecb377987693b5888
--- /dev/null
+++ b/benchmark/PUMaC_Power/2019Power-SOLS.pdf
@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:69f6d5724317be5d568314b187a464b4272cf0bf8b1a7bd83c2c287ffcf76062
+size 515553
diff --git a/benchmark/PUMaC_Power/2019Power.pdf b/benchmark/PUMaC_Power/2019Power.pdf
new file mode 100644
index 0000000000000000000000000000000000000000..6efa4176a2f9c67ecaf9f551b3ddad9520e4517c
--- /dev/null
+++ b/benchmark/PUMaC_Power/2019Power.pdf
@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:571282d53629fb7f07f33906538f73d336697765765d41e55098965a978f70c0
+size 620827
diff --git a/benchmark/PUMaC_Power/PUMACPowerRound-6.pdf b/benchmark/PUMaC_Power/PUMACPowerRound-6.pdf
new file mode 100644
index 0000000000000000000000000000000000000000..4272a14c2da76b0398cbda038c2f9f47f9b88f09
--- /dev/null
+++ b/benchmark/PUMaC_Power/PUMACPowerRound-6.pdf
@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:0aef09acf2013fb5a01c77494dfb8083681d0ed2dd9da1e7e455b83bc07edaed
+size 819673
diff --git a/benchmark/PUMaC_Power/PUMaC2010_Power.pdf b/benchmark/PUMaC_Power/PUMaC2010_Power.pdf
new file mode 100644
index 0000000000000000000000000000000000000000..e790e9690a81cdf2b94dfd4d92830cafc2254d54
--- /dev/null
+++ b/benchmark/PUMaC_Power/PUMaC2010_Power.pdf
@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:320c312e03f6c47388781c345c0cfe521c0af13dfb8c87a4fa4df9dbdbf70e28
+size 154618
diff --git a/benchmark/PUMaC_Power/PUMaC2010_PowerSol.pdf b/benchmark/PUMaC_Power/PUMaC2010_PowerSol.pdf
new file mode 100644
index 0000000000000000000000000000000000000000..df00c7729a432dbde795ec86780b07892993eabd
--- /dev/null
+++ b/benchmark/PUMaC_Power/PUMaC2010_PowerSol.pdf
@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:97396533af61c4752199a0bbf45343711073ef94437e57465b4d3ff67bb2c8bb
+size 171569
diff --git a/benchmark/PUMaC_Power/PUMaC2011_Power.pdf b/benchmark/PUMaC_Power/PUMaC2011_Power.pdf
new file mode 100644
index 0000000000000000000000000000000000000000..68d5dc733f031367c5a8454c0a6c84b31a801af7
--- /dev/null
+++ b/benchmark/PUMaC_Power/PUMaC2011_Power.pdf
@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:2523ceaa00fe58030aaf583d8deab6c8a4ef78156965d3e39c12486f9e1ccda7
+size 530939
diff --git a/benchmark/PUMaC_Power/PUMaC2011_PowerSol.pdf b/benchmark/PUMaC_Power/PUMaC2011_PowerSol.pdf
new file mode 100644
index 0000000000000000000000000000000000000000..0ca1760e602b7c8beb1ea541ff02a11da251ce14
--- /dev/null
+++ b/benchmark/PUMaC_Power/PUMaC2011_PowerSol.pdf
@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:0dc0c4d6af245de2b34b8d6f8f84686d2e33f0ca274dfeb91751d3001c78a004
+size 1454780
diff --git a/benchmark/PUMaC_Power/PUMaC2012_Power.pdf b/benchmark/PUMaC_Power/PUMaC2012_Power.pdf
new file mode 100644
index 0000000000000000000000000000000000000000..5336dee94e01fd151eb85a316b50bd00a0b0da70
--- /dev/null
+++ b/benchmark/PUMaC_Power/PUMaC2012_Power.pdf
@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:763b4d5e64febb7394dc9def50023f1ff834b7bd2466ad3502c1aa2df1ca013b
+size 975759
diff --git a/benchmark/PUMaC_Power/PUMaC2012_PowerSol.pdf b/benchmark/PUMaC_Power/PUMaC2012_PowerSol.pdf
new file mode 100644
index 0000000000000000000000000000000000000000..48b8f89babdff2ef2e3a04dfe63eec8975ced31e
--- /dev/null
+++ b/benchmark/PUMaC_Power/PUMaC2012_PowerSol.pdf
@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:960b976463bde59bb26e99aeb2f8dadc5315e51cf2d11907c0aba3cd7a429f09
+size 1021817
diff --git a/benchmark/PUMaC_Power/PUMaC2013_Power.pdf b/benchmark/PUMaC_Power/PUMaC2013_Power.pdf
new file mode 100644
index 0000000000000000000000000000000000000000..83fe4a24eb3d610a52f82c3f0c3343d4bc39573e
--- /dev/null
+++ b/benchmark/PUMaC_Power/PUMaC2013_Power.pdf
@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:43a9805048226995c47348db5040447b767d10ab1d3aba94c1eefc29c854a94b
+size 444996
diff --git a/benchmark/PUMaC_Power/PUMaC2013_PowerSol.pdf b/benchmark/PUMaC_Power/PUMaC2013_PowerSol.pdf
new file mode 100644
index 0000000000000000000000000000000000000000..6ccadd620d337d90e4d72bb248b00d240d6341f6
--- /dev/null
+++ b/benchmark/PUMaC_Power/PUMaC2013_PowerSol.pdf
@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:b8f69b7d5c63b5daf392685a8b2a42ffb451bf03ff95d7fbbf43c15baf47c162
+size 649008
diff --git a/benchmark/PUMaC_Power/PUMaC2014_Power.pdf b/benchmark/PUMaC_Power/PUMaC2014_Power.pdf
new file mode 100644
index 0000000000000000000000000000000000000000..7c16b5029a72b94b4d0f51294eb29abf4656dee5
--- /dev/null
+++ b/benchmark/PUMaC_Power/PUMaC2014_Power.pdf
@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:184844fb8497bda0cbbcc00c692d3f63e1395e9ecd3336f0b754abfb599c5df6
+size 309111
diff --git a/benchmark/PUMaC_Power/PUMaC2015_Power.pdf b/benchmark/PUMaC_Power/PUMaC2015_Power.pdf
new file mode 100644
index 0000000000000000000000000000000000000000..78bb76fddc6622eaa1d11d584adb5491a1f504d1
--- /dev/null
+++ b/benchmark/PUMaC_Power/PUMaC2015_Power.pdf
@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:6b2f3190b9a90964ffd37c7a60e126a56b1dca263123b884b038824a5b3e5b33
+size 1290555
diff --git a/benchmark/PUMaC_Power/PUMaC2015_PowerwithSol.pdf b/benchmark/PUMaC_Power/PUMaC2015_PowerwithSol.pdf
new file mode 100644
index 0000000000000000000000000000000000000000..0ee3abde95bedf9808e92a36cc0f8aa753611595
--- /dev/null
+++ b/benchmark/PUMaC_Power/PUMaC2015_PowerwithSol.pdf
@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:bad6f689d68f00dc7c92cd371c96dec9f908ba9a3637d6bf3cbdfb4b69fba411
+size 1405666
diff --git a/benchmark/PUMaC_Power/PUMaC2016_Power.pdf b/benchmark/PUMaC_Power/PUMaC2016_Power.pdf
new file mode 100644
index 0000000000000000000000000000000000000000..c83d366a79029e8e78c897d2ac1aa05bd580080a
--- /dev/null
+++ b/benchmark/PUMaC_Power/PUMaC2016_Power.pdf
@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:cd03866da1e034bcc932005158288c0b1516ffd4703ec5abf50467c34becb0e9
+size 529251
diff --git a/benchmark/PUMaC_Power/PUMaC2017_Power.pdf b/benchmark/PUMaC_Power/PUMaC2017_Power.pdf
new file mode 100644
index 0000000000000000000000000000000000000000..3c80c75f9080a59961b8799a05a02284d76754ce
--- /dev/null
+++ b/benchmark/PUMaC_Power/PUMaC2017_Power.pdf
@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:0b4b72f1a0afd03ce5ca3484a7884d54b51747492097d6e5c1f3f81b3b862f36
+size 438940
diff --git a/benchmark/PUMaC_Power/PUMaC2017_Power_Solutions.pdf b/benchmark/PUMaC_Power/PUMaC2017_Power_Solutions.pdf
new file mode 100644
index 0000000000000000000000000000000000000000..a731e890a6265b3b3e426a43c7256d9971d61337
--- /dev/null
+++ b/benchmark/PUMaC_Power/PUMaC2017_Power_Solutions.pdf
@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:6e9724855076ff07a58ae517c6bed6f8f102b0201564a7369ce433aead2ace16
+size 317844
diff --git a/benchmark/PUMaC_Power/PUMaC2023PowerRoundMarch28Update.pdf b/benchmark/PUMaC_Power/PUMaC2023PowerRoundMarch28Update.pdf
new file mode 100644
index 0000000000000000000000000000000000000000..d1e7cd7346aa8c2250bc81ab9331c1c325393615
--- /dev/null
+++ b/benchmark/PUMaC_Power/PUMaC2023PowerRoundMarch28Update.pdf
@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:2ad13388d3a1122a5f5b03429dd877d229ece2d0e11c011e765a5753495f2dec
+size 453365
diff --git a/benchmark/PUMaC_Power/PUMaC_2020__Power_Round___Problem_Statements-f5gg.pdf b/benchmark/PUMaC_Power/PUMaC_2020__Power_Round___Problem_Statements-f5gg.pdf
new file mode 100644
index 0000000000000000000000000000000000000000..0eea3abdec8be37cfa3fd2ae137e2aa93f971a40
--- /dev/null
+++ b/benchmark/PUMaC_Power/PUMaC_2020__Power_Round___Problem_Statements-f5gg.pdf
@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:b276e87b178c9e3ed99fbbab48d9b039a6190eb4e5a4a2945094d6a69f4d48af
+size 354490
diff --git a/benchmark/PUMaC_Power/PUMaC_2020__Power_Round___Solutions.pdf b/benchmark/PUMaC_Power/PUMaC_2020__Power_Round___Solutions.pdf
new file mode 100644
index 0000000000000000000000000000000000000000..5b22f1fc6d85075fa75bc4f81b9c72006900e31e
--- /dev/null
+++ b/benchmark/PUMaC_Power/PUMaC_2020__Power_Round___Solutions.pdf
@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:b2812f06f15559c4c62c40cd05da28947c5edf1bba68b91362feee49a123c7b3
+size 417016
diff --git a/benchmark/PUMaC_Power/PUMaC_2023_Power_Round_Sol.pdf b/benchmark/PUMaC_Power/PUMaC_2023_Power_Round_Sol.pdf
new file mode 100644
index 0000000000000000000000000000000000000000..5e5e25c15c1bb3430d10f4dcb65c6d53404c873c
--- /dev/null
+++ b/benchmark/PUMaC_Power/PUMaC_2023_Power_Round_Sol.pdf
@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:47b90c584661cccb3ef8111c6da68c93cbc63a4e4b43b412234e1c248f685b4a
+size 371664
diff --git a/benchmark/PUMaC_Power/Power-Round.pdf b/benchmark/PUMaC_Power/Power-Round.pdf
new file mode 100644
index 0000000000000000000000000000000000000000..c6d8600de4fdf43f40f61df0005d1cf1edae2b0e
--- /dev/null
+++ b/benchmark/PUMaC_Power/Power-Round.pdf
@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:b798917b5986d901f01fa65ca2299853b7e66e842d0135c68406832113c9e5b1
+size 364374
diff --git a/benchmark/PUMaC_Power/Power-Solutions.pdf b/benchmark/PUMaC_Power/Power-Solutions.pdf
new file mode 100644
index 0000000000000000000000000000000000000000..ce61908fc48718a7a63c235436f5dc5837c9d408
--- /dev/null
+++ b/benchmark/PUMaC_Power/Power-Solutions.pdf
@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:9a80893f8287f4945dde98a6197301180b7fbd74bb394de694020c69333c463d
+size 692908
diff --git a/benchmark/PUMaC_Power/Power2024_Updated_1117.pdf b/benchmark/PUMaC_Power/Power2024_Updated_1117.pdf
new file mode 100644
index 0000000000000000000000000000000000000000..fb62959b48d32874ac1206ea8765d11753892f44
--- /dev/null
+++ b/benchmark/PUMaC_Power/Power2024_Updated_1117.pdf
@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:c0b9a15b54ba0e2f1f295c0083c47587d2022781b53fcae048eea9d14fc42605
+size 1124490
diff --git a/benchmark/PUMaC_Power/Power_2025.pdf b/benchmark/PUMaC_Power/Power_2025.pdf
new file mode 100644
index 0000000000000000000000000000000000000000..0db647f04a6b0e7df60e74896ab66e49fb27be51
--- /dev/null
+++ b/benchmark/PUMaC_Power/Power_2025.pdf
@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:c5ef7e35fc3c15bd2552503a71a3fcca3828c6b96029b341bce5e1992bc0f086
+size 825042
diff --git a/benchmark/PUMaC_Power/pumac2021powersols.pdf b/benchmark/PUMaC_Power/pumac2021powersols.pdf
new file mode 100644
index 0000000000000000000000000000000000000000..d7a1b08208a0ae25f6bed187b72f9d55a106ab41
--- /dev/null
+++ b/benchmark/PUMaC_Power/pumac2021powersols.pdf
@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:0dd01315a987ae7e270b37a17b664d282b01c4f53a757c41ed3e469b5e0c355e
+size 288222
diff --git a/benchmark/PUMaC_Power/pumac2021v20.pdf b/benchmark/PUMaC_Power/pumac2021v20.pdf
new file mode 100644
index 0000000000000000000000000000000000000000..f61d7b1c22fbe83d731b0209fcd06ec4cd066825
--- /dev/null
+++ b/benchmark/PUMaC_Power/pumac2021v20.pdf
@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:7a0b658973cecb204e577eff998c477fd99148813467f16e0c22d5a634ba66ae
+size 869711
diff --git a/benchmark/README.md b/benchmark/README.md
index 100d8489ce07f47bf983e7bbdb0fb66f96c42538..76d10468356972ce0ce40928636043657bb355e1 100644
--- a/benchmark/README.md
+++ b/benchmark/README.md
@@ -1,700 +1,261 @@
-# Agent Olympiad — competition benchmark suite
+# Agent Olympiad — team competition benchmark suite
 
