v2: revert value-altering substitutions from generation

12,120 rows (3.72%), 13,754 characters. A substitution is reverted only where the numeral still parses as a number and that number is different: a digit reading as a different digit (10,650), the currency marker reading as a digit (2,006), a decimal point vanishing or reading as a comma (1,026), a thousands separator reading as a decimal point (72). Row count, row order and every label are unchanged. Letter obfuscation is untouched. Adds a research-only LICENSE and the full change log under corrections/.

LICENSE

@@ -0,0 +1,43 @@
+BitAbuse / BitViper / BitCore — Research Use Only
+
+Copyright (c) 2025 Hanyong Lee, Chaelyn Lee, Yongjae Lee, Jaesung Lee
+
+These datasets accompany:
+
+    Hanyong Lee, Chaelyn Lee, Yongjae Lee, and Jaesung Lee. 2025.
+    BitAbuse: A Dataset of Visually Perturbed Texts for Defending Phishing Attacks.
+    Findings of the Association for Computational Linguistics: NAACL 2025,
+    pages 4367-4384. https://doi.org/10.18653/v1/2025.findings-naacl.247
+
+TERMS
+
+1. Research use only. As stated in the paper's Limitations and Ethics Statement:
+   "The datasets and models used in this paper are publicly available, but they
+   should not be used for purposes other than research." No other use is
+   permitted. In particular, no use may conduct, assist, or improve phishing,
+   fraud, or other attacks.
+
+2. Attribution. Cite the paper above in any work that uses these datasets.
+
+3. No redistribution. Do not mirror or republish these datasets elsewhere. Link
+   to the Hugging Face repositories instead, so that users receive the current
+   revision rather than a stale copy. Quoting individual examples in a paper or
+   report is not redistribution.
+
+4. Upstream data. The corpus derives from phishing reports collected from the
+   Bitcoin Abuse Database (https://www.bitcoinabuse.com, accessed 2023-04-30),
+   which was acquired by TRM Labs in October 2023 and folded into Chainabuse.
+   Nothing here grants rights over that upstream material, and users are
+   responsible for their own compliance with its terms.
+
+5. Content and identifiers. The texts are real phishing, sextortion and scam
+   messages, and contain abusive and sexually coercive language. They also carry
+   identifiers as they appeared in the original reports: bitcoin wallet
+   addresses in payment demands, email addresses, URLs, and residual mail
+   headers. Do not treat any of it as safe to reuse, and do not attempt to
+   contact, identify, or authenticate against anyone or anything named in the
+   data.
+
+6. No warranty. The datasets are provided "as is", without warranty of any kind,
+   express or implied. The authors accept no liability for any claim, damage or
+   other liability arising from their use.

