Dataset Viewer
Auto-converted to Parquet Duplicate
null_id
stringlengths
23
49
corpus
stringclasses
2 values
null_model
stringclasses
8 values
seed
int64
0
2
signs
listlengths
1
25
length
int64
1
25
cisi-within_inscription_shuffle-s0-0
cisi
within_inscription_shuffle
0
[ "P094", "P320", "P145" ]
3
cisi-within_inscription_shuffle-s0-1
cisi
within_inscription_shuffle
0
[ "P378", "P384", "P201", "P065" ]
4
cisi-within_inscription_shuffle-s0-2
cisi
within_inscription_shuffle
0
[ "P346", "P026", "P151", "P378" ]
4
cisi-within_inscription_shuffle-s0-3
cisi
within_inscription_shuffle
0
[ "P341", "P142", "P324" ]
3
cisi-within_inscription_shuffle-s0-4
cisi
within_inscription_shuffle
0
[ "P324", "P272", "P256", "P121", "P385", "P122", "P091", "P145" ]
8
cisi-within_inscription_shuffle-s0-5
cisi
within_inscription_shuffle
0
[ "P210", "P127", "P060", "P324", "P332", "P364", "P062" ]
7
cisi-within_inscription_shuffle-s0-6
cisi
within_inscription_shuffle
0
[ "P385", "P086", "P122", "P123" ]
4
cisi-within_inscription_shuffle-s0-7
cisi
within_inscription_shuffle
0
[ "P103", "P086", "P324", "P122", "P108", "P309", "P194" ]
7
cisi-within_inscription_shuffle-s0-8
cisi
within_inscription_shuffle
0
[ "P050", "P217", "P092" ]
3
cisi-within_inscription_shuffle-s0-9
cisi
within_inscription_shuffle
0
[ "P122", "P301", "P062", "P230", "P385" ]
5
cisi-within_inscription_shuffle-s0-10
cisi
within_inscription_shuffle
0
[ "P060", "P353", "P122", "P062", "P201", "P050", "P230", "P324" ]
8
cisi-within_inscription_shuffle-s0-11
cisi
within_inscription_shuffle
0
[ "P301", "P230", "P230", "P120" ]
4
cisi-within_inscription_shuffle-s0-12
cisi
within_inscription_shuffle
0
[ "P147", "P120", "P124", "P316", "P217", "P050" ]
6
cisi-within_inscription_shuffle-s0-13
cisi
within_inscription_shuffle
0
[ "P324", "P260", "P013", "P160", "P047", "P120", "P326", "P194" ]
8
cisi-within_inscription_shuffle-s0-14
cisi
within_inscription_shuffle
0
[ "P098", "P181", "P217", "P000" ]
4
cisi-within_inscription_shuffle-s0-15
cisi
within_inscription_shuffle
0
[ "P385", "P050", "P124", "P122", "P217" ]
5
cisi-within_inscription_shuffle-s0-16
cisi
within_inscription_shuffle
0
[ "P268", "P230", "P268", "P301", "P062", "P217", "P122" ]
7
cisi-within_inscription_shuffle-s0-17
cisi
within_inscription_shuffle
0
[ "P122", "P324", "P086", "P053", "P230", "P082" ]
6
cisi-within_inscription_shuffle-s0-18
cisi
within_inscription_shuffle
0
[ "P180", "P128", "P086" ]
3
cisi-within_inscription_shuffle-s0-19
cisi
within_inscription_shuffle
0
[ "P251", "P325", "P217", "P070", "P058" ]
5
cisi-within_inscription_shuffle-s0-20
cisi
within_inscription_shuffle
0
[ "P313", "P324", "P358", "P060", "P145" ]
5
cisi-within_inscription_shuffle-s0-21
cisi
within_inscription_shuffle
0
[ "P095", "P230", "P385", "P122", "P050", "P073", "P009", "P145", "P324" ]
9
cisi-within_inscription_shuffle-s0-22
cisi
within_inscription_shuffle
0
[ "P215", "P075", "P369", "P253", "P122", "P324", "P091" ]
7
cisi-within_inscription_shuffle-s0-23