 Last updated: 2026-07-22
 
-A curated collection of **real team-competition datasets** for benchmarking multi-agent AI teams. All datasets in the external suite live flat in this folder and are organized below by **domain**.
+**Multi-agent / team contests only.** Solo Origin=1 packs (AIME, OlympicArena, HealthBench, HLE, …) were removed from this folder.
 
-The canonical project-wide tracker is
-[`agent-olympiad/docs/DATA_COLLECTION.md`](../agent-olympiad/docs/DATA_COLLECTION.md).
-It preserves the original 20 source-collected team competitions and aligns them
-with the **46** local datasets documented here. Counts are kept separate because
-the two inventories can overlap conceptually. Deep-dive simulator notes also live in
-[`from_zhongzheng/olympiad/README.md`](../agent-olympiad/from_zhongzheng/olympiad/README.md)
-and [`from_zhongzheng/art/README.md`](../agent-olympiad/from_zhongzheng/art/README.md).
+Canonical trackers: [`from_zhongzheng/README.md`](../agent-olympiad/from_zhongzheng/README.md) · [`olympiad/README.md`](../agent-olympiad/from_zhongzheng/olympiad/README.md) · [`art/README.md`](../agent-olympiad/from_zhongzheng/art/README.md).
 
-**Grading type**
-- **RULE** — deterministic auto-grading: exact/normalized match, hidden test cases, `eval.py`, or flag match. No LLM judge.
-- **RUBRIC** — free-form deliverable scored against an explicit rubric (official human scores where available, otherwise a validated LLM-judge panel).
-
-> **Data hosting.** The three original benchmarks (`OlympicArena`, `modeling_agent`, `MLAgentBench`) are mirrored on the private HF dataset [`akaaafk/multiagent_bench`](https://huggingface.co/datasets/akaaafk/multiagent_bench). Everything else is pulled from its own public HF / GitHub source — see [Downloading](#downloading). Large or Docker-based datasets are fetched on demand.
+**Grading**
+- **RULE** — exact/normalized match, hidden tests, flags
+- **RUBRIC** — open deliverable scored by rubric / jury / calibrated LLM judges
 
 ---
 
 ## Summary
 
-| Datasets | Domains | Rule-based | Rubric |
-|:--------:|:-------:|:----------:|:------:|
-| **46** | **13** | **21** | **25** |
-
-| Domain | Datasets | Rule | Rubric |
-|--------|:--------:|:----:|:------:|
-| [STEM olympiad](#1-stem-olympiad) | 5 | 4 | 1 |
-| [Competitive programming](#2-competitive-programming) | 3 | 3 | 0 |
-| [Software & ML engineering](#3-software--ml-engineering) | 3 | 3 | 0 |
-| [AI/ML research](#4-aiml-research) | 2 | 0 | 2 |
-| [Applied math modeling](#5-applied-math-modeling) | 1 | 0 | 1 |
-| [Cybersecurity](#6-cybersecurity) | 2 | 2 | 0 |
-| [Medicine](#7-medicine) | 2 | 0 | 2 |
-| [Business & finance](#8-business--finance) | 3 | 0 | 3 |
-| [Law](#9-law) | 1 | 0 | 1 |
-| [Ethics, debate & general knowledge](#10-ethics-debate--general-knowledge) | 4 | 1 | 3 |
-| [Frontier exams, reasoning & agents](#11-frontier-exams-reasoning--agents) | 4 | 4 | 0 |
-| [Writing & professional deliverables](#12-writing--professional-deliverables) | 4 | 0 | 4 |
-| [International olympiads & team contests](#13-international-olympiads--team-contests) | 12 | 4 | 8 |
-| **Total** | **46** | **21** | **25** |
-
-> `wharton_investment`, `iGeo`, `IOAI`, `History_Olympiad`, and `IEO` are hybrids (RULE + RUBRIC); in the domain table they are counted under **Rubric**. Local top-level folders = **46**, matching this README.
-
-Access legend below each table: ✅ downloaded locally · ⏳ fetch on demand (large runtime assets such as Kaggle CSVs / Docker images).
-
----
-
-## 1. STEM olympiad
-
-Closed-book science/math contests with closed-form answers → fully auto-graded. `iGeo` also includes rubric-scored fieldwork.
-
-### `OlympicArena` — OlympicArena (GAIR, NeurIPS 2024) · RULE ✅
-
-| | |
-|---|---|
-| **Domain** | STEM olympiads — Math, Physics, Chemistry, Biology, Geography, Astronomy, CS |
-| **Questions** | 10,615 (test 9,977 answers withheld · val 638 public) |
-| **Team setup** | Individual in origin; adaptable to per-discipline specialist agents |
-| **Answer type** | Single/multi value, expression, interval, code (CS ships test cases) |
-| **Grading** | Answer-type-aware rule match; `val` local, `test` via official platform |
-| **Data** | `OlympicArena/` — 7 discipline folders × val/test parquet (~254 MB) |
-| **Link** | [HF GAIR/OlympicArena](https://huggingface.co/datasets/GAIR/OlympicArena) · [GitHub](https://github.com/GAIR-NLP/OlympicArena) · [arXiv:2406.12753](https://arxiv.org/abs/2406.12753) |
-
-### `OlympiadBench` — Olympiad math + physics (ACL 2024) · RULE ✅
-
-| | |
-|---|---|
-| **Domain** | Olympiad mathematics + physics (bilingual EN/ZH, text + multimodal) |
-| **Questions** | 8,476 |
-| **Answer type** | Final numeric / expression answer |
-| **Grading** | Final-answer match against gold (use upstream answer-checker) |
-| **Data** | `OlympiadBench/` (~104 MB) |
-| **Link** | [HF Hothan/OlympiadBench](https://huggingface.co/datasets/Hothan/OlympiadBench) · [arXiv:2402.14008](https://arxiv.org/abs/2402.14008) |
-
-### `AIME` — American Invitational Mathematics Examination · RULE ✅
-
-| | |
-|---|---|
-| **Domain** | Competition mathematics |
-| **Questions** | 90 (AIME 2024 + 2025 + 2026, 30 each) |
-| **Answer type** | Integer 0–999, no partial credit |
-| **Grading** | Deterministic integer match via `math-verify` |
-| **Data** | `AIME/aime24/`, `AIME/aime25/`, `AIME/aime26/` |
-| **Link** | [HF math-ai/aime24](https://huggingface.co/datasets/math-ai/aime24) · [aime25](https://huggingface.co/datasets/math-ai/aime25) · [aime26](https://huggingface.co/datasets/math-ai/aime26) |
-
-### `science_bowl` — DOE National Science Bowl · RULE ✅
-
-| | |
-|---|---|
-| **Domain** | Science buzzer quiz — physics, chemistry, biology, earth science, energy, math |
-| **Questions** | **~23,691** across 508 PDFs (11,846 Toss-Up + 11,845 Bonus) |
-| **Team setup** | 4 agents + 1 alternate |
-| **Answer type** | Short spoken answers; Toss-Up = no conferring, Bonus = team conferring |
-| **Grading** | Normalized exact match; built-in individual-vs-team comparison by round type |
-| **Data** | `science_bowl/` — `HS-Sample-Questions/` (17 sets) + `MS-Sample-Questions/` (16 sets) |
-| **Link** | [science.osti.gov NSB](https://science.osti.gov/wdts/nsb/Regional-Competitions/Resources) |
-
-### `iGeo` — International Geography Olympiad · RULE + RUBRIC ✅
+| Datasets | Rule-based | Rubric / hybrid |
+|:--------:|:----------:|:---------------:|
+| **24** | **9** | **15** |
 
