Add per-sample metrics + dataset-viewer card

metrics/*.jsonl: one row per sample (path, source, in/out bytes, kept%,
opaque in→out, and the input/output code), sharded under 10MB so the HF
Dataset Viewer renders a browsable in→out table without LFS. README card
points the viewer at metrics/*.jsonl and embeds the aggregate scoreboard.

README.md

@@ -1,3 +1,93 @@
 ---
 license: mit
+pretty_name: devirt-corpus
+tags:
+- code
+- javascript
+- deobfuscation
+- obfuscation
+size_categories:
+- 1K<n<10K
+configs:
+- config_name: default
+  data_files:
+  - split: train
+    path: metrics/*.jsonl
 ---
+
+# devirt-corpus
+
+A corpus of obfuscated JavaScript with **per-sample deobfuscation metrics** for
+the [`devirt-core`](https://github.com/devirt-dev/devirt-core) engine. Every
+sample is run through the deobfuscator; the input (obfuscated) bytes, the output
+(deobfuscated) bytes, and readability scores are published in `metrics.jsonl`.
+
+👉 **Use the Dataset Viewer above** to browse all 9,514 samples — filter by
+`source`, sort by `kept_pct`/`opaque_out`, and read each sample's `input` next
+to its `output`.
+
+## What's measured
+
+| column | meaning |
+|---|---|
+| `path` | sample path in this repo |
+| `source` | technique / origin (e.g. `generated/strings_rc4`, `real/tranco`) |
+| `in_bytes` / `out_bytes` | obfuscated input vs. deobfuscated output size |
+| `kept_pct` | `out_bytes / in_bytes` — lower means more dead weight removed (on real *minified* input this exceeds 100% because the engine reformats/expands it; readability, not size, is the win there) |
+| `opaque_in` → `opaque_out` | % of identifiers that look machine-generated (hex `_0x…` or 1–2 chars) before vs. after — **the headline readability metric** |
+| `input` / `output` | the actual obfuscated and deobfuscated source |
+
+## Scoreboard
+
+Aggregate over the whole corpus, by source. Generated with
+`cargo run --release --bin report -- --markdown` in `devirt-core`.
+
+| source | files | in bytes | out bytes | kept% | opaque% in→out |
+|---|--:|--:|--:|--:|--:|
+| `generated/cf_dead` | 500 | 754109 | 61635 | 8% | 51.6% → 3.9% |
+| `generated/controlflow` | 500 | 762150 | 63133 | 8% | 51.5% → 4.2% |
+| `generated/controlflow_half` | 500 | 705909 | 63089 | 8% | 51.3% → 4.2% |
+| `generated/deadcode` | 500 | 637045 | 61819 | 9% | 50.9% → 4.0% |
+| `generated/minimal` | 500 | 638935 | 60273 | 9% | 50.9% → 3.7% |
+| `generated/names_hex` | 500 | 642247 | 64967 | 10% | 52.6% → 11.3% |
+| `generated/names_mangled` | 500 | 406652 | 58959 | 14% | 52.7% → 11.8% |
+| `generated/numbers` | 500 | 838299 | 61909 | 7% | 39.3% → 4.0% |
+| `generated/objectkeys` | 500 | 641608 | 61241 | 9% | 51.0% → 4.0% |
+| `generated/sa_cf` | 500 | 2019731 | 61280 | 3% | 60.0% → 3.7% |
+| `generated/sa_numbers` | 500 | 1519510 | 61188 | 4% | 44.3% → 3.7% |
+| `generated/splitstrings` | 500 | 646336 | 60023 | 9% | 51.0% → 3.7% |
+| `generated/strings_b64` | 500 | 1164235 | 61346 | 5% | 58.2% → 3.7% |
+| `generated/strings_indexshift` | 500 | 1894140 | 61255 | 3% | 60.8% → 3.7% |
+| `generated/strings_none` | 500 | 647016 | 59997 | 9% | 50.9% → 3.7% |
+| `generated/strings_rc4` | 500 | 1876836 | 61506 | 3% | 60.4% → 3.7% |
+| `generated/strings_rotate` | 500 | 1164235 | 61346 | 5% | 58.2% → 3.7% |
+| `generated/strings_wrappers` | 500 | 2098975 | 61665 | 2% | 61.0% → 3.6% |
+| `generated/strong` | 500 | 2604260 | 108691 | 4% | 48.4% → 3.7% |
+| `real/tranco` | 14 | 601126 | 982004 | 163% | 49.6% → 7.4% |
+| **TOTAL** | **9514** | **22263354** | **2197326** | **9%** | **52.9% → 4.6%** |
+
+**Headline:** across 9,514 samples the engine cuts machine-generated identifiers
+from **52.9% to 4.6%**. The `names_hex` / `names_mangled` rows (~11%) are the
+honest weak spot — pure identifier renaming can't be reversed to semantic names.
+
+## Layout
+
+- `generated/<profile>/<seed>.js` — 500 deterministic seeds × 19
+  [javascript-obfuscator](https://github.com/javascript-obfuscator/javascript-obfuscator)
+  profiles.
+- `real/<source>/…` — real-world scripts (crawled top sites; npm bundles; HTTP
+  Archive), filtered + deduped.
+- `seeds/` — the plain source programs the generated set is built from.
+- `metrics/*.jsonl` — the per-sample table above; powers the Dataset Viewer
+  (sharded to stay under HF's 10 MB-per-file limit).
+
+Regenerate the scores from scratch:
+
+```sh
+git clone https://github.com/devirt-dev/devirt-core && cd devirt-core
+scripts/pull-corpus.sh
+cargo run --release --bin report -- --json samples/metrics.jsonl --markdown
+# shard for HF (>10 MB files need LFS otherwise):
+mkdir -p samples/metrics && split -C 9000000 -d --additional-suffix=.jsonl \
+  samples/metrics.jsonl samples/metrics/part- && rm samples/metrics.jsonl
+```