Publish canonical GLM-4.7 PIE C++ post-training dataset

.gitattributes

@@ -58,3 +58,5 @@ saved_model/**/* filter=lfs diff=lfs merge=lfs -text
 # Video files - compressed
 *.mp4 filter=lfs diff=lfs merge=lfs -text
 *.webm filter=lfs diff=lfs merge=lfs -text
+grpo/train.jsonl filter=lfs diff=lfs merge=lfs -text
+sft/train.jsonl filter=lfs diff=lfs merge=lfs -text

README.md

@@ -0,0 +1,65 @@
+---
+license: other
+task_categories:
+  - text-generation
+language:
+  - en
+tags:
+  - code
+  - code-optimization
+  - cpp
+  - sft
+  - grpo
+configs:
+  - config_name: sft
+    data_files:
+      - split: train
+        path: sft/train.jsonl
+  - config_name: grpo
+    data_files:
+      - split: train
+        path: grpo/train.jsonl
+  - config_name: evaluation
+    data_files:
+      - split: validation
+        path: eval/validation.jsonl
+---
+
+# GLM-4.7-Flash PIE C++ Post-Training Data
+
+The exact prepared dataset used for the GLM-4.7-Flash C++ performance
+post-training runs.
+
+## Splits
+
+| File | Rows | Purpose |
+| --- | ---: | --- |
+| `sft/train.jsonl` | 7,864 | Supervised fine-tuning |
+| `grpo/train.jsonl` | 7,887 | GRPO prompt and reward evaluation |
+| `eval/validation.jsonl` | 1,259 | Full held-out evaluation |
+| `eval/validation_mini126.jsonl` | 126 | Fast evaluation |
+| `eval/validation_mini4.jsonl` | 4 | Smoke evaluation |
+
+Each row carries a stable task ID, problem ID, split, prompt, label, and task
+metadata. SFT rows additionally contain the user and assistant messages used by
+the trainer.
+
+## Preparation
+
+The source tasks were filtered by compiling and executing the oracle answer in
+the C++ sandbox. Training rows were retained when the response format was valid,
+all tests passed, and the reward result was `correct`.
+
+The source preparation run scored 8,350 training tasks, retained 7,887, and
+rejected 463. The complete preparation receipt is in `manifest.json`.
+
+## Provenance
+
+This is a prepared derivative of the
+[PIE C++ performance dataset](https://github.com/madaan/pie-perf), which is
+based on IBM Project CodeNet. Use and redistribution are subject to the
+applicable upstream dataset terms.
+
+Training code:
+[TokenBender/browser-is-all-you-need](https://github.com/tokenbender/browser-is-all-you-need/tree/client/glm47-h100-posttraining)
+

SHA256SUMS

@@ -0,0 +1,7 @@
+62587948eb1e5e56a8ad40e120f1ff2158632d58af954f3a37c6bd03f822d562  ./README.md
+9297d35ed42a807239bcea812931045673475e4389220ea0b732115788dade6e  ./eval/validation.jsonl
+e1efd4bc975310dd34946c4ec9705187f82d67e7ab30d41643f063b48b372929  ./eval/validation_mini126.jsonl
+83ea467298740f35f49242c909c9f8d636c7f23ac69d62dac8e3bc106a69d33f  ./eval/validation_mini4.jsonl
+216e961127630ef687a2908b8e822d17f39dc248fcb1ab912840e15598ab1101  ./grpo/train.jsonl
+9660523f3044ee30b73a01d68f516a1b08a3a404abec4fa653cc43e0e1ab0cd0  ./manifest.json
+047007ba4f575ae694d29ef9cb8294b31fc1e7263e6c21a85be58fb81ba7afa6  ./sft/train.jsonl