-| | |
-|---|---|
-| **Domain** | Geography — human, physical, environmental, fieldwork |
-| **Questions** | **600** (WRT **505** · FWE **55** · MMT **40**) |
-| **Team setup** | Individual WRT/MMT; recommend 3 roles (map / data / writer) for fieldwork sims |
-| **Answer type** | Short written responses, map/data interpretation, field reports; MMT = MCQ |
-| **Grading** | RULE via marking schemes for keyed items; RUBRIC for open FWE writeups |
-| **Data** | `iGeo/igeo_questions.json` (+ source PDFs) |
-| **Link** | [geoolympiad.org document library](https://geoolympiad.org/document-library/) |
+| Domain | Datasets |
+|--------|:--------:|
+| [Programming & CTF](#1-programming--ctf) | 3 |
+| [AI olympiad](#2-ai-olympiad) | 1 |
+| [Math modeling & proofs](#3-math-modeling--proofs) | 2 |
+| [Lab / chemistry / astronomy / linguistics](#4-lab--chemistry--astronomy--linguistics) | 4 |
+| [Business, law, ethics, debate](#5-business-law-ethics-debate) | 8 |
+| [History, writing, quiz, puzzle, ARC](#6-history-writing-quiz-puzzle-arc) | 6 |
+| **Total** | **24** |
 
 ---
 
-## 2. Competitive programming
-
-Code judged by hidden test cases; ICPC-style team rules fit a shared-machine agent team.
-
-### `USACO` — USA Computing Olympiad · RULE ✅
-
-| | |
-|---|---|
-| **Questions** | 520 (bronze 148 · silver 144 · gold 142 · platinum 86) |
-| **Answer type** | Program; graded on hidden test cases (pass/fail per subtask) |
-| **Grading** | Run submission against hidden cases; score = fraction passed |
-| **Data** | `USACO/data/` — 12 parquet shards (~3.8 GB) |
-| **Link** | [HF codegenning/usacobench_formatted](https://huggingface.co/datasets/codegenning/usacobench_formatted) |
+## 1. Programming & CTF
 
-### `LiveCodeBench` — LeetCode / AtCoder / Codeforces · RULE ✅
-
-| | |
-|---|---|
-| **Questions** | ~1,055, timestamped for contamination-free windows |
-| **Grading** | Hidden test cases; evaluate on the window after each model's cutoff |
-| **Data** | `LiveCodeBench/` (~4.5 GB) |
-| **Link** | [HF livecodebench/code_generation_lite](https://huggingface.co/datasets/livecodebench/code_generation_lite) · [arXiv:2403.07974](https://arxiv.org/abs/2403.07974) |
-
-### `CodeContests` — Codeforces / ICPC-style (AlphaCode data) · RULE ✅
+### `CodeContests` — ICPC-style (AlphaCode) · RULE ✅
 
 | | |
 |---|---|
 | **Questions** | **13,610** (train 13,328 · valid 117 · test 165) |
-| **Team setup** | ICPC analogue: 3 agents, one shared submission queue, penalty per wrong submit |
-| **Grading** | Hidden test cases |
+| **Team** | Origin **3** · shared machine + penalty |
 | **Data** | `CodeContests/` (~7.6 GB) |
 | **Link** | [HF deepmind/code_contests](https://huggingface.co/datasets/deepmind/code_contests) |
 
----
-
-## 3. Software & ML engineering
-
-Interactive environments; deliverable is a code artifact scored automatically.
-
-### `SWE-bench_Verified` — real GitHub issue fixes · RULE ✅
+### `NYU_CTF_Bench` — CSAW CTF · RULE ✅
 
 | | |
 |---|---|
-| **Questions** | 500 (human-validated subset of SWE-bench) |
-| **Answer type** | Patch against a real repository |
-| **Grading** | Apply patch → run `FAIL_TO_PASS` + `PASS_TO_PASS` tests (per-repo Docker) |
-| **Team setup** | Locator / patcher / reviewer on a shared checkout |
-| **Data** | `SWE-bench_Verified/` (~2 MB metadata; repos cloned at eval time) |
-| **Link** | [HF princeton-nlp/SWE-bench_Verified](https://huggingface.co/datasets/princeton-nlp/SWE-bench_Verified) |
+| **Questions** | 200 across 6 categories |
+| **Team** | Origin **4–6** |
+| **Data** | `NYU_CTF_Bench/` |
+| **Link** | [NYU-LLM-CTF/NYU_CTF_Bench](https://github.com/NYU-LLM-CTF/NYU_CTF_Bench) |
 
-### `MLAgentBench` — ML experimentation tasks (Stanford SNAP, ICML 2024) · RULE ✅
+### `Cybench` — professional CTF · RULE ✅
 
 | | |
 |---|---|
-| **Questions** | 15 tasks (Kaggle challenges, classic datasets, research + speedup tasks) |
-| **Answer type** | Improved code + `submission.csv` |
-| **Grading** | Hidden `eval.py`; >10% improvement over baseline = success |
-| **Team setup** | Engineer / researcher / reviewer sharing one workspace |
-| **Data** | `MLAgentBench/benchmarks/` — per-task `env/` + hidden `scripts/` (task defs ~0.5 MB) |
-| **Link** | [GitHub](https://github.com/snap-stanford/MLAgentBench) · [arXiv:2310.03302](https://arxiv.org/abs/2310.03302) |
-
-### `MLE-bench` — 75 real Kaggle competitions (OpenAI) · RULE ✅
-
-| | |
-|---|---|
-| **Questions** | 75 official competitions (Lite = Low complexity = 22) across tabular, CV, NLP, signal; local clone has 82 competition folders (7 extras beyond `split75`) |
-| **Answer type** | `submission.csv` per competition, built by training real models on real data |
-| **Grading** | `mlebench grade` scores against the competition's **private leaderboard**; medal thresholds (bronze/silver/gold) as if competing at the time |
-| **Team setup** | Kaggle-team analogue: EDA / feature engineer / modeler / experiment runner sharing one workspace |
-| **Data** | `MLE-bench/` — harness/repo local; competition CSVs still need Kaggle API + `mlebench prepare` (100s of GB) |
-| **Link** | [repo openai/mle-bench](https://github.com/openai/mle-bench) · [arXiv:2410.07095](https://arxiv.org/abs/2410.07095) |
+| **Questions** | **40** challenge packs |
+| **Team** | Origin **4–6** |
+| **Data** | `Cybench/benchmark/` |
+| **Link** | [andyzorigin/cybench](https://github.com/andyzorigin/cybench) |
 
 ---
 
-## 4. AI/ML research
-
-Open-ended research deliverables graded by structured review rubrics.
-
-### `MLR-Bench` — ML research tasks (NeurIPS/ICLR/ICML workshops) · RUBRIC ✅
+## 2. AI olympiad
 
-| | |
-|---|---|
-| **Questions** | 201 tasks across 9 ML topics |
-| **Answer type** | Idea → proposal → experiments → paper (stepwise or end-to-end) |
-| **Grading** | MLR-Judge rubric on up to 9 dims (novelty, soundness, clarity, feasibility, …), dual-judge averaged |
-| **Team setup** | Ideator / proposer / experimenter / writer; add a verifier (coding agents fabricate ~80% of results) |
-| **Data** | `MLR-Bench/` (~0.6 MB) |
-| **Link** | [HF chchenhui/mlrbench-tasks](https://huggingface.co/datasets/chchenhui/mlrbench-tasks) · [arXiv:2505.19955](https://arxiv.org/abs/2505.19955) |
-
-### `PaperBench` — replicate 20 ICML 2024 papers (OpenAI) · RUBRIC ✅
+### `IOAI` — International Olympiad in AI · RULE + RUBRIC ✅
 
 | | |
 |---|---|
-| **Questions** | 20 eval papers, 8,316 leaf criteria (local also ships 3 development papers → 23 / 8,921) |
-| **Answer type** | Full replication codebase built from the paper |
-| **Grading** | Hierarchical author-co-designed rubric tree (leaf = binary, weighted up to a Replication Score); rubric hidden from solver |
-| **Team setup** | Read-paper / implement / run-experiments in a sandbox (long-horizon like MLAgentBench) |
-| **Data** | `PaperBench/` — papers + rubrics (~13 MB; heavy LFS blobs may still be pulled on demand) |
-| **Link** | [repo openai/preparedness](https://github.com/openai/preparedness/tree/main/project/paperbench) · [arXiv:2504.01848](https://arxiv.org/abs/2504.01848) |
+| **Questions** | **17** tasks (2024–2026) |
+| **Team** | Individual + **Team Challenge** (Rec 3–5) |
+| **Data** | `IOAI/` — see local `README.md` |
+| **Link** | [IOAI-official](https://ioai-official.org/resources/) |
 