README.md

@@ -1,4 +1,7 @@
 ---
+license: other
+license_name: research-only
+license_link: LICENSE
 dataset_info:
   features:
   - name: id
@@ -9,13 +12,170 @@ dataset_info:
     dtype: string
   splits:
   - name: train
-    num_bytes: 72259353
+    num_bytes: 72254030
     num_examples: 325580
-  download_size: 45871484
-  dataset_size: 72259353
+  download_size: 45679005
+  dataset_size: 72254030
 configs:
 - config_name: default
   data_files:
   - split: train
     path: data/train-*
+size_categories:
+- 100K<n<1M
 ---
+
+# BitAbuse
+
+Homoglyph-obfuscated abusive messages paired with their originals.
+`text` is the obfuscated form, `label` the original.
+
+BitAbuse is the union of two parts:
+[BitCore](https://huggingface.co/datasets/AutoML/bitcore), restored by hand, and
+[BitViper](https://huggingface.co/datasets/AutoML/bitviper), generated.
+
+```text
+bitabuse (325,580)  =  bitcore (26,591)  +  bitviper (298,989)
+```
+
+## Revisions
+
+| revision | rows | note |
+|---|---:|---|
+| `main` | 325,580 | original release, 2024-08-16 |
+| `v2` | 325,580 | value-altering substitutions reverted, 2026-08 |
+
+`main` is unchanged. Loading without a revision keeps the original behaviour.
+
+```python
+load_dataset("AutoML/bitabuse", revision="v2")
+```
+
+## What v2 changes
+
+Generation writes `text` from `label` by drawing, for each character, one of about
+twenty randomised look-alike targets. For most characters that is the point: the pair
+stays the same message, obfuscated. But a **numeral** can be drawn onto something that
+still parses as a numeral, and then the pair stops being an obfuscation and becomes two
+different claims.
+
+```text
+label   same old bs as usual for the last 3.5 months.
+main    sảme oⅠd ƅs aş usuaӀ for the lɵsț 3͒5 ӕoņths.     <- reads as 35 months
+v2      sảme oⅠd ƅs aş usuaӀ for the lɵsț 3.5 ӕoņths.
+```
+
+The decimal point was drawn onto U+0352 COMBINING FERMATA, which attaches to the
+preceding character and disappears. Neither a reader nor a parser can notice.
+
+**The rule: revert a substitution when the numeral still parses as a number and that
+number is different. Nothing else.**
+
+| class | occurrences | example |
+|---|---:|---|
+| a digit reads as a different digit | 10,650 | `100%` → `109%` |
+| the currency marker reads as a digit | 2,006 | `740$` → `7409` |
+| a decimal point vanishes or reads as a comma | 1,026 | `3.5` → `3͒5`, `12.000` → `12،000` |
+| a thousands separator reads as a decimal point | 72 | `$9,524` → `$9․524` |
+
+**12,120 rows (3.72%), 13,754 characters.** Row count, row order and every `label` are
+unchanged, so indices into `main` remain valid.
+
+Every affected row is in the BitViper part. BitCore, whose pairs did not go through
+generation, carries none of them. The same correction is published as
+[`AutoML/bitviper`](https://huggingface.co/datasets/AutoML/bitviper) revision `v2`;
+BitCore is unaffected and has no v2.
+
+## What v2 does not change
+
+The test is whether the numeral still parses, not which codepoint was drawn.
+
+```text
+0 → u          kept       does not parse as a number; visibly corrupt
+0 → ս          kept       likewise, and identical in kind to 0 → u
+0 → 9          reverted   parses, and the value changed
+2 → ߄          reverted   NKO DIGIT FOUR parses as four
+0 → ߀          kept       NKO DIGIT ZERO parses as zero, the same value
+5,000 → 5‚000  kept       a low-9 quote still reads as a comma; the same number
+1,350 → 1350   kept       a separator that vanishes costs no value
+hello, world → hello. world    kept    not inside a numeral at all
+```
+
+**Letter substitutions are never reverted.** They are the obfuscation the dataset exists
+to study.
+
+```text
+thinking → fhļnking     visited → vjsited     know → ķhow
+```
+
+`main` carries 5,000,593 substitutions, of which 4,759,927 replace a letter. All of them
+stay, along with punctuation obfuscation outside numerals.
+
+## What v2 still does not catch
+
+The correction is deliberately narrow, and these remain in `v2`:
+
+- **Currency swaps. 700 characters** where `$` was drawn onto `฿`, `₮` or `₵`. The number
+  survives and the currency changes. A reader sees a different symbol rather than a
+  different number, so these are treated as obfuscation.
+- **Numerals broken beyond reading.** Where a digit was drawn onto a letter or a symbol,
+  the amount is unreadable rather than wrong, and v2 leaves it. `$500` → `ȿ5ȣ0` stays.
+- **Wallet and bitcoin addresses**, which are obfuscated throughout in both revisions.
+  Do not treat any address in `text` as valid in either revision.
+
+If your task needs a guaranteed-clean numeral, use `label`.
+
+## Known effect
+
+**27 rows now have `text == label`.** Every substitution in those rows was a
+value-altering one, so reverting them leaves the row uncorrupted. No row in `main` had
+`text == label`. They are kept rather than dropped so that row indices stay aligned
+with `main`. Filter them out if your task requires every row to be obfuscated.
+
+```python
+ds = ds.filter(lambda r: r["text"] != r["label"])
+```
+
+## License and citation
+
+Research use only. See `LICENSE`, and the paper's Ethics Statement: "The datasets
+and models used in this paper are publicly available, but they should not be used
+for purposes other than research."
+
+The corpus derives from phishing reports collected from the Bitcoin Abuse Database.
+Nothing here grants rights over that upstream material.
+
+**Content warning.** These are real phishing, sextortion and scam messages. They
+contain abusive and sexually coercive language, and carry identifiers from the
+original reports — wallet addresses, email addresses, URLs and residual mail
+headers. Do not attempt to contact, identify, or authenticate against anything
+named in the data.
+
+Redistribution is not permitted; link here rather than mirroring. Quoting
+individual examples in a paper is not redistribution.
+
+```bibtex
+@inproceedings{lee-etal-2025-bitabuse,
+    title = "{B}it{A}buse: A Dataset of Visually Perturbed Texts for Defending Phishing Attacks",
+    author = "Lee, Hanyong  and Lee, Chaelyn  and Lee, Yongjae  and Lee, Jaesung",
+    editor = "Chiruzzo, Luis  and Ritter, Alan  and Wang, Lu",
+    booktitle = "Findings of the Association for Computational Linguistics: NAACL 2025",
+    month = apr,
+    year = "2025",
+    address = "Albuquerque, New Mexico",
+    publisher = "Association for Computational Linguistics",
+    url = "https://aclanthology.org/2025.findings-naacl.247/",
+    doi = "10.18653/v1/2025.findings-naacl.247",
+    pages = "4367--4384",
+    ISBN = "979-8-89176-195-7",
+}
+```
+
+## Reproducing
+
+`corrections/v2_changes.csv` lists every reverted character with its row id and position.
+The correction is deterministic and derives entirely from `main`: for each position
+where `label` holds a digit, `$`, or a separator between two digits, the drawn character
+is classified as reading like a numeral, a decimal point, a comma, nothing at all
+(combining marks and format characters), or visible damage; the position is reverted only
+when the numeral still parses and its value differs.

data/train-00000-of-00001.parquet

@@ -1,3 +1,3 @@
 version https://git-lfs.github.com/spec/v1
-oid sha256:1a3649e8ada7af83154fcc3ace8621c8dd3d6af81933e0e2c6a27cbfad02f15f
-size 45871484
+oid sha256:eb50c9670fcf81a2a0c48f0820bde0ef9c6d61e110abb7d9241ac5ea7891c159
+size 45679005