cisi
within_inscription_shuffle
0
[ "P050", "P324", "P123", "P276", "P086" ]
5
cisi-within_inscription_shuffle-s0-24
cisi
within_inscription_shuffle
0
[ "P004", "P154", "P123", "P226" ]
4
cisi-within_inscription_shuffle-s0-25
cisi
within_inscription_shuffle
0
[ "P058", "P238", "P324", "P122", "P054", "P071" ]
6
cisi-within_inscription_shuffle-s0-26
cisi
within_inscription_shuffle
0
[ "P000", "P385", "P122" ]
3
cisi-within_inscription_shuffle-s0-27
cisi
within_inscription_shuffle
0
[ "P349", "P325", "P324", "P145", "P120", "P156" ]
6
cisi-within_inscription_shuffle-s0-28
cisi
within_inscription_shuffle
0
[ "P283", "P332", "P122", "P205", "P324", "P095" ]
6
cisi-within_inscription_shuffle-s0-29
cisi
within_inscription_shuffle
0
[ "P205", "P364", "P011", "P194", "P086", "P122", "P076", "P352" ]
8
cisi-within_inscription_shuffle-s0-30
cisi
within_inscription_shuffle
0
[ "P325", "P288", "P009", "P044", "P324", "P349" ]
6
cisi-within_inscription_shuffle-s0-31
cisi
within_inscription_shuffle
0
[ "P230", "P120", "P234", "P000", "P256" ]
5
cisi-within_inscription_shuffle-s0-32
cisi
within_inscription_shuffle
0
[ "P123", "P226", "P004" ]
3
cisi-within_inscription_shuffle-s0-33
cisi
within_inscription_shuffle
0
[ "P301", "P073", "P062", "P369", "P230", "P174", "P110", "P050" ]
8
cisi-within_inscription_shuffle-s0-34
cisi
within_inscription_shuffle
0
[ "P145", "P390", "P256" ]
3
cisi-within_inscription_shuffle-s0-35
cisi
within_inscription_shuffle
0
[ "P325", "P061", "P294", "P368", "P147" ]
5
cisi-within_inscription_shuffle-s0-36
cisi
within_inscription_shuffle
0
[ "P221", "P324", "P145", "P201", "P283", "P050", "P309" ]
7
cisi-within_inscription_shuffle-s0-37
cisi
within_inscription_shuffle
0
[ "P022", "P385", "P011", "P122" ]
4
cisi-within_inscription_shuffle-s0-38
cisi
within_inscription_shuffle
0
[ "P324", "P145", "P122", "P082", "P170", "P349" ]
6
cisi-within_inscription_shuffle-s0-39
cisi
within_inscription_shuffle
0
[ "P122", "P324", "P062", "P092", "P009", "P385" ]
6
cisi-within_inscription_shuffle-s0-40
cisi
within_inscription_shuffle
0
[ "P324", "P098", "P031", "P384" ]
4
cisi-within_inscription_shuffle-s0-41
cisi
within_inscription_shuffle
0
[ "P368", "P324" ]
2
cisi-within_inscription_shuffle-s0-42
cisi
within_inscription_shuffle
0
[ "P393", "P201", "P050" ]
3
cisi-within_inscription_shuffle-s0-43
cisi
within_inscription_shuffle
0
[ "P385", "P122", "P096", "P324" ]
4
cisi-within_inscription_shuffle-s0-44
cisi
within_inscription_shuffle
0
[ "P086", "P230", "P126", "P120" ]
4
cisi-within_inscription_shuffle-s0-45
cisi
within_inscription_shuffle
0
[ "P325", "P324", "P166", "P290", "P136" ]
5
cisi-within_inscription_shuffle-s0-46
cisi
within_inscription_shuffle
0
[ "P324", "P349", "P123", "P234", "P230", "P332" ]
6
cisi-within_inscription_shuffle-s0-47
cisi
within_inscription_shuffle
0
[ "P385", "P122", "P031", "P324", "P204" ]
5
cisi-within_inscription_shuffle-s0-48
cisi
within_inscription_shuffle
0
[ "P364", "P110", "P160", "P303", "P324" ]