 ---
 
-## 5. Applied math modeling
+## 3. Math modeling & proofs
 
 ### `modeling_agent` — HiMCM / MCM / ICM / IM2C / MidMCM · RUBRIC ✅
 
 | | |
 |---|---|
-| **Domain** | Applied math modeling — statistics, optimization, simulation, data analysis |
-| **Questions** | 68 (2001–2025; Middle 3 · High 21 · Undergrad 44) |
-| **Team setup** | 3–4 agents (modeler / coder / writer); `decomposition` field seeds the split |
-| **Answer type** | Multi-day open-book modeling paper (assumptions, model, analysis, recommendations) |
-| **Grading** | LLM-judge panel over per-problem `requirements` grading points; `eval_roles` as judge personas |
-| **Data** | `modeling_agent/modeling_data_final.json` — 68 entries with `requirements`, `eval_roles`, `decomposition` |
-| **Link** | [COMAP contests](https://www.comap.com/contests) |
-
----
-
-## 6. Cybersecurity
-
-Capture-the-flag with exact flag matching inside dockerized challenge environments.
-
-### `NYU_CTF_Bench` — CSAW CTF (NeurIPS 2024) · RULE ✅
-
-| | |
-|---|---|
-| **Questions** | 200 challenges across 6 categories (web, pwn, forensics, rev, crypto, misc) |
-| **Team setup** | CTF team split by category |
-| **Grading** | Exact flag match inside the challenge's Docker environment; fully automatic |
-| **Data** | `NYU_CTF_Bench/` (~2.5 GB). Windows note: a few challenge folder names with `?` / trailing `...` were sanitized on extract |
-| **Link** | [repo NYU-LLM-CTF/NYU_CTF_Bench](https://github.com/NYU-LLM-CTF/NYU_CTF_Bench) · [site](https://nyu-llm-ctf.github.io/) |
-
-### `Cybench` — professional CTF (HackTheBox / Sekai etc.) · RULE ✅
-
-| | |
-|---|---|
-| **Questions** | 40 tasks with subtask decomposition |
-| **Grading** | Flag match + per-subtask checks in Docker |
-| **Data** | `Cybench/benchmark/` — 40 challenge packs (`README` + `challenge`/`dist` + `metadata/`); Docker images still pulled for interactive web/pwn |
-| **Link** | [repo andyzorigin/cybench](https://github.com/andyzorigin/cybench) |
-
----
-
-## 7. Medicine
-
-### `HealthBench` — realistic clinical conversations (OpenAI) · RUBRIC ✅
-
-| | |
-|---|---|
-| **Questions** | 5,000 conversations (subsets: consensus 3,671 · hard 1,000) |
-| **Answer type** | Free-form response to a health conversation |
-| **Grading** | 48,562 physician-written weighted criteria per conversation; model-based grader checks each met/unmet → weighted score |
-| **Team setup** | Clinician / safety-reviewer / communicator; `red_teaming` split stresses safety coordination |
-| **Data** | `HealthBench/` — JSONL (`oss_eval`, `consensus`, `hard`) (~246 MB) |
-| **Link** | [HF openai/healthbench](https://huggingface.co/datasets/openai/healthbench) · [arXiv:2505.08775](https://arxiv.org/abs/2505.08775) |
-
-### `HealthBench-Professional` — physician-authored professional tier (OpenAI) · RUBRIC ✅
-
-| | |
-|---|---|
-| **Questions** | 525 examples across `consult` / `writing` / `research` use cases (incl. `red_teaming` type) |
-| **Answer type** | Free-form clinical response; each example ships a physician-written reference response |
-| **Grading** | Per-example physician-written `rubric_items` (criterion + points), model-graded met/unmet → weighted score |
-| **Team setup** | Harder tier than HealthBench: specialist / safety-reviewer / medical-writer roles; physician response doubles as a Tier-2 gold reference |
-| **Data** | `HealthBench-Professional/` — `healthbench_professional_eval.jsonl` (~2.8 MB) |
-| **Link** | [HF openai/healthbench-professional](https://huggingface.co/datasets/openai/healthbench-professional) |
-
----
-
-## 8. Business & finance
-
-Open-ended decks / reports graded against competition rubrics; past winners as gold references.
-
-### `gcch_harvard` — Global Case Competition at Harvard · RUBRIC ✅
-
-| | |
-|---|---|
-| **Domain** | Business case — finance / M&A / strategy |
-| **Questions** | 7 cases (2018–2026) |
-| **Team setup** | 2–5 agents, unrestricted internet & software |
-| **Answer type** | Slide deck + pitch |
-| **Grading** | Pairwise LLM-judge vs 9 winning decks under official criteria |
-| **Data** | `gcch_harvard/` — 7 case PDFs + 9 winning decks + past-editions page |
-| **Link** | [thecasecompetition.org](https://www.thecasecompetition.org/past-editions) |
-
-### `cfa_research_challenge` — CFA Institute Research Challenge · RUBRIC ✅
-
-| | |
-|---|---|
-| **Domain** | Finance — equity research reports |
-| **Questions** | 19 inferred subject-company tasks (`tasks.json`) + 19 champion report/presentation pairs (2008–2026), 38 PDFs |
-| **Team setup** | 3–5 agents (valuation / industry / writing / defense) |
-| **Answer type** | ~10-page BUY/SELL/HOLD equity research report + oral defense |
-| **Grading** | Industry rubric (valuation, writing, presentation); pairwise vs champion reports |
-| **Data** | `cfa_research_challenge/` — `tasks.json` + 38 PDFs (Commonwealth Bank, Vestas, Canadian Tire, …) |
-| **Link** | [cfainstitute.org past champions](https://www.cfainstitute.org/insights/events/research-challenge/past-champions) |
-| **Notes** | Task: for a host-assigned listed company, write ~10-page equity research (BUY/SELL/HOLD) from public info and defend orally. Official site publishes winners only—no yearly problem PDFs; companies in `tasks.json` are inferred from champion reports. |
-
-### `wharton_investment` — Wharton Global HS Investment Competition · RUBRIC + RULE ✅
+| **Questions** | 68 |
+| **Team** | Origin **3–4** |
+| **Data** | `modeling_agent/modeling_data_final.json` |
+| **Link** | [COMAP](https://www.comap.com/contests) |
 
-| | |
-|---|---|
-| **Domain** | Finance — long-horizon investment strategy |
-| **Questions** | 4 client case studies (2019–22 + current) |
-| **Team setup** | 4–7 agents (analyst / PM) + online trading simulator |
-| **Answer type** | 10-week simulated portfolio + strategy report & defense |
-| **Grading** | Portfolio-replay against real historical market data (RULE) + strategy rubric panel (RUBRIC) |
-| **Data** | `wharton_investment/` — 4 case-study HTML files |
-| **Link** | [globalyouth.wharton.upenn.edu](https://globalyouth.wharton.upenn.edu/investment-competition/) |
-
----
-
-## 9. Law
-
-### `vis_moot` — Willem C. Vis International Commercial Arbitration Moot · RUBRIC ✅
+### `PUMaC_Power` — PUMaC Power Round · RUBRIC ✅
 
 | | |
 |---|---|
-| **Questions** | 7 Problems (26th–33rd editions, 2019–2026; 60–90 pp. case records) |
-| **Team setup** | 2–8 agents, unrestricted legal research |
-| **Answer type** | Claimant & Respondent memoranda + oral pleading |
-| **Grading** | Memorandum + oral rubrics; pairwise vs Pace-database winning memoranda |
-| **Data** | `vis_moot/` — 7 Problems + PO2 + 33rd-edition rules (8 PDFs) |
-| **Link** | [vismoot.org](https://www.vismoot.org/previous-moots/) |
+| **Questions** | ~19 Power packets (2007–2025) |
+| **Team** | Origin **8** |
+| **Data** | `PUMaC_Power/` |
+| **Link** | [PUMaC Archives](https://jason-shi-f9dm.squarespace.com/archives) |
 
 ---
 
-## 10. Ethics, debate & general knowledge
+## 4. Lab / chemistry / astronomy / linguistics
 
-Humanities contests; mostly adversarial team formats judged by rubric (one auto-graded quiz).
-
-### `debatebench` — WUDC / BP Debate (DebateBench) · RUBRIC ✅
+### `EOES` — EOES / EUSO · RUBRIC ✅
 
 | | |
 |---|---|
-| **Questions** | 360 scored speeches across 45 debates (`scores.xlsx`); 32 debates also have full transcripts |
-| **Team setup** | 8 agents (4 teams × 2); 15-min prep, 7-min speeches |
-| **Grading** | **Official human judge speaker scores + team rankings** (`scores.xlsx`) — the calibration anchor for all rubric judges |
-| **Data** | `debatebench/` — `cleaned_speeches/` (32 transcript files) + `scores.xlsx` |
-| **Link** | [HF utkarsh2105/DebateBench](https://huggingface.co/datasets/utkarsh2105/DebateBench) · [arXiv:2502.06279](https://arxiv.org/abs/2502.06279) |
+| **Questions** | **2 exp/year** · ~90 PDFs / ~172 MB |
+| **Team** | Origin **3** |
+| **Data** | `EOES/` — see `INVENTORY.md` |
+| **Link** | [eoes.science](https://www.eoes.science/Previous%20olympiads/previous.html) |
 
-### `ethics_bowl_appe` — APPE Intercollegiate Ethics Bowl · RUBRIC ✅
+### `IChTo` — International Chemistry Tournament · RUBRIC ✅
 
 | | |
 |---|---|
-| **Questions** | ~140 cases across 9 case sets (2021–2026) |
-| **Team setup** | 5 agents; adversarial: presentation → commentary → response → judge Q&A |
-| **Grading** | Panel of ≥3 LLM judges with official scoring guides; team-vs-team → Elo |
-| **Data** | `ethics_bowl_appe/` — 9 case sets + rules + judging/scoring guides (14 files) |
-| **Link** | [appe-ethics.org](https://www.appe-ethics.org/cases-rules-guidelines/) |
+| **Questions** | **106** · 9 sets |
+| **Team** | Origin **3+** (Reporter / Opponent / Reviewer) |
+| **Data** | `IChTo/` |
+| **Link** | [ichto.org](http://ichto.org/en/problems/) |
 
