README.md

---
license: mit
language:
- en
pretty_name: Indus Script Corpora (indus_website and CISI)
tags:
- indus-script
- indus-valley
- undeciphered-script
- archaeology
- historical-linguistics
- computational-epigraphy
size_categories:
- 1K<n<10K
task_categories:
- text-classification
- other
configs:
- config_name: indus_website
  data_files: indus_website_real_corpus.csv
- config_name: cisi_primary
  data_files: cisi_real_corpus.csv
- config_name: cisi_hierarchical
  data_files: cisi_real_corpus_hierarchical.csv
- config_name: cisi_allograph
  data_files: cisi_real_corpus_allograph.csv
---

# Indus Script Corpora (indus_website and CISI)

## Dataset Description

Two independently sourced Indus Valley script corpora, packaged together
as four configs of one dataset since they share a schema and are
routinely used side by side for comparison. Both are derived, cleaned
CSVs built for computational analysis, not new digitization efforts.
All credit for the underlying transcription work belongs to the two
source repositories named below; this card documents what was changed,
checked, and fixed in producing these versions, so downstream users
don't have to reverse-engineer that from code.

**This dataset does not decipher or interpret the Indus script.** It is
raw structural data (which signs appear, in what order, at which site)
for anyone doing their own statistical or computational analysis.

### The four configs

| Config | Corpus | N (usable) | Signs | Unit |
|---|---|---|---|---|
| `indus_website` | 2,543 inscriptions, 52 sites | 2,543 | 592 | inscription |
| `cisi_primary` | 179 Mohenjo-daro unicorn seals (104 non-damaged) | 179 | 182 (142 non-damaged) | inscription |
| `cisi_hierarchical` | same 179 seals, allographs split at 3+ attestations | 179 | 234 (185 non-damaged) | inscription |
| `cisi_allograph` | same 179 seals, every allograph a distinct sign | 179 | 337 (230 non-damaged) | inscription |

Load a specific config with, e.g.:
```python
load_dataset("<repo>", "cisi_hierarchical")
```

**Do not directly compare raw statistics (entropy, vocabulary size,
uniqueness) between `indus_website` and any `cisi_*` config without
adjusting for scale.** Their vocabularies differ by roughly 3 to 4
times (142 to 337 signs vs. 592), and such statistics are strongly
vocabulary-size-dependent. The three `cisi_*` configs are directly
comparable to each other (same 179 underlying inscriptions), and are
specifically designed to be compared to each other; see "Why three CISI
granularities" below.

### Companion resources

Built as part of
[joyboseroy/indus_decipher](https://github.com/joyboseroy/indus_decipher),
a falsification-oriented statistical toolkit for the Indus script. That
repository's `CORPUS_REGISTRY.md` and `CLAIM_EVIDENCE_MATRIX.md`
document, in much more depth than this card, every real corpus this
project uses (including a fifth corpus, the actual M77/IDF-80 corpus,
not distributed here due to its more involved access terms), the
specific bugs found and fixed while working with each one, and exactly
what has and has not been established using them.

## Dataset Structure

### Data Instances

`indus_website`:
```
inscription_id,sign_sequence,site,object_type,line_count,damaged,reading_direction,motif,mean_uncertainty,cisi_number
S2531,G140 G8 G342,Mohenjo-daro,seal,1,False,R-L,Bull1:W,0.0,M-1734
```

`cisi_primary` / `cisi_hierarchical` / `cisi_allograph` (same schema,
sign IDs differ in granularity):
```
inscription_id,sign_sequence,site,object_type,line_count,damaged,reading_direction,motif,mean_uncertainty
M-1A,P121 P202 P385 P073 P108,Mohenjo-daro,seal,1,False,L-R,unicorn_I,0.0
```

### Data Fields

All four configs share one schema:

| Field | Type | Description |
|---|---|---|
| `inscription_id` | string | Real ID from the config's own source (database key for `indus_website`; real CISI ID, e.g. `"M-1A"`, for the three `cisi_*` configs) |
| `sign_sequence` | string | Space-separated sign IDs, left-to-right as stored (see `reading_direction`) |
| `site` | string | Excavation site. 52 distinct sites in `indus_website`; always `"Mohenjo-daro"` in the `cisi_*` configs |
| `object_type` | string | Seal, tablet, etc. where recorded; `"unknown"` otherwise |
| `line_count` | int | Number of physical lines on the object |
| `damaged` | bool | `indus_website`: a per-inscription flag from the source. `cisi_*`: true if ANY grapheme in the inscription has a nonzero damage code |
| `reading_direction` | string | `"R-L"` means use `sign_sequence` as stored; `"L-R"` means reverse it before use. A loader-convention flag, not a claim about the script's actual reading direction. `indus_website` has a genuine mix of both values (per-inscription, source-provided); all `cisi_*` rows are `"L-R"` |
| `motif` | string | `indus_website`: iconographic field-symbol code (e.g. `"Bull1:W"`), `"unknown"` for 921 of 2,543 rows. `cisi_*`: always a unicorn variant (e.g. `"unicorn_I"`) |
| `mean_uncertainty` | float | `indus_website`: always `0.0` (no per-grapheme uncertainty field in that source). `cisi_*`: mean, across an inscription's graphemes, of the original annotator's 0-100 subjective uncertainty score, distinct from `damaged` |
| `cisi_number` | string | `indus_website` only. Real Mahadevan/CISI catalog number (e.g. `"M-1734"`, `"H-322"`) for 2,375 of 2,543 rows (93%); `"unknown"` otherwise. Not applicable to the `cisi_*` configs, whose `inscription_id` already is the real CISI number |