5
cisi-within_inscription_shuffle-s0-49
cisi
within_inscription_shuffle
0
[ "P255", "P000" ]
2
cisi-within_inscription_shuffle-s0-50
cisi
within_inscription_shuffle
0
[ "P122", "P385", "P123", "P086" ]
4
cisi-within_inscription_shuffle-s0-51
cisi
within_inscription_shuffle
0
[ "P086", "P126" ]
2
cisi-within_inscription_shuffle-s0-52
cisi
within_inscription_shuffle
0
[ "P205", "P217", "P058", "P120" ]
4
cisi-within_inscription_shuffle-s0-53
cisi
within_inscription_shuffle
0
[ "P013", "P324", "P359", "P010" ]
4
cisi-within_inscription_shuffle-s1-0
cisi
within_inscription_shuffle
1
[ "P145", "P320", "P094" ]
3
cisi-within_inscription_shuffle-s1-1
cisi
within_inscription_shuffle
1
[ "P378", "P201", "P065", "P384" ]
4
cisi-within_inscription_shuffle-s1-2
cisi
within_inscription_shuffle
1
[ "P346", "P026", "P151", "P378" ]
4
cisi-within_inscription_shuffle-s1-3
cisi
within_inscription_shuffle
1
[ "P324", "P341", "P142" ]
3
cisi-within_inscription_shuffle-s1-4
cisi
within_inscription_shuffle
1
[ "P272", "P122", "P145", "P121", "P324", "P256", "P385", "P091" ]
8
cisi-within_inscription_shuffle-s1-5
cisi
within_inscription_shuffle
1
[ "P324", "P210", "P127", "P364", "P060", "P332", "P062" ]
7
cisi-within_inscription_shuffle-s1-6
cisi
within_inscription_shuffle
1
[ "P122", "P123", "P086", "P385" ]
4
cisi-within_inscription_shuffle-s1-7
cisi
within_inscription_shuffle
1
[ "P194", "P103", "P086", "P324", "P108", "P309", "P122" ]
7
cisi-within_inscription_shuffle-s1-8
cisi
within_inscription_shuffle
1
[ "P050", "P092", "P217" ]
3
cisi-within_inscription_shuffle-s1-9
cisi
within_inscription_shuffle
1
[ "P122", "P385", "P062", "P301", "P230" ]
5
cisi-within_inscription_shuffle-s1-10
cisi
within_inscription_shuffle
1
[ "P050", "P353", "P060", "P122", "P230", "P324", "P201", "P062" ]
8
cisi-within_inscription_shuffle-s1-11
cisi
within_inscription_shuffle
1
[ "P120", "P230", "P230", "P301" ]
4
cisi-within_inscription_shuffle-s1-12
cisi
within_inscription_shuffle
1
[ "P316", "P217", "P147", "P124", "P120", "P050" ]
6
cisi-within_inscription_shuffle-s1-13
cisi
within_inscription_shuffle
1
[ "P047", "P120", "P013", "P160", "P260", "P194", "P326", "P324" ]
8
cisi-within_inscription_shuffle-s1-14
cisi
within_inscription_shuffle
1
[ "P181", "P000", "P217", "P098" ]
4
cisi-within_inscription_shuffle-s1-15
cisi
within_inscription_shuffle
1
[ "P385", "P217", "P124", "P122", "P050" ]
5
cisi-within_inscription_shuffle-s1-16
cisi
within_inscription_shuffle
1
[ "P217", "P062", "P301", "P268", "P268", "P230", "P122" ]
7
cisi-within_inscription_shuffle-s1-17
cisi
within_inscription_shuffle
1
[ "P082", "P230", "P053", "P324", "P122", "P086" ]
6
cisi-within_inscription_shuffle-s1-18
cisi
within_inscription_shuffle
1
[ "P180", "P086", "P128" ]
3
cisi-within_inscription_shuffle-s1-19
cisi
within_inscription_shuffle
1
[ "P325", "P251", "P217", "P058", "P070" ]
5
cisi-within_inscription_shuffle-s1-20
cisi
within_inscription_shuffle
1
[ "P145", "P060", "P313", "P358", "P324" ]
5