-### `ethics_bowl_nhseb` — National High School Ethics Bowl · RUBRIC ✅
+### `IOAA` — Astronomy & Astrophysics (incl. Group) · RULE ✅
 
 | | |
 |---|---|
-| **Questions** | 15 cases (2025–26 national set) |
-| **Team setup** | 3–7 agents; dialogic (non-adversarial) format |
-| **Grading** | Judge rubric, similar to APPE |
-| **Data** | `ethics_bowl_nhseb/` — `2025-26-national-case-set.pdf` |
-| **Link** | [nhseb.org/case-library](https://nhseb.org/case-library) |
+| **Questions** | ~174–176 PDFs; 16 Group/Team packets |
+| **Team** | Individual + **Group 5** |
+| **Data** | `IOAA/` |
+| **Link** | [ioaastrophysics.org](https://ioaastrophysics.org/resources/problems-from-past-ioaa) |
 
-### `qanta` — QANTA Quiz Bowl · RULE ✅
+### `IOL` — International Linguistics Olympiad · RULE ✅
 
 | | |
 |---|---|
-| **Domain** | General knowledge (humanities + sciences) |
-| **Questions** | ~100,000 pyramidal tossups with answers + metadata |
-| **Team setup** | 4 agents; per-category specialization + buzz-timing decision |
-| **Answer type** | Short answer at chosen buzz point |
-| **Grading** | Automatic: answer accuracy + buzz-position credit |
-| **Data** | `qanta/` — `qanta.train.2021.12.20.json` (169 MB) + `qanta.dev` |
-| **Link** | [qanta-org.github.io](https://qanta-org.github.io/data-and-code/) · [HF mirror](https://huggingface.co/datasets/TasnimKabir12/qanta) |
+| **Questions** | **130** source · 555 HF records |
+| **Team** | Individual + **Team Contest 4** |
+| **Data** | `IOL/` |
+| **Link** | [ioling.org](https://ioling.org/problems/by_year/) · [HF agurung/ioling](https://huggingface.co/datasets/agurung/ioling) |
 
 ---
 
-## 11. Frontier exams, reasoning & agents
-
-Closed-ended frontier benchmarks — every answer is exact-match verifiable, so all four are fully auto-graded.
-
-### `HLE` — Humanity's Last Exam (CAIS + Scale AI) · RULE ✅
-
-| | |
-|---|---|
-| **Questions** | 2,500 expert-written questions across 100+ subjects (math, sciences, humanities; multimodal) |
-| **Answer type** | Multiple-choice or short answer; each has an unambiguous, non-searchable solution |
-| **Grading** | Exact match / MC key — designed for automated grading |
-| **Team setup** | Per-subject specialist agents + an aggregator that routes questions |
-| **Data** | `HLE/` — 2,500 rows (~262 MiB); gated source access approved |
-| **Link** | [HF cais/hle](https://huggingface.co/datasets/cais/hle) · [lastexam.ai](https://lastexam.ai) · [arXiv:2501.14249](https://arxiv.org/abs/2501.14249) |
+## 5. Business, law, ethics, debate
 
-### `GPQA` — graduate-level Google-proof QA · RULE ✅
+### `gcch_harvard` · RUBRIC ✅
 
 | | |
 |---|---|
-| **Questions** | 448 (Diamond subset: 198) — biology, physics, chemistry, written & validated by PhDs |
-| **Answer type** | 4-option multiple choice; non-experts with web access score only ~34% |
-| **Grading** | MC key exact match |
-| **Team setup** | Domain-specialist debate: 3 discipline agents argue to consensus before answering |
-| **Data** | `GPQA/` — main 448 · Diamond 198 · extended 546 (~8 MiB); gated source access approved |
-| **Link** | [HF Idavidrein/gpqa](https://huggingface.co/datasets/Idavidrein/gpqa) · [arXiv:2311.12022](https://arxiv.org/abs/2311.12022) |
+| **Questions** | 7 cases |
+| **Team** | Origin **2–5** |
+| **Data** | `gcch_harvard/` |
 
-### `ARC-AGI-2` — ARC Prize abstraction puzzles · RULE ✅
+### `cfa_research_challenge` · RUBRIC ✅
 
 | | |
 |---|---|
-| **Questions** | 1,000 public training + 120 public eval tasks (few-shot grid-transformation puzzles) |
-| **Answer type** | Output grid for each test input |
-| **Grading** | Exact grid match (the official ARC Prize competition metric) |
-| **Team setup** | Real prize competition run by teams: hypothesis-generator / verifier / programmatic-solver agents |
-| **Data** | `ARC-AGI-2/` (~6 MB) |
-| **Link** | [repo arcprize/ARC-AGI-2](https://github.com/arcprize/ARC-AGI-2) · [arcprize.org](https://arcprize.org) |
+| **Questions** | **19** inferred tasks + champion reports |
+| **Team** | Origin **3–5** |
+| **Data** | `cfa_research_challenge/` — `tasks.json` + PDFs |
 
-### `GAIA` — general AI assistant benchmark (Meta AI + HF) · RULE ✅
+### `wharton_investment` · RULE + RUBRIC ✅
 
 | | |
 |---|---|
-| **Questions** | 466 real-world questions in 3 difficulty levels; many ship an attached file |
-| **Answer type** | Short single-correct answer, reached via web browsing / tool use / multimodal reasoning |
-| **Grading** | Official exact-match scorer; validation split (165) has public answers, test via leaderboard |
-| **Team setup** | Orchestrator + browser / file-reader / calculator tool agents — the canonical agent-team benchmark |
-| **Data** | `GAIA/` — validation 165 + test 301 with attachments (118 files, ~105 MiB); gated source access approved |
-| **Link** | [HF gaia-benchmark/GAIA](https://huggingface.co/datasets/gaia-benchmark/GAIA) · [arXiv:2311.12983](https://arxiv.org/abs/2311.12983) |
-
----
-
-## 12. Writing & professional deliverables
+| **Questions** | 4 cases |
+| **Team** | Origin **4–7** |
+| **Data** | `wharton_investment/` |
 
-Free-form deliverables where **every instance carries its own written rubric** — the most judge-friendly rubric tier.
-
-### `WritingBench` — 1,000 real-world writing queries · RUBRIC ✅
+### `IEO` — International Economics Olympiad · RULE + RUBRIC ✅
 
 | | |
 |---|---|
-| **Questions** | 1,000 queries across 6 domains / 100 subdomains (creative, professional, technical, …) with style/format/length requirements |
-| **Answer type** | Free-form writing, tens to thousands of words |
-| **Grading** | 5 instance-specific criteria per query (name + description + scoring rubric), 10-point scale; score via LLM judge or the released [critic model](https://huggingface.co/AQuarterMile/WritingBench-Critic-Model-Qwen-7B) |
-| **Team setup** | Writer / editor / fact-checker pipeline vs solo writer |
-| **Data** | `WritingBench/benchmark_all.jsonl` (queries + criteria) |
-| **Link** | [repo X-PLUG/WritingBench](https://github.com/X-PLUG/WritingBench) · [arXiv:2503.05244](https://arxiv.org/abs/2503.05244) |
+| **Questions** | economics + **team business case** PDFs |
+| **Team** | Case Rec **3–5** |
+| **Data** | `IEO/` |
 
-### `ProfBench` — professional report generation (NVIDIA) · RUBRIC ✅
+### `vis_moot` · RUBRIC ✅
 
 | | |
 |---|---|
-| **Questions** | 40 tasks × 4 domains (Physics PhD, Chemistry PhD, Finance MBA, Consulting MBA) |
-| **Answer type** | Comprehensive professional report answering a junior-colleague-style assignment |
-| **Grading** | 3,000+ expert-human-written criteria (15–59 per task) over reasoning / extraction / style; LLM judge with published bias controls |
-| **Team setup** | Consulting-team analogue: researcher / analyst / report-writer |
-| **Data** | `ProfBench/test.jsonl` |
-| **Link** | [HF nvidia/ProfBench](https://huggingface.co/datasets/nvidia/ProfBench) · [repo NVlabs/ProfBench](https://github.com/NVlabs/ProfBench) · [arXiv:2510.18941](https://arxiv.org/abs/2510.18941) |
+| **Questions** | 7 Problems |
+| **Team** | Origin **2–8** |
+| **Data** | `vis_moot/` |
 
-### `BiGGen-Bench` — instance-rubric generation benchmark · RUBRIC ✅
+### `ethics_bowl_appe` / `ethics_bowl_nhseb` · RUBRIC ✅
 
 | | |
 |---|---|
-| **Questions** | 765 instances over 9 capabilities / 77 tasks (reasoning, planning, refinement, tool use, safety, …) |
-| **Answer type** | Free-form generation |
-| **Grading** | Each instance ships its own 5-point scoring rubric + reference answer; scored by LLM/Prometheus judges |
-| **Team setup** | Capability-routed roles; doubles as a cheap judge-calibration set (per-instance rubrics + references) |
-| **Data** | `BiGGen-Bench/` (~1 MB parquet) |
-| **Link** | [HF prometheus-eval/BiGGen-Bench](https://huggingface.co/datasets/prometheus-eval/BiGGen-Bench) · [arXiv:2406.05761](https://arxiv.org/abs/2406.05761) |
+| **Questions** | ~140 APPE cases · 15 NHSEB cases |
+| **Team** | APPE **5** · NHSEB **3–7** |
+| **Data** | `ethics_bowl_appe/` · `ethics_bowl_nhseb/` |
 
