The full dataset viewer is not available (click to read why). Only showing a preview of the rows.
Error code: DatasetGenerationError
Exception: TypeError
Message: Couldn't cast array of type string to null
Traceback: Traceback (most recent call last):
File "/usr/local/lib/python3.14/site-packages/datasets/builder.py", line 1837, in _prepare_split_single
writer.write_table(table)
~~~~~~~~~~~~~~~~~~^^^^^^^
File "/usr/local/lib/python3.14/site-packages/datasets/arrow_writer.py", line 765, in write_table
self._write_table(pa_table, writer_batch_size=writer_batch_size)
~~~~~~~~~~~~~~~~~^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
File "/usr/local/lib/python3.14/site-packages/datasets/arrow_writer.py", line 773, in _write_table
pa_table = table_cast(pa_table, self._schema)
File "/usr/local/lib/python3.14/site-packages/datasets/table.py", line 2369, in table_cast
return cast_table_to_schema(table, schema)
File "/usr/local/lib/python3.14/site-packages/datasets/table.py", line 2303, in cast_table_to_schema
cast_array_to_feature(
~~~~~~~~~~~~~~~~~~~~~^
table[name] if name in table_column_names else pa.array([None] * len(table), type=schema.field(name).type),
^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
feature,
^^^^^^^^
)
^
File "/usr/local/lib/python3.14/site-packages/datasets/table.py", line 1852, in wrapper
return pa.chunked_array([func(chunk, *args, **kwargs) for chunk in array.chunks])
~~~~^^^^^^^^^^^^^^^^^^^^^^^^
File "/usr/local/lib/python3.14/site-packages/datasets/table.py", line 2143, in cast_array_to_feature
return array_cast(
array,
...<2 lines>...
allow_decimal_to_str=allow_decimal_to_str,
)
File "/usr/local/lib/python3.14/site-packages/datasets/table.py", line 1854, in wrapper
return func(array, *args, **kwargs)
File "/usr/local/lib/python3.14/site-packages/datasets/table.py", line 2005, in array_cast
raise TypeError(f"Couldn't cast array of type {_short_str(array.type)} to {_short_str(pa_type)}")
TypeError: Couldn't cast array of type string to null
The above exception was the direct cause of the following exception:
Traceback (most recent call last):
File "/src/services/worker/src/worker/job_runners/config/parquet_and_info.py", line 1369, in compute_config_parquet_and_info_response
parquet_operations, partial, estimated_dataset_info = stream_convert_to_parquet(
~~~~~~~~~~~~~~~~~~~~~~~~~^
builder, max_dataset_size_bytes=max_dataset_size_bytes
^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
)
^
File "/src/services/worker/src/worker/job_runners/config/parquet_and_info.py", line 948, in stream_convert_to_parquet
builder._prepare_split(split_generator=splits_generators[split], file_format="parquet")
~~~~~~~~~~~~~~~~~~~~~~^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
File "/usr/local/lib/python3.14/site-packages/datasets/builder.py", line 1683, in _prepare_split
for job_id, done, content in self._prepare_split_single(
~~~~~~~~~~~~~~~~~~~~~~~~~~^
gen_kwargs=gen_kwargs, job_id=job_id, **_prepare_split_args
^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
):
^
File "/usr/local/lib/python3.14/site-packages/datasets/builder.py", line 1869, in _prepare_split_single
raise DatasetGenerationError("An error occurred while generating the dataset") from e
datasets.exceptions.DatasetGenerationError: An error occurred while generating the datasetNeed help to make the dataset viewer work? Make sure to review how to configure the dataset viewer, and open a discussion for direct support.
secret string | word string | tier string | expected_bands list | subtype null | note null |
|---|---|---|---|---|---|
auntie | relative | near | [
1,
2
] | null | null |
auntie | guardian | near | [
1,
2
] | null | null |
auntie | bracket | mid | [
3
] | null | null |
auntie | niece | mid | [
3
] | null | null |
auntie | tap | far | [
4,
5
] | null | null |
auntie | hallucinating | far | [
4,
5
] | null | null |
auntie | drugged | far | [
4,
5
] | null | null |
auntie | carpenter | far | [
4,
5
] | null | null |
auntie | unchanged | far | [
4,
5
] | null | null |
basket | hamper | near | [
1,
2
] | null | null |
basket | crate | near | [
1,
2
] | null | null |
basket | weave | mid | [
3
] | null | null |
basket | picnic | mid | [
3
] | null | null |
basket | trumpeting | far | [
4,
5
] | null | null |
basket | helpline | far | [
4,
5
] | null | null |
basket | baraka | far | [
4,
5
] | null | null |
basket | platt | far | [
4,
5
] | null | null |
basket | loop | far | [
4,
5
] | null | null |
border | frontier | near | [
1,
2
] | null | null |
border | fence | mid | [
3
] | null | null |
border | checkpoint | mid | [
3
] | null | null |
border | ulf | far | [
4,
5
] | null | null |
border | rahman | far | [
4,
5
] | null | null |
border | grout | far | [
4,
5
] | null | null |
border | misty | far | [
4,
5
] | null | null |
border | biodegradable | far | [
4,
5
] | null | null |
chimney | flue | near | [
1,
2
] | null | null |
chimney | exhaust | near | [
1,
2
] | null | null |
chimney | soot | mid | [
3
] | null | null |
chimney | duct | mid | [
3
] | null | null |
chimney | nipple | far | [
4,
5
] | null | null |
chimney | assuage | far | [
4,
5
] | null | null |
chimney | flapper | far | [
4,
5
] | null | null |
chimney | montmartre | far | [
4,
5
] | null | null |
chimney | astray | far | [
4,
5
] | null | null |
countryside | rural | near | [
1,
2
] | null | null |
countryside | farmland | near | [
1,
2
] | null | null |