cisi-within_inscription_shuffle-s1-21
cisi
within_inscription_shuffle
1
[ "P230", "P324", "P009", "P050", "P145", "P073", "P095", "P122", "P385" ]
9
cisi-within_inscription_shuffle-s1-22
cisi
within_inscription_shuffle
1
[ "P075", "P253", "P369", "P091", "P122", "P215", "P324" ]
7
cisi-within_inscription_shuffle-s1-23
cisi
within_inscription_shuffle
1
[ "P324", "P123", "P276", "P050", "P086" ]
5
cisi-within_inscription_shuffle-s1-24
cisi
within_inscription_shuffle
1
[ "P123", "P154", "P226", "P004" ]
4
cisi-within_inscription_shuffle-s1-25
cisi
within_inscription_shuffle
1
[ "P071", "P058", "P324", "P054", "P238", "P122" ]
6
cisi-within_inscription_shuffle-s1-26
cisi
within_inscription_shuffle
1
[ "P385", "P000", "P122" ]
3
cisi-within_inscription_shuffle-s1-27
cisi
within_inscription_shuffle
1
[ "P324", "P120", "P145", "P156", "P325", "P349" ]
6
cisi-within_inscription_shuffle-s1-28
cisi
within_inscription_shuffle
1
[ "P205", "P122", "P324", "P095", "P332", "P283" ]
6
cisi-within_inscription_shuffle-s1-29
cisi
within_inscription_shuffle
1
[ "P205", "P364", "P011", "P086", "P194", "P076", "P352", "P122" ]
8
cisi-within_inscription_shuffle-s1-30
cisi
within_inscription_shuffle
1
[ "P325", "P349", "P009", "P288", "P324", "P044" ]
6
cisi-within_inscription_shuffle-s1-31
cisi
within_inscription_shuffle
1
[ "P000", "P230", "P256", "P234", "P120" ]
5
cisi-within_inscription_shuffle-s1-32
cisi
within_inscription_shuffle
1
[ "P226", "P004", "P123" ]
3
cisi-within_inscription_shuffle-s1-33
cisi
within_inscription_shuffle
1
[ "P110", "P073", "P050", "P062", "P369", "P230", "P301", "P174" ]
8
cisi-within_inscription_shuffle-s1-34
cisi
within_inscription_shuffle
1
[ "P256", "P145", "P390" ]
3
cisi-within_inscription_shuffle-s1-35
cisi
within_inscription_shuffle
1
[ "P294", "P147", "P325", "P061", "P368" ]
5
cisi-within_inscription_shuffle-s1-36
cisi
within_inscription_shuffle
1
[ "P309", "P145", "P283", "P050", "P201", "P324", "P221" ]
7
cisi-within_inscription_shuffle-s1-37
cisi
within_inscription_shuffle
1
[ "P011", "P385", "P022", "P122" ]
4
cisi-within_inscription_shuffle-s1-38
cisi
within_inscription_shuffle
1
[ "P349", "P122", "P324", "P170", "P082", "P145" ]
6
cisi-within_inscription_shuffle-s1-39
cisi
within_inscription_shuffle
1
[ "P062", "P092", "P009", "P324", "P122", "P385" ]
6
cisi-within_inscription_shuffle-s1-40
cisi
within_inscription_shuffle
1
[ "P031", "P324", "P098", "P384" ]
4
cisi-within_inscription_shuffle-s1-41
cisi
within_inscription_shuffle
1
[ "P368", "P324" ]
2
cisi-within_inscription_shuffle-s1-42
cisi
within_inscription_shuffle
1
[ "P393", "P201", "P050" ]
3
cisi-within_inscription_shuffle-s1-43
cisi
within_inscription_shuffle
1
[ "P385", "P324", "P096", "P122" ]
4
cisi-within_inscription_shuffle-s1-44
cisi
within_inscription_shuffle
1
[ "P086", "P230", "P126", "P120" ]
4
cisi-within_inscription_shuffle-s1-45
cisi
within_inscription_shuffle
1
[ "P290", "P166", "P324", "P325", "P136" ]
5
End of preview. Expand in Data Studio