-### `FLASK` — fine-grained alignment-skill rubric eval (KAIST, ICLR 2024) · RUBRIC ✅
+### `debatebench` · RUBRIC ✅
 
 | | |
 |---|---|
-| **Questions** | ~1,700 instances compiled from 120 datasets |
-| **Answer type** | Free-form response |
-| **Grading** | Each instance annotated with required skills (from a 12-skill taxonomy), domain, difficulty + instance-specific 5-point score rubrics |
-| **Team setup** | Skill-specialized agents; skill-level score breakdown localizes which teammate helped |
-| **Data** | `FLASK/` (~5 MB) |
-| **Link** | [repo kaistAI/FLASK](https://github.com/kaistAI/FLASK) · [arXiv:2307.10928](https://arxiv.org/abs/2307.10928) |
+| **Questions** | 360 scored speeches / 45 debates |
+| **Team** | Origin **8** (4×2) |
+| **Data** | `debatebench/` |
 
 ---
 
-## 13. International olympiads & team contests
+## 6. History, writing, quiz, puzzle, ARC
 
-Local PDF / HF packs for real multi-agent olympiads (also tracked in `from_zhongzheng/olympiad/`). Hybrids with both keyed and open parts are counted under **Rubric** in the summary table (same convention as `iGeo` / `wharton_investment`).
-
-### `IOAI` — International Olympiad in AI · RULE + RUBRIC ✅
+### `History_Olympiad` · RULE + RUBRIC ✅
 
 | | |
 |---|---|
-| **Domain** | AI olympiad — ML / CV / NLP Individual + Team Challenge |
-| **Questions** | **17** tasks (2024–2026; `IOAI2025/` HF mirror not double-counted) |
-| **Team setup** | Individual Rec 3 · Team Challenge Rec 3–5 |
-| **Grading** | Hidden-test metrics (Individual) + jury/rubric (Team Challenge) |
-| **Data** | `IOAI/` — see local `README.md`; `IOAI-2024/` · `IOAI-2025/` · `IOAI-2026/` + `IOAI2025/` |
-| **Link** | [HF IOAI2025](https://huggingface.co/datasets/IOAI-official/IOAI2025) · [ioai-official.org](https://ioai-official.org/resources/) |
+| **Questions** | ~90 PDFs (6 editions) |
+| **Team** | Bee + **Bowl team** |
+| **Data** | `History_Olympiad/` |
 
-### `EOES` — EOES / EUSO experimental science · RUBRIC ✅
+### `WSC_Writing` · RUBRIC ✅
 
 | | |
 |---|---|
-| **Domain** | 3-person integrated lab practical (physics + chemistry + biology) |
-| **Questions** | **2 experiments/year**; local **~90 PDFs / ~172 MB** (EUSO partial + EOES 2021/23–25; 2022 tasks unpublished) |
-| **Team setup** | Origin/Rec **3** (experimentalist / analyst / scribe) |
-| **Grading** | Official marking schemes; rubric-style open responses |
-| **Data** | `EOES/` — `euso/<year>/`, `eoes/<year>/`, mirrors; see `INVENTORY.md` |
-| **Link** | [EUSO](http://euso.eu/about/experiments/) · [EOES](https://www.eoes.science/Previous%20olympiads/previous.html) |
+| **Questions** | **42** guiding Qs |
+| **Team** | Origin **3** |
+| **Data** | `WSC_Writing/` |
 
-### `IChTo` — International Chemistry Tournament · RUBRIC ✅
+### `science_bowl` · RULE ✅
 
 | | |
 |---|---|
-| **Domain** | IYPT-style chemistry fights (Reporter / Opponent / Reviewer) |
-| **Questions** | **106** open problems · 9 sets (2017–2024 & 2026: 12 each; 2025: 10) |
-| **Team setup** | 3+ with role rotation |
-| **Grading** | Jury rubric during oral fights |
-| **Data** | `IChTo/` — official problem-set PDFs (2017–2026) |
-| **Link** | [ichto.org/en/problems](http://ichto.org/en/problems/) |
+| **Questions** | **~23,691** |
+| **Team** | 4 + alternate |
+| **Data** | `science_bowl/` |
 
-### `PUMaC_Power` — PUMaC Power Round · RUBRIC ✅
+### `qanta` · RULE ✅
 
 | | |
 |---|---|
-| **Domain** | 8-person week-long team proof packet |
-| **Questions** | ~19 Power packets (2007–2025) |
-| **Team setup** | Origin/Rec **8** |
-| **Grading** | Proof rubric; official solutions as reference |
-| **Data** | `PUMaC_Power/` — problem (+ solution) PDFs |
-| **Link** | [PUMaC Archives](https://jason-shi-f9dm.squarespace.com/archives) |
+| **Questions** | ~100,000 |
+| **Team** | 4 |
+| **Data** | `qanta/` |
 
-### `MIT_Mystery_Hunt` — MIT Mystery Hunt · RULE ✅
+### `MIT_Mystery_Hunt` · RULE ✅
 
 | | |
 |---|---|
-| **Domain** | Large-team puzzle hunt → metas → coin |
-| **Questions** | **4,202** keyed answers; puzzle body/media mirror under `puzzles/` |
-| **Team setup** | Origin 5–150 · Rec 8–20 for lab sims |
-| **Grading** | Exact / normalized answer match; metas in `metapuzzles.yml` |
-| **Data** | `MIT_Mystery_Hunt/` — `mh_answers/`, `answers.tsv`, `puzzles/`, local README |
-| **Link** | [puzzles.mit.edu](https://puzzles.mit.edu/) · [mh_answers](https://github.com/dgulotta/mh_answers) |
+| **Questions** | **4,202** keyed answers |
+| **Team** | Origin **5–150** · Rec 8–20 |
+| **Data** | `MIT_Mystery_Hunt/` |
 
-### `IOL` — International Linguistics Olympiad · RULE ✅
+### `ARC-AGI-2` · RULE ✅
 
 | | |
 |---|---|
-| **Domain** | Linguistics olympiad (+ 4-person Team Contest) |
-| **Questions** | **130** source problems · 555 HF records (478 text-strict) |
-| **Team setup** | Individual Rec 2–3 · Team Rec **4** |
-| **Grading** | Match official / HF answer units |
-| **Data** | `IOL/` — `ioling_hf/` + `official_pdfs/` |
-| **Link** | [ioling.org](https://ioling.org/problems/by_year/) · [HF agurung/ioling](https://huggingface.co/datasets/agurung/ioling) |
-
-### `IESO` — International Earth Science Olympiad · RULE ✅
-
-| | |
-|---|---|
-| **Domain** | Earth science theory + practical |
-| **Questions** | ~125 theory/practical PDFs (+ keys) |
-| **Team setup** | Rec 2–3 on practicals |
-| **Grading** | Official keys |
-| **Data** | `IESO/` — theory/practical PDFs + keys |
-| **Link** | [Past IESO questions](http://www.geosocindia.org/index.php/ieso/Questions_From_Past_IESOs) |
-
-### `IPO` — International Philosophy Olympiad · RUBRIC ✅
-
-| | |
-|---|---|
-| **Domain** | Timed philosophy essay olympiad |
-| **Questions** | ~95 structured Topic prompts (`topics.jsonl`; pick 1 of ~4/year) |
-| **Team setup** | Rec 3 (drafter / critic / editor) |
-| **Grading** | IPO / national jury rubric |
-| **Data** | `IPO/topics/` — yearly topic files + `topics.jsonl`; `essays/` = samples (not prompts) |
-| **Link** | [philosophy-olympiad.org](https://www.philosophy-olympiad.org/) |
-
-### `History_Olympiad` — International History Olympiad · RULE + RUBRIC ✅
-
-| | |
-|---|---|
-| **Domain** | History bee/bowl + written historiography exams |
-| **Questions** | ~90 exam/key/rubric PDFs across 6 editions (2014, 2018–2019, 2022–2024) |
-| **Team setup** | Bee Rec 2–3 · Bowl Rec **4** |
-| **Grading** | RULE for keyed bees; RUBRIC for written exams |
-| **Data** | `History_Olympiad/` — year folders of exams/keys/rubrics |
-| **Link** | [historyolympiad.com/resources](https://www.historyolympiad.com/resources/) |
-
-### `IEO` — International Economics Olympiad · RULE + RUBRIC ✅
-
-| | |
-|---|---|
-| **Domain** | Economics exam + team business case |
-| **Questions** | ~1 open economics set (2018) + ~10+ business-case / task PDFs |
-| **Team setup** | Econ Rec 2–3 · Case Rec 3–5 |
-| **Grading** | RULE for economics keys; RUBRIC for case deck/pitch |
-| **Data** | `IEO/` — syllabus/regulations + task PDFs |
-| **Link** | [ieo-official.org/prepare](https://ieo-official.org/prepare) |
-
-### `IOAA` — Intl. Olympiad on Astronomy & Astrophysics · RULE ✅
-
-| | |
-|---|---|
-| **Domain** | Astronomy & astrophysics (incl. 5-person Group round) |
-| **Questions** | ~20 years · ~174–176 PDFs (2007–2025); 16 Group/Team packets |
-| **Team setup** | Individual Rec 2–3 · Group Rec **5** |
-| **Grading** | Official marking / numeric tolerances |
-| **Data** | `IOAA/` — theory / data / observation / group PDFs (~209 MB) |
-| **Link** | [IOAA past problems](https://ioaastrophysics.org/resources/problems-from-past-ioaa) |
-
-### `WSC_Writing` — World Scholar's Cup Collaborative Writing · RUBRIC ✅
-
-| | |
-|---|---|
-| **Domain** | 3-person staged collaborative essay writing |
-| **Questions** | **42** public guiding questions; live rounds use 3–4 sealed prompts/session |
-| **Team setup** | Origin/Rec **3** (plan → draft → peer edit; no devices) |
-| **Grading** | WSC collaborative-writing rubric |
-| **Data** | `WSC_Writing/` — guiding questions + discussion prompts |
-| **Link** | [scholarscup.org/events](https://scholarscup.org/events/) |
-
----
-
-## Evaluation design
-
-Datasets sit in three grading tiers. Every Tier 2/3 score must be produced by a judge that passed the calibration gate.
-
-**Tier 1 — rule-based auto-grading** (all RULE datasets above, incl. `HLE`, `GPQA`, `GAIA`, `ARC-AGI-2`, `MLE-bench`, `IOL`, `IESO`, `IOAA`, `MIT_Mystery_Hunt`): exact/normalized match, hidden test cases, `eval.py`, leaderboard rank, or flag match. No judge; fully reproducible.
-
-**Tier 2 — reference-anchored comparison** (`debatebench`, `gcch_harvard`, `cfa_research_challenge`, `vis_moot`, `PaperBench`): gold samples or official human scores exist. `debatebench` is the **calibration set** — validate any LLM adjudicator against its 360 human-scored speeches in `scores.xlsx` (Spearman vs speaker scores; pairwise ranking accuracy) before trusting it elsewhere. Comparisons run pairwise, position-swapped, vs winning samples → win rate.
-
-**Tier 3 — rubric-judge panel** (`modeling_agent`, `MLR-Bench`, `HealthBench`, `HealthBench-Professional`, `ethics_bowl_*`, `wharton_investment` report half, `EOES`, `IChTo`, `PUMaC_Power`, `IPO`, `WSC_Writing`, `IOAI` Team Challenge, `History_Olympiad` written tracks, `IEO` case, `iGeo` fieldwork, and the instance-rubric sets `WritingBench`, `ProfBench`, `BiGGen-Bench`, `FLASK`): no single gold reference. Run the real match/deliverable structure, score with a panel of ≥3 judges using the official rubric/criteria; adversarial contests aggregate team-vs-team outcomes into Elo. The instance-rubric sets are the easiest tier-3 case — each item ships its own written criteria, so the judge prompt is fully determined by the data (`BiGGen-Bench` also works as a secondary judge-calibration set alongside `debatebench`).
-
-**Judge validation & bias controls**
-- **Gate:** judge must pass `debatebench` calibration (target Spearman ≥ 0.5) before its Tier 2/3 scores count.
-- **Position bias:** pairwise comparisons run twice with order swapped; disagreement = tie.
-- **Verbosity bias:** enforce each contest's own length/time caps before judging.
-- **Panel:** ≥3 judge models/prompts, averaged; report inter-judge agreement.
-
-**Team-level metrics (all tiers):** log turn/token budget consumed, inter-agent messages, role-specialization entropy, and a solo-agent baseline at equal budget — the headline result is the **team-vs-solo delta**, not the absolute score.
-
-**Contamination:** prefer withheld/newest splits — OlympicArena `test`, AIME 2026, LiveCodeBench post-cutoff windows, newest case sets / 33rd Vis Problem / current Wharton case / IOAI 2025–2026 — for headline numbers; use older years as development splits.
+| **Questions** | 1,000 train + 120 eval |
+| **Team** | ARC Prize **team** format |
+| **Data** | `ARC-AGI-2/` |
+| **Link** | [arcprize/ARC-AGI-2](https://github.com/arcprize/ARC-AGI-2) |
 