eval/validation_mini4.jsonl

@@ -0,0 +1,4 @@
+{"label": "pie_cpp_test_001506", "metadata": {"data_source": "pie-cpp-perf", "gem5_cycles": 11, "hidden_test_count": 1, "problem_id": "p03766", "reference_metric": "runtime_cpu_ns", "reference_value": 11, "source_task_path": "/workspace/runs/pie-tasks-full-20260706/test/pie_cpp_test_001506.json", "split": "test", "subset": "eval", "task_id": "pie_cpp_test_001506", "task_path": "tasks/test/pie_cpp_test_001506.json", "visible_test_count": 1}, "problem_id": "p03766", "prompt": "Optimize the following C++20 program while preserving its exact behavior.\nReturn exactly one <reasoning>...</reasoning> block followed by exactly one fenced cpp code block.\nDo not mention or rely on hidden tests. The program must read stdin and write stdout.\n\nVisible tests:\nInput:\n2\n\nExpected output:\n4\n\n\nProgram:\n```cpp\n#include <bits/stdc++.h>\n\nusing namespace std;\n\n#define int long long\n\n#define in ({int x=0;int c=getchar(),n=0;for(;!isdigit(c);c=getchar()) n=(c=='-');for(;isdigit(c);c=getchar()) x=x*10+c-'0';n?-x:x;})\n\nmt19937 rng(chrono::steady_clock::now().time_since_epoch().count());\n\nint rnd(int l,int r){return l+rng()%(r-l+1);}\n\n#define fasty ios_base::sync_with_stdio(0),cin.tie(0);\n\n#define forinc(a,b,c) for(int a=b,_c=c;a<=_c;++a)\n\n#define fordec(a,b,c) for(int a=b,_c=c;a>=_c;--a)\n\n#define forv(a,b) for(auto&a:b)\n\n#define fi first\n\n#define se second\n\n#define pb push_back\n\n#define ii pair<int,int>\n\n#define mt make_tuple\n\n#define all(a) a.begin(),a.end()\n\n#define reset(f, x) memset(f, x, sizeof(f))\n\n#define bit(x,i) ((x>>(i-1))&1)\n\n#define on(x,i) (x|(1ll<<(i-1)))\n\n#define off(x,i) (x&~(1<<(i-1)))\n\n#define gg exit(0);\n\n\n\n\n\nconst int M=1000000007;\n\n\n\nint n,ret=1,tmp=1;\n\nint f[1000010];\n\n\n\nmain(){\n\n    #define task \"TASK\"\n\n    //freopen(task\".inp\",\"r\",stdin);\n\n    //freopen(task\".out\",\"w\",stdout);\n\n\n\n    n=in;\n\n    forinc(i,1,n){\n\n        f[i]=(f[i]+f[i-1])%M;\n\n        tmp=(tmp+f[i])%M;\n\n        f[i+3]=(f[i+3]+tmp)%M;\n\n        ret=(ret+tmp%M*(n-1)%M*(i==n ? 1 : n)%M)%M;\n\n    }\n\n    cout<<ret;\n\n}\n```", "split": "test", "task_id": "pie_cpp_test_001506"}
+{"label": "pie_cpp_test_001531", "metadata": {"data_source": "pie-cpp-perf", "gem5_cycles": 6, "hidden_test_count": 1, "problem_id": "p03766", "reference_metric": "runtime_cpu_ns", "reference_value": 6, "source_task_path": "/workspace/runs/pie-tasks-full-20260706/test/pie_cpp_test_001531.json", "split": "test", "subset": "eval", "task_id": "pie_cpp_test_001531", "task_path": "tasks/test/pie_cpp_test_001531.json", "visible_test_count": 1}, "problem_id": "p03766", "prompt": "Optimize the following C++20 program while preserving its exact behavior.\nReturn exactly one <reasoning>...</reasoning> block followed by exactly one fenced cpp code block.\nDo not mention or rely on hidden tests. The program must read stdin and write stdout.\n\nVisible tests:\nInput:\n2\n\nExpected output:\n4\n\n\nProgram:\n```cpp\n#include <bits/stdc++.h>\n\nusing namespace std;\n\n\n\nint main(){\n\n    constexpr long MOD = 1000000007;\n\n    long N;\n\n    cin >> N;\n\n    long dp{N};\n\n    vector<long> sum(4, (N - 1) * N % MOD);\n\n    sum[0] = 0;\n\n    for(long i = 3, j = 2; i != N + 3; ++i, ++j)sum[i & 3] = (sum[j & 3] - 1 + (dp += sum[(j ^ 2) & 3])) % MOD;\n\n    cout << dp % MOD << endl;\n\n    return 0;\n\n}\n```", "split": "test", "task_id": "pie_cpp_test_001531"}