Indus Script Sequences with Null Baselines

Sign sequences from two digitised Indus inscription corpora, plus matched null corpora built with several different null models. The point of the dataset is the nulls: a claim that the Indus script has "structure" or "linguistic" properties only means something relative to a stated null, and this dataset gives ready-made nulls with a fixed train/test protocol so such claims can be compared.

Everything is built by build.py from the indus_decipher repository, so the nulls can be regenerated exactly.

What this is, and what it is not

  • The real sequences are transcriptions from two existing public sources. This dataset does not add new epigraphic data, and it was not checked against the original publications. Corrections in the upstream sources are not tracked here.
  • The value added is a fixed train/test protocol and a set of reproducible null corpora, so that claims about "structure" in the script can be compared on equal terms. The nulls are synthetic and can be regenerated from the code, so they are a convenience and a reference, not irreplaceable data.
  • The reference baseline is deliberately simple. It mostly shows that shuffled sequences are easy to detect.
  • This is a niche resource for people studying the Indus script or undeciphered scripts. It makes no claim about decipherment.

Why this helps Indus script researchers

Papers on the Indus script often argue that the sign sequences show "structure" (entropy, n-gram regularities, positional preferences), and that this favours a linguistic reading. Whether that follows depends on what the structure is compared against, and each paper builds its own ad hoc comparison. Three things make those comparisons hard to check:

  1. Nulls differ between papers (shuffles, matched random sequences, Markov chains), and often the text does not say which part of the real structure each one preserves.
  2. The nulls are often fitted on the same sequences they are later compared with, which inflates how real-like they look.
  3. Few corpora are public, and each has its own sign numbering, so results are hard to rerun.

This dataset addresses the first two directly. Every null is labelled by what it preserves and destroys, and every null is fitted on the train half only, so any statistic can be computed on held-out real inscriptions against a documented set of nulls. It does not address the third beyond packaging two public corpora in one schema.

Concrete uses:

  • Test a new statistic or classifier against eight named nulls without writing generators.
  • Report a result as "separates real from bigram_markov and position_preserving_shuffle", which says much more than "differs from random".
  • Compare two published claims on identical data and splits.
  • Check whether an apparent structural finding is already produced by order-1 or order-2 dependency (the Markov nulls).
  • Teach or prototype sequence analysis on undeciphered-script data without cleaning raw files.

Quick start

import pandas as pd
real = pd.read_parquet("real.parquet")
nulls = pd.read_parquet("nulls.parquet")

c = "indus_website"
test = real[(real.corpus == c) & (real.split == "test") & (~real.damaged)]
shuf = nulls[(nulls.corpus == c) & (nulls.null_model == "position_preserving_shuffle")]
print(len(test), len(shuf), test.signs.iloc[0][:5], shuf.signs.iloc[0][:5])

Worked example: four statistics with bootstrap intervals

Script: example_stats.py. Values are on indus_website with 95% bootstrap intervals over sequences, in brackets. The bigram model and the top-5 first and last signs are fit on the real train half. "Last sign in top 5" is the share of sequences ending in one of the five most common final signs in train, and likewise for first. Nulls pool their three seeds.

sequences bigram cross-entropy (bits/token) last sign in top 5 first sign in top 5 adjacent repeat
real, train (in-sample) 4.51 [4.46, 4.57] 0.593 [0.566, 0.621] 0.215 [0.192, 0.240] 0.012 [0.006, 0.018]
real, test (held out) 5.80 [5.72, 5.89] 0.571 [0.545, 0.597] 0.182 [0.163, 0.203] 0.013 [0.007, 0.019]
bigram_markov 4.52 [4.49, 4.55] 0.589 [0.572, 0.605] 0.220 [0.206, 0.234] 0.012 [0.008, 0.015]
trigram_markov 4.51 [4.48, 4.54] 0.595 [0.579, 0.611] 0.215 [0.202, 0.229] 0.013 [0.010, 0.017]
position_preserving_shuffle 6.14 [6.09, 6.19] 0.593 [0.577, 0.609] 0.215 [0.202, 0.228] 0.014 [0.011, 0.018]
within_inscription_shuffle 7.56 [7.51, 7.60] 0.212 [0.198, 0.226] 0.071 [0.063, 0.079] 0.014 [0.010, 0.018]
global_shuffle 8.15 [8.12, 8.19] 0.188 [0.174, 0.201] 0.073 [0.065, 0.081] 0.073 [0.065, 0.081]
unigram_matched 7.21 [7.18, 7.25] 0.607 [0.592, 0.624] 0.224 [0.211, 0.238] 0.087 [0.078, 0.097]
motif_stratified 7.22 [7.19, 7.26] 0.591 [0.574, 0.607] 0.207 [0.194, 0.221] 0.094 [0.084, 0.104]
site_stratified 7.21 [7.18, 7.25] 0.601 [0.585, 0.618] 0.219 [0.206, 0.233] 0.086 [0.076, 0.095]