 ---
 
 ## Downloading
 
-Locally-present datasets (✅) are already in this folder. To (re)download any dataset into `benchmark/`:
-
 ```bash
-# STEM olympiad + programming + engineering (rule-based)
-hf download Hothan/OlympiadBench             --repo-type dataset --local-dir benchmark/OlympiadBench
-hf download math-ai/aime24                   --repo-type dataset --local-dir benchmark/AIME/aime24
-hf download math-ai/aime25                   --repo-type dataset --local-dir benchmark/AIME/aime25
-hf download math-ai/aime26                   --repo-type dataset --local-dir benchmark/AIME/aime26
-hf download codegenning/usacobench_formatted --repo-type dataset --local-dir benchmark/USACO           # 3.8 GB
-hf download princeton-nlp/SWE-bench_Verified --repo-type dataset --local-dir benchmark/SWE-bench_Verified
-
-# Large — pull only when needed
-hf download livecodebench/code_generation_lite --repo-type dataset --local-dir benchmark/LiveCodeBench  # 4.4 GB
-hf download deepmind/code_contests             --repo-type dataset --local-dir benchmark/CodeContests    # 7.6 GB
-
-# Research + medicine (rubric)
-hf download chchenhui/mlrbench-tasks           --repo-type dataset --local-dir benchmark/MLR-Bench
-hf download openai/healthbench                 --repo-type dataset --local-dir benchmark/HealthBench
-hf download openai/healthbench-professional    --repo-type dataset --local-dir benchmark/HealthBench-Professional
-
-# Writing & professional deliverables (rubric, all small)
-hf download nvidia/ProfBench                   --repo-type dataset --local-dir benchmark/ProfBench
-hf download prometheus-eval/BiGGen-Bench       --repo-type dataset --local-dir benchmark/BiGGen-Bench
-curl -Lo benchmark/WritingBench/benchmark_all.jsonl https://raw.githubusercontent.com/X-PLUG/WritingBench/main/benchmark_query/benchmark_all.jsonl
-
-# International olympiads (section 13)
+hf download deepmind/code_contests --repo-type dataset --local-dir benchmark/CodeContests
 hf download IOAI-official/IOAI2025 --repo-type dataset --local-dir benchmark/IOAI/IOAI2025
-hf download agurung/ioling         --repo-type dataset --local-dir benchmark/IOL/ioling_hf
+hf download agurung/ioling --repo-type dataset --local-dir benchmark/IOL/ioling_hf
 git clone --depth 1 https://github.com/dgulotta/mh_answers.git benchmark/MIT_Mystery_Hunt/mh_answers
-# EOES / IChTo / PUMaC / IESO / IPO / History / IEO / IOAA / WSC / iGeo:
-#   official PDF / HTML archives (see from_zhongzheng/olympiad deep-dive cards)
-
-# Gated on HF — visit the dataset page, accept the terms, then:
-#   hf download cais/hle            --repo-type dataset --local-dir benchmark/HLE     # 274 MB
-#   hf download Idavidrein/gpqa     --repo-type dataset --local-dir benchmark/GPQA    # 9 MB
-#   hf download gaia-benchmark/GAIA --repo-type dataset --local-dir benchmark/GAIA
-
-# Original three (private mirror)
-#   hf download akaaafk/multiagent_bench --repo-type dataset --include "olympiad/*" --local-dir <tmp>   # OlympicArena, MLAgentBench, modeling_agent
-
-# Docker / GitHub-based (not plain HF pulls)
-#   NYU CTF Bench : pip install nyuctf ; challenges cloned on first CTFDataset(split=...) call
-#   Cybench       : local packs in benchmark/Cybench/benchmark/; Docker images for full eval
-#   PaperBench    : git clone https://github.com/openai/preparedness --filter=blob:none ; git lfs fetch --include "project/paperbench/data/**"
-#   MLE-bench     : git clone https://github.com/openai/mle-bench ; Kaggle API creds ; mlebench prepare --lite
-#   ARC-AGI-2     : git clone https://github.com/arcprize/ARC-AGI-2
-#   FLASK         : git clone https://github.com/kaistAI/FLASK
+git clone --depth 1 https://github.com/arcprize/ARC-AGI-2.git benchmark/ARC-AGI-2
+# CTF / PDF olympiads: see from_zhongzheng/olympiad deep-dive cards
 ```
diff --git a/benchmark/WSC_Writing/from_repo_raw/rubric/wsc_writing_rubric.pdf b/benchmark/WSC_Writing/from_repo_raw/rubric/wsc_writing_rubric.pdf
new file mode 100644
index 0000000000000000000000000000000000000000..62350b41a70108ce49045bd9b02dff67707dfb6d
--- /dev/null
+++ b/benchmark/WSC_Writing/from_repo_raw/rubric/wsc_writing_rubric.pdf
@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:1084f231ad3a7cfea338e828c0ace1eaceb73621e551011fa664d7a88c3b6cba
+size 171991
diff --git a/benchmark/qanta/qanta.train.2021.12.20.json b/benchmark/qanta/qanta.train.2021.12.20.json
new file mode 100644
index 0000000000000000000000000000000000000000..bd8cc420ad15467c76203e60deb74b85093bd273
--- /dev/null
+++ b/benchmark/qanta/qanta.train.2021.12.20.json
@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:65b5c424fabdc0a700f6ef1468c4bcc21313501f41dfd2db2b00470d01f6861d
+size 177570997
diff --git a/benchmark/science_bowl/MS-Sample-Questions/Sample-Set-10/10A_MS_Reg_2016.pdf b/benchmark/science_bowl/MS-Sample-Questions/Sample-Set-10/10A_MS_Reg_2016.pdf
new file mode 100644
index 0000000000000000000000000000000000000000..fb6744a83571740f713a4d5e88aa905d2f87c952
--- /dev/null
+++ b/benchmark/science_bowl/MS-Sample-Questions/Sample-Set-10/10A_MS_Reg_2016.pdf
@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:fec29a05794869db4e1fe0fa6c3927630b9c212e84c9918d405a89042527dbde
+size 235369
diff --git a/benchmark/science_bowl/MS-Sample-Questions/Sample-Set-10/11A_MS_Reg_2016.pdf b/benchmark/science_bowl/MS-Sample-Questions/Sample-Set-10/11A_MS_Reg_2016.pdf
new file mode 100644
index 0000000000000000000000000000000000000000..2ee9d9a697a7e16dfc668b934eb07ace0dd6f393
--- /dev/null
+++ b/benchmark/science_bowl/MS-Sample-Questions/Sample-Set-10/11A_MS_Reg_2016.pdf
@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:c8ebb767541f0cb413d332ef1a48d0958be5022716ab47ab2c640271777152d5
+size 278698
diff --git a/benchmark/science_bowl/MS-Sample-Questions/Sample-Set-10/12A_MS_Reg_2016.pdf b/benchmark/science_bowl/MS-Sample-Questions/Sample-Set-10/12A_MS_Reg_2016.pdf
new file mode 100644
index 0000000000000000000000000000000000000000..3372c65e36e6eff624831c6055ac2d28fad6a58f
--- /dev/null
+++ b/benchmark/science_bowl/MS-Sample-Questions/Sample-Set-10/12A_MS_Reg_2016.pdf
@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:339d17ce4d0d247108b4af6cab74f41422f34809e30112d928217a2919d36144
+size 237485
diff --git a/benchmark/science_bowl/MS-Sample-Questions/Sample-Set-10/14A_MS_Reg_2016.pdf b/benchmark/science_bowl/MS-Sample-Questions/Sample-Set-10/14A_MS_Reg_2016.pdf
new file mode 100644
index 0000000000000000000000000000000000000000..0dbf552a9ea5fd08d31594faf41b3aa5f4e0e324
--- /dev/null
+++ b/benchmark/science_bowl/MS-Sample-Questions/Sample-Set-10/14A_MS_Reg_2016.pdf
@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:44d6d96f2a3417f491c8c5415a4b3f0552b014e2bf0cab7d298de253c93ad33f
+size 236422
diff --git a/benchmark/science_bowl/MS-Sample-Questions/Sample-Set-10/15A_MS_Reg_2016.pdf b/benchmark/science_bowl/MS-Sample-Questions/Sample-Set-10/15A_MS_Reg_2016.pdf
new file mode 100644
index 0000000000000000000000000000000000000000..9383c9391ad162dc9e91976068a1a6dd4d9dad63
--- /dev/null
+++ b/benchmark/science_bowl/MS-Sample-Questions/Sample-Set-10/15A_MS_Reg_2016.pdf