### Data Splits

No splits in any config.

## Dataset Creation

### Source Data

`indus_website`: MySQL dump (`population-script.sql`) from
[yajnadevam/indus-website](https://github.com/yajnadevam/indus-website).
Scale matches the "ICIT/Yajnadevam digitization" cited in independent
literature on Indus sign-system statistics.

`cisi_*`: JSON files from
[mayig/indus-valley-script-corpus](https://github.com/mayig/indus-valley-script-corpus)
(MIT licensed), an in-progress, transparently documented
hand-transcription of Parpola's CISI corpus, credited to Parpola,
Wells, and Fuls's sign lists.

### Corrections made during conversion

- **Reading direction (`indus_website`).** The original conversion had
  sign order backwards. Caught by a diagnostic checking for a published
  fingerprint (final sign position should show lower entropy than
  initial position); confirmed correct after reversal.
- **CISI/Mahadevan numbers (`indus_website`).** The source SQL has a
  `CISI` column that an earlier version of the conversion script never
  extracted. Fixed: 93% of rows now carry this real crosswalk number,
  spot-checked against physical seal photographs for consistent sign
  counts.
- **Reading direction (`cisi_*`).** Same kind of fix, independently
  confirmed for this separate source.

### Why three CISI granularities

The `cisi_hierarchical` config exists to distinguish two competing
explanations for a result found using `cisi_allograph`: does allograph
choice genuinely matter, or is naive full fragmentation just too sparse
on 104 inscriptions to trust? `cisi_hierarchical`'s 3-occurrence
threshold was built specifically to test this. It did not resolve to
"it's just sparsity," which is why all three configs are provided
rather than one being treated as the "real" version.

## Considerations for Using the Data

- **The three `cisi_*` configs give different answers to the same
  structural question, and that is a documented finding, not an
  inconsistency to average away.** Using a basic language-vs-non-language
  statistical classifier (see companion repo), `cisi_primary` classifies
  this corpus one way; `cisi_hierarchical` and `cisi_allograph` both
  classify it the other way, and this is not simply because the latter
  two are too sparse (`cisi_hierarchical` was built to be conservative
  and still agrees with the fully fragmented version). Tracing why found
  that the single most common sign at primary granularity is actually
  several independently well-attested, visually distinct forms bundled
  under one ID. Report all three when it matters, don't pick the one
  that supports your preferred conclusion.
- **`cisi_*` is small (104 non-damaged inscriptions) and entirely one
  motif at one site.** Statistics requiring higher-order sequential
  structure did not produce a reliable signal on it in the companion
  repository's own testing, most likely due to sample size. It is also
  not a representative cross-section of the Indus corpus as a whole.
- **`motif` and `cisi_number` are both partial in `indus_website`.**
  Treat `"unknown"` as missing data, not a meaningful third category.
- This dataset alone cannot support a claim that the Indus script is or
  is not a language, an administrative code, or any other specific kind
  of system. See the companion repository's synthetic-continuum and
  external-language-calibration experiments for why that inference is
  harder than it looks.
- **`indus_website`'s source repository maintainer is also associated
  with an independently disputed decipherment claim for an unrelated
  script.** The raw glyph sequences and site metadata here are ordinary
  transcription data, checked directly (reading-direction fingerprint,
  CISI-number cross-reference) rather than taken on trust; no
  sign-to-meaning mapping from that other work is used anywhere here.

## Additional Information

### Licensing

`indus_website`: source repository license terms apply to the
underlying transcription data; this derived CSV is provided under MIT
for the conversion and structuring work. `cisi_*`: source repository is
MIT licensed; this derived CSV is provided under the same terms.

### Citation

If you use this dataset, please cite the relevant original source:

```
indus_website config: yajnadevam/indus-website (GitHub), underlying ICIT/Yajnadevam digitization
cisi_* configs: mayig/indus-valley-script-corpus (GitHub), digitization of Parpola's CISI corpus
```

and, if relevant, the companion analysis toolkit:

```
joyboseroy/indus_decipher (GitHub)
```