+{"label": "pie_cpp_test_001533", "metadata": {"data_source": "pie-cpp-perf", "gem5_cycles": 17, "hidden_test_count": 1, "problem_id": "p03766", "reference_metric": "runtime_cpu_ns", "reference_value": 17, "source_task_path": "/workspace/runs/pie-tasks-full-20260706/test/pie_cpp_test_001533.json", "split": "test", "subset": "eval", "task_id": "pie_cpp_test_001533", "task_path": "tasks/test/pie_cpp_test_001533.json", "visible_test_count": 1}, "problem_id": "p03766", "prompt": "Optimize the following C++20 program while preserving its exact behavior.\nReturn exactly one <reasoning>...</reasoning> block followed by exactly one fenced cpp code block.\nDo not mention or rely on hidden tests. The program must read stdin and write stdout.\n\nVisible tests:\nInput:\n2\n\nExpected output:\n4\n\n\nProgram:\n```cpp\n#include<bits/stdc++.h>\n\n#define re register\n\n#define N 1000005\n\n#define mod 1000000007\n\nusing namespace std;\n\ntemplate<typename _int>\n\ninline void read(re _int&x){\n\n\tre _int res=0,flag=1;re char opt;\n\n\twhile((opt=getchar())<'0'||opt>'9')if(opt=='-')flag=-1;\n\n\twhile(opt>='0'&&opt<='9'){res=(res<<1)+(res<<3)+opt-48;opt=getchar();}\n\n\tx=res*flag;\n\n}\n\ntypedef long long ll; \n\nll n,dp[N],sum[N],ans;\n\nint main(void){\n\n\tre ll i;*dp=1;*sum=1;\n\n\tread(n);ans=(n>2);\n\n\tif(n==1)return puts(\"1\"),0;\n\n\tfor(i=1;i<=n;++i){\n\n\t\t(dp[i]+=(sum[i-1]-((i>=2)?dp[i-2]:0)+mod)%mod)%=mod;\n\n//\t\tprintf(\"%lld %lld\\n\",sum[i-1],((i>=2)?dp[i-2]:0));\n\n\t\tsum[i]=(sum[i-1]+dp[i])%mod;\n\n\t\tif(i<n-1)(ans+=(1ll*(n-1)*(n-1)%mod+min(i+2,n-1))*dp[i]%mod)%=mod;\n\n\t}\n\n\tprintf(\"%lld\\n\",((ans+1ll*n*dp[n-1]%mod)%mod+1ll*(n-1)*(n-1)%mod+1)%mod);\n\n\treturn 0;\n\n}\n```", "split": "test", "task_id": "pie_cpp_test_001533"}
+{"label": "pie_cpp_test_001549", "metadata": {"data_source": "pie-cpp-perf", "gem5_cycles": 11, "hidden_test_count": 1, "problem_id": "p03766", "reference_metric": "runtime_cpu_ns", "reference_value": 11, "source_task_path": "/workspace/runs/pie-tasks-full-20260706/test/pie_cpp_test_001549.json", "split": "test", "subset": "eval", "task_id": "pie_cpp_test_001549", "task_path": "tasks/test/pie_cpp_test_001549.json", "visible_test_count": 1}, "problem_id": "p03766", "prompt": "Optimize the following C++20 program while preserving its exact behavior.\nReturn exactly one <reasoning>...</reasoning> block followed by exactly one fenced cpp code block.\nDo not mention or rely on hidden tests. The program must read stdin and write stdout.\n\nVisible tests:\nInput:\n2\n\nExpected output:\n4\n\n\nProgram:\n```cpp\n#include<bits/stdc++.h>\n\nusing namespace std;\n\n#ifndef ONLINE_JUDGE\n\n#define dbg(x...) do{cout << \"\\033[32;1m\" << #x << \"->\" ; err(x);} while(0)\n\nvoid err(){cout << \"\\033[39;0m\" << endl;}\n\ntemplate<template<typename...> class T,typename t,typename... A>\n\nvoid err(T<t> a,A... x){for (auto v:a) cout << v << ' '; err(x...);}\n\ntemplate<typename T,typename... A>\n\nvoid err(T a,A... x){cout << a << ' '; err(x...);}\n\n#else\n\n#define dbg(...)\n\n#endif\n\ntypedef long long ll;\n\ntypedef pair<int,int> pi;\n\ntypedef vector<int> vi;\n\ntemplate<class T> using vc=vector<T>;\n\ntemplate<class T> using vvc=vc<vc<T>>;\n\ntemplate<class T> void mkuni(vector<T>&v)\n\n{\n\n    sort(v.begin(),v.end());\n\n    v.erase(unique(v.begin(),v.end()),v.end());\n\n}\n\ntemplate<class T>\n\nvoid print(T x,int suc=1)\n\n{\n\n    cout<<x;\n\n    if(suc==1) cout<<'\\n';\n\n    else cout<<' ';\n\n}\n\ntemplate<class T>\n\nvoid print(const vector<T>&v,int suc=1)\n\n{\n\n    for(int i=0;i<v.size();i++)\n\n        print(v[i],i==(int)(v.size())-1?suc:2);\n\n}\n\nconst int INF=0x3f3f3f3f;\n\nconst int mod=1e9+7;\n\nconst int maxn=1e6+7;\n\nll dp[maxn];\n\nll pre[maxn];\n\nint main()\n\n{\n\n\tint n;\n\n    cin>>n;\n\n    dp[1]=n;\n\n    pre[1]=n;\n\n    for(int i=2;i<=n;i++)\n\n    {\n\n        dp[i]=(dp[i-1]+1LL*(n-1)*(n-1)%mod)%mod;\n\n        if(i>=3) dp[i]=(dp[i]+pre[i-3])%mod;\n\n        if(i==2) dp[i]=(dp[i]+n-1)%mod;\n\n        else dp[i]=(dp[i]+n-i+2);\n\n        pre[i]=(pre[i-1]+dp[i])%mod;\n\n        //dbg(i,dp[i]);\n\n    }\n\n    cout<<dp[n]<<'\\n';\n\n}\n```", "split": "test", "task_id": "pie_cpp_test_001549"}