How to read it:

  • The nulls are built from train inscriptions, so the fair comparator is the in-sample real row. The held-out row shows what generalising costs (for example 5.80 against 4.51 bits). Comparing held-out real text to the nulls would make the Markov nulls look more real-like than they are.
  • Bigram cross-entropy. Only the Markov nulls match real (about 4.5 bits), and they do so by construction. Every shuffle and matched null is far higher (6.1 to 8.2 bits), including the shuffle that keeps each inscription's first and last sign. Order dependence inside inscriptions is not reproduced by shuffling.
  • First and last sign concentration. The nulls that keep or sample these (position-preserving, matched, stratified, Markov) match real. The two plain shuffles do not (about 0.19 to 0.21 for the last sign against 0.59). A finding that real inscriptions have concentrated opening and closing signs is therefore not evidence of anything those nulls do not already produce.
  • Adjacent repeats. Real data, the Markov nulls and the shuffles that keep each inscription's own signs all sit near 1%. The unigram-matched and stratified nulls give 8 to 9%, because they draw signs without regard to what a given inscription contains. Rarity of repeats mostly reflects that single inscriptions seldom reuse a sign.

These four statistics were picked as illustrations, on one corpus. They are not a finding about the script, and a Markov null matching real on a statistic it was fitted on is expected, not informative.

Contents

real (2,722 rows): one row per inscription.

field meaning
inscription_id id in the source corpus
corpus cisi (179 inscriptions, Mohenjo-daro only) or indus_website (2,543 inscriptions, 52 sites)
signs list of sign IDs in canonical order (see Direction)
length number of signs
site, object_type, line_count, motif source metadata, unknown where absent
damaged source damage flag (always False in indus_website, so it carries no information there)
reading_direction_stored direction label in the source file
mean_uncertainty annotator uncertainty, 0 to 100 (CISI only, 0 elsewhere)
split train or test, 50/50, assigned by hashing the sequence so duplicates stay on one side

Sign IDs are only comparable within a corpus. CISI uses Parpola numbers (P121), indus_website uses its own G numbers. No images are included.

nulls (29,736 rows): synthetic sequences.

field meaning
null_id, corpus, seed identifiers. Three seeds (0, 1, 2) per null model
null_model see below
signs, length the generated sequence

All nulls are fitted on the train half of non-damaged real inscriptions only, so held-out test inscriptions cannot leak into them.

null model what it keeps what it destroys
within_inscription_shuffle each inscription's own sign multiset all order
global_shuffle corpus unigram frequencies and the length distribution order and co-occurrence
position_preserving_shuffle first and last sign of each inscription middle order
unigram_matched marginal sign and length distributions order, and sign-to-position structure beyond the marginals
bigram_markov real order-1 transitions and stopping dependency beyond order 1
trigram_markov real order-2 transitions and stopping dependency beyond order 2
motif_stratified, site_stratified the above, within motif or site strata cross-stratum structure

Full definitions are in the docstrings of the source modules in indus_decipher.

Task and reference baseline

Task: given held-out real test inscriptions and the nulls, score each sequence so that real ones rank above null ones. Metric is ROC AUC per null model (0.5 means indistinguishable).

Reference scorer (baseline.py): mean per-token log-probability under a bigram model minus that under a unigram model, both fit on the real train half. Intended as a floor, not a strong detector.

null model cisi (50 real test) indus_website (1,358 real test)
global_shuffle 0.966 0.930
within_inscription_shuffle 0.877 0.831
unigram_matched 0.800 0.841
motif_stratified 0.820 0.841
site_stratified 0.816 0.840
position_preserving_shuffle 0.474 0.571
bigram_markov 0.065 0.266
trigram_markov 0.040 0.238

AUC well below 0.5 for the Markov nulls is expected here. They were sampled from the same train bigram table the scorer uses, so they look more typical to it than real held-out text does. It shows that a detector built from a null's own statistics cannot test that null. A better detector needs features the generator did not use. The nulls are also fitted on train inscriptions and scored by a train-fit model, so they carry an in-sample advantage over the held-out real sequences. The AUCs against the shuffles are therefore conservative.