@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:292c059460fb9926a4ae87553792c426af2f04497601139055d413652c05e410
+size 238351
diff --git a/benchmark/science_bowl/MS-Sample-Questions/Sample-Set-10/16A_MS_Reg_2016.pdf b/benchmark/science_bowl/MS-Sample-Questions/Sample-Set-10/16A_MS_Reg_2016.pdf
new file mode 100644
index 0000000000000000000000000000000000000000..ae638c072bfa81707ff8f74886df914466a886b6
--- /dev/null
+++ b/benchmark/science_bowl/MS-Sample-Questions/Sample-Set-10/16A_MS_Reg_2016.pdf
@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:3aaf140d02298e4387529b8f733a5b0bc661779876f6201a9a7d7a8da78247eb
+size 236600
diff --git a/benchmark/science_bowl/MS-Sample-Questions/Sample-Set-10/17A_MS_Reg_2016.pdf b/benchmark/science_bowl/MS-Sample-Questions/Sample-Set-10/17A_MS_Reg_2016.pdf
new file mode 100644
index 0000000000000000000000000000000000000000..c4350c7cb738cd9b6e7ef23724d9f7dcced5fd03
--- /dev/null
+++ b/benchmark/science_bowl/MS-Sample-Questions/Sample-Set-10/17A_MS_Reg_2016.pdf
@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:fc5d029912b2d58d4ac39ce42a5c94b2c99a0ee9b0e72f6ad2c44223273b83bc
+size 238052
diff --git a/benchmark/science_bowl/MS-Sample-Questions/Sample-Set-10/1A_MS_Reg_2016.pdf b/benchmark/science_bowl/MS-Sample-Questions/Sample-Set-10/1A_MS_Reg_2016.pdf
new file mode 100644
index 0000000000000000000000000000000000000000..672cec380b089e5012d0ab0aa8b989efefb19457
--- /dev/null
+++ b/benchmark/science_bowl/MS-Sample-Questions/Sample-Set-10/1A_MS_Reg_2016.pdf
@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:3992d449edefce9e938ca1b59e32b928972e1e7e5fad7acb71b3a3b7f71d82bd
+size 233527
diff --git a/benchmark/science_bowl/MS-Sample-Questions/Sample-Set-10/2A_MS_Reg_2016.pdf b/benchmark/science_bowl/MS-Sample-Questions/Sample-Set-10/2A_MS_Reg_2016.pdf
new file mode 100644
index 0000000000000000000000000000000000000000..f28baaa0d371ba59bcfbc3885cd2c9c5219823d9
--- /dev/null
+++ b/benchmark/science_bowl/MS-Sample-Questions/Sample-Set-10/2A_MS_Reg_2016.pdf
@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:68d021d46d4738c2a0dcd01428135049c5feca282133b892357824d1bcae535b
+size 231822
diff --git a/benchmark/science_bowl/MS-Sample-Questions/Sample-Set-10/3A_MS_Reg_2016.pdf b/benchmark/science_bowl/MS-Sample-Questions/Sample-Set-10/3A_MS_Reg_2016.pdf
new file mode 100644
index 0000000000000000000000000000000000000000..2352db48a7982ccb3f3e310439527db2860174bc
--- /dev/null
+++ b/benchmark/science_bowl/MS-Sample-Questions/Sample-Set-10/3A_MS_Reg_2016.pdf
@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:deb723c4488d22c5c57d1826f5fc66b84f51d3b5d3ac9d8a6a5d9e196ae1e7b6
+size 282383
diff --git a/benchmark/science_bowl/MS-Sample-Questions/Sample-Set-10/4A_MS_Reg_2016.pdf b/benchmark/science_bowl/MS-Sample-Questions/Sample-Set-10/4A_MS_Reg_2016.pdf
new file mode 100644
index 0000000000000000000000000000000000000000..4400b1cc2aae75c21bd6068370b958aece189b1c
--- /dev/null
+++ b/benchmark/science_bowl/MS-Sample-Questions/Sample-Set-10/4A_MS_Reg_2016.pdf
@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:7f13e29240bcc2739bdca76e6cd1cfe8391e20eb1c1385275ad0620695bac52a
+size 235383
diff --git a/benchmark/science_bowl/MS-Sample-Questions/Sample-Set-10/6A_MS_Reg_2016.pdf b/benchmark/science_bowl/MS-Sample-Questions/Sample-Set-10/6A_MS_Reg_2016.pdf
new file mode 100644
index 0000000000000000000000000000000000000000..7be99356477f66a7e392b536925170a425c0c1ce
--- /dev/null
+++ b/benchmark/science_bowl/MS-Sample-Questions/Sample-Set-10/6A_MS_Reg_2016.pdf
@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:1eabb16a07a504e1b72333cfe652a66dbca19e9a3eda36c52d4f93cd3728c64f
+size 237092
diff --git a/benchmark/vis_moot/26th-2019-problem-incl-PO2.pdf b/benchmark/vis_moot/26th-2019-problem-incl-PO2.pdf
new file mode 100644
index 0000000000000000000000000000000000000000..5a4ff4e951f40939f95531c6b2bc52c58cafffa9
--- /dev/null
+++ b/benchmark/vis_moot/26th-2019-problem-incl-PO2.pdf
@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:7dc6b4b137b003ad90a1bc4de1747123ae8712824f49e00a0203b4c3ba62f55d
+size 1101747
diff --git a/benchmark/vis_moot/28th-2021-problem-incl-appendix.pdf b/benchmark/vis_moot/28th-2021-problem-incl-appendix.pdf
new file mode 100644
index 0000000000000000000000000000000000000000..af4c674c02709809777fcc4b837c478fe2705909
--- /dev/null
+++ b/benchmark/vis_moot/28th-2021-problem-incl-appendix.pdf
@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:9f575f7d7d14a0c5bbe2e0b7d0be17751ff28fca206174bfdd4fc55ccd8b1872
+size 1838641
diff --git a/benchmark/vis_moot/29th-2022-problem-incl-PO2.pdf b/benchmark/vis_moot/29th-2022-problem-incl-PO2.pdf
new file mode 100644
index 0000000000000000000000000000000000000000..d03426aa88c38c60e7dbceef9bc8362c065f38b6
--- /dev/null
+++ b/benchmark/vis_moot/29th-2022-problem-incl-PO2.pdf
@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:9f031ee5258db4dcd468babfa6625830319dbaeaef240f8da25078c6d7eabf58
+size 1788076
diff --git a/benchmark/vis_moot/30th-2023-problem.pdf b/benchmark/vis_moot/30th-2023-problem.pdf
new file mode 100644
index 0000000000000000000000000000000000000000..b6e6398ad253cbaf3c1cd1dccc5bb4f8f0c0b33a
--- /dev/null
+++ b/benchmark/vis_moot/30th-2023-problem.pdf
@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:ea233400c4744c39fddc0c1290c7d2a54e00dafd50fd436b241d9dbb30c8f422
+size 2190593
diff --git a/benchmark/vis_moot/31st-2024-problem-incl-PO2.pdf b/benchmark/vis_moot/31st-2024-problem-incl-PO2.pdf
new file mode 100644
index 0000000000000000000000000000000000000000..4337a7056948d7b3ff819dab73118405c3fd96a0
--- /dev/null
+++ b/benchmark/vis_moot/31st-2024-problem-incl-PO2.pdf
@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:c99d9bbc68c37df8ea63be608d7671e8e8d917b9f5480ed02affd508cb64a8ac
+size 1811131
diff --git a/benchmark/vis_moot/32nd-2025-problem-incl-PO2.pdf b/benchmark/vis_moot/32nd-2025-problem-incl-PO2.pdf
new file mode 100644
index 0000000000000000000000000000000000000000..6251d7ff74f7e8e97b75b0741f9dca8c9c7ebb54
--- /dev/null
+++ b/benchmark/vis_moot/32nd-2025-problem-incl-PO2.pdf
@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:39b1eaf9c239d7439dac0b48edd598ceca6d9a51d93236b1dea119d0fa350df2
+size 2096056
diff --git a/benchmark/vis_moot/33rd-2026-problem-incl-PO2.pdf b/benchmark/vis_moot/33rd-2026-problem-incl-PO2.pdf
new file mode 100644
index 0000000000000000000000000000000000000000..bc18dec977b22c9d149863f8d9a3eee6e21ffecd
--- /dev/null
+++ b/benchmark/vis_moot/33rd-2026-problem-incl-PO2.pdf
@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:d56a8188402ac438d34066272ae206608cc669b6e07039fc38b7ce4b6a4e0ca9
+size 2529834
diff --git a/benchmark/vis_moot/33rd-2026-rules.pdf b/benchmark/vis_moot/33rd-2026-rules.pdf
new file mode 100644
index 0000000000000000000000000000000000000000..65ab81edd63fe80f538132622fcdf5993056bcc4
--- /dev/null
+++ b/benchmark/vis_moot/33rd-2026-rules.pdf
@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:81ccbc49f2de0a7196da341efaf93e2d8bfe3a0bb98f40946d6576b1c62d9e6d
+size 811655