Upload README.md with huggingface_hub

README.md

@@ -1,33 +1,89 @@
 ---
-dataset_info:
-  features:
-  - name: task_name
-    dtype: string
-  - name: flag
-    dtype: string
-  - name: flag_x86_64
-    dtype: string
-  - name: flag_aarch64_linux
-    dtype: string
-  - name: flag_riscv
-    dtype: string
-  - name: x86_64
-    dtype: string
-  - name: aarch64_linux
-    dtype: string
-  - name: riscv
-    dtype: string
-  - name: test_c
-    dtype: string
-  splits:
-  - name: test
-    num_bytes: 765634
-    num_examples: 10
-  download_size: 120241
-  dataset_size: 765634
-configs:
-- config_name: default
-  data_files:
-  - split: test
-    path: data/test-*
+language:
+  - en
+license: mit
+task_categories:
+  - text-generation
+tags:
+  - assembly
+  - security
+  - constant-time
+  - side-channels
+  - binsec
+size_categories:
+  - n<1K
 ---
+
+# RelBench-O2: Constant-Time Assembly at -O2
+
+## Overview
+
+This dataset collects the **rel_bench** constant-time programs, re-cut into one
+self-contained program per record, compiled with LLVM/Clang-17 at **-O2** for three
+targets, and checked for constant-timeness with `binsec -sse -checkct`.
+
+Constant-time source code is not enough: the optimizer can reintroduce a secret-dependent
+branch or memory access that the -O0 build does not have, and it does so *differently per
+target*. The intended tasks are **transpilation** (translate the assembly to another ISA)
+and **classification/repair** (decide whether a build leaks, and fix it if it does).
+
+## Dataset Structure
+
+| Column | Description |
+|--------|-------------|
+| `task_name` | Sample identifier, e.g. `ct_select_1` |
+| `flag` | BinSec constant-time verdict: `secure`, `insecure`, or `unknown` -- the `x86_64` build's verdict |
+| `flag_x86_64` | BinSec verdict for the x86-64 `-O2` build |
+| `flag_aarch64_linux` | BinSec verdict for the AArch64 `-O2` build |
+| `flag_riscv` | BinSec verdict for the RISC-V `-O2` build |
+| `x86_64` | x86-64 assembly at -O2 (**Intel syntax**, built with `-masm=intel`) |
+| `aarch64_linux` | AArch64 assembly at -O2 |
+| `riscv` | RISC-V RV64 (`rv64gc`/`lp64d`) assembly at -O2 |
+| `test_c` | C source of the BinSec harness (`test.c`) -- tags inputs secret/public through the `high_input_N` / `low_input_N` markers and calls the function under analysis |
+
+## Statistics
+
+- **Total records**: 10
+
+| flag | count |
+|---|---|
+| `secure` | 7 |
+| `insecure` | 3 |
+
+## The constant-time oracle
+
+Each build is analysed by BinSec's `checkct` plugin:
+
+```bash
+binsec -sse -checkct -sse-script "libsym.ini,checkct.cfg" -sse-depth 50000 test_bin
+```
+
+`libsym.ini` replaces the `high_input_N` markers with `secret` and the `low_input_N`
+markers with `nondet` (public), so BinSec knows which bytes are secret. It then reports:
+
+- `secure` -- no control flow or memory access depends on a secret
+- `insecure` -- at least one secret-dependent branch or address was found
+- `unknown` -- BinSec could not decide (path cut, depth limit, unresolved symbol)
+
+**The verdict is per build.** The same program can be secure on one target and insecure on
+another, because the three backends make different branch/`cmov`/`csel` choices at -O2.
+This dataset's single `flag` column reports the `x86_64` build's verdict; the per-build verdicts are kept
+in `flag_x86_64`, `flag_aarch64_linux` and `flag_riscv`.
+
+## Usage
+
+```python
+from datasets import load_dataset
+
+ds = load_dataset("Akirayasha/relbench_O2", split="test")
+
+record = ds[0]
+src_asm = record["x86_64"]  # source-side assembly
+tgt_asm = record["riscv"]   # reference target-side assembly
+leaky   = record["flag_riscv"] == "insecure"   # or flag / flag_x86_64 / flag_aarch64_linux
+harness = record["test_c"]  # rebuild + recheck a candidate against BinSec
+```
+
+## Citation
+
+Part of the CISC-to-RISC transpilation research project at MBZUAI.