Direction

signs follows the canonical order from the indus_decipher loader (right to left), which reverses the stored order for rows labelled L-R. 2,440 of 2,543 indus_website rows and all CISI rows carry that label, so for almost every row signs is the reverse of the file order. Use reading_direction_stored if you need the original order.

Limitations

  • The CISI half has only 54 non-damaged train inscriptions and 50 non-damaged test inscriptions, so CISI nulls are weak and its AUCs are noisy. Treat CISI as a small demonstration set. With one site, site_stratified for CISI probably reduces to an unstratified null (not checked).
  • indus_website contains many repeated inscriptions (1,948 unique sequences among 2,543 rows). Splits are grouped by sequence, but repeats still inflate counts.
  • The M77 corpus used in indus_decipher is not included because it is not publicly redistributable.
  • Sign inventories differ between corpora, so results do not transfer between them without a sign mapping.
  • Nulls test structure, not meaning. A real corpus beating every null here shows it is not simple noise. It does not show the script encodes language.

Licensing and provenance

Rows carry the terms of their source. The dataset should not be treated as having one blanket licence.

  • cisi: derived from the MIT-licensed hand digitisation at mayig/indus-valley-script-corpus of a subset of the Corpus of Indus Seals and Inscriptions (CISI, Parpola and colleagues). The digitiser thanks Asko Parpola for the underlying corpus. Please cite both the digitisation and CISI.
  • indus_website: derived from the SQL data in yajnadevam/indus-website. That repository carries a GPL-3.0 licence file, but no separate licence for the data was found, and its upstream provenance is not documented there. Anyone redistributing or building on these rows should check with the maintainer. The maintainer of that repository is also the author of a disputed decipherment claim. This dataset uses only the transcription data, takes no position on that claim, and uses no sign-to-meaning mapping from it.
  • nulls: generated by code in the indus_decipher repository (MIT), from the train halves above, so they inherit the same source terms.
  • M77 (Mahadevan 1977) is used inside indus_decipher but is not publicly redistributable, so it is not included.

If a rights holder objects to any part, open an issue on the GitHub repository and that part will be removed.

Sources and references

Data sources

Methods this dataset's nulls and baseline build on

  • Rao, R. P. N., Yadav, N., Vahia, M. N., Joglekar, H., Adhikari, R., and Mahadevan, I. (2009). Entropic Evidence for Linguistic Structure in the Indus Script. Science, 324(5931), 1165.
  • Rao, R. P. N., Yadav, N., Vahia, M. N., Joglekar, H., Adhikari, R., and Mahadevan, I. (2009). A Markov Model of the Indus Script. PNAS, 106(33), 13685-13690.
  • Yadav, N., Joglekar, H., Rao, R. P. N., Vahia, M. N., Adhikari, R., and Mahadevan, I. (2010). Statistical Analysis of the Indus Script Using n-Grams. PLOS ONE, 5(3), e9506. https://doi.org/10.1371/journal.pone.0009506
  • Sinha, S., Izhar, A. M., Pan, R. K., and Wells, B. K. (2010). Network Analysis of a Corpus of Undeciphered Indus Civilization Inscriptions Indicates Syntactic Organization. Preprint (2010), listed on arXiv under the same title.
  • Farmer, S., Sproat, R., and Witzel, M. (2004). The Collapse of the Indus-Script Thesis: The Myth of a Literate Harappan Civilization. Electronic Journal of Vedic Studies, 11(2).

Code and companion work

Reproducing

git clone https://github.com/joyboseroy/indus_decipher
python build.py --repo ./indus_decipher --out .
python baseline.py --data . --out baseline_auc.csv

Citation

Joy Bose. Indus Script Sequences with Null Baselines. Hugging Face Datasets, 2026. Built from indus_decipher: https://github.com/joyboseroy/indus_decipher

Downloads last month
38