grpo/train.jsonl

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:216e961127630ef687a2908b8e822d17f39dc248fcb1ab912840e15598ab1101
+size 19084087

manifest.json

@@ -0,0 +1,40 @@
+{
+  "counts": {
+    "copied_tasks": 9146,
+    "eval": 1259,
+    "train": 7887
+  },
+  "data_source": "pie-cpp-perf",
+  "eval_splits": [
+    "validation",
+    "test"
+  ],
+  "files": {
+    "eval": "eval/validation.jsonl",
+    "grpo_train": "grpo/train.jsonl",
+    "oracle_filter": "oracle_filter/train.jsonl",
+    "sft_train": "sft/train.jsonl"
+  },
+  "filter_train_oracle_full_marks": true,
+  "kind": "slime-cpp-perf-dataset",
+  "oracle_filter": {
+    "dropped": 463,
+    "keep_rule": "strict format, reward reason == correct, and all tests pass",
+    "kept": 7887,
+    "reason_counts": {
+      "compile_error": 261,
+      "correct": 7887,
+      "missing_runtime": 2,
+      "sanitizer_error": 9,
+      "tests_failed": 170,
+      "timeout": 21
+    },
+    "scored": 8350
+  },
+  "output_dir": "/workspace/runs/glm47-pie-data-sft-full",
+  "profile": "glm47-sft-full",
+  "run_id": "glm47_sft_data_20260706",
+  "schema_version": 1,
+  "sort_by_size": true,
+  "source_tasks_dir": "/workspace/runs/pie-tasks-full-20260706"
+}

sft/train.jsonl

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:047007ba4f575ae694d29ef9cb8294b31fc1e7263e6c21a85be58fb81ba7afa6
+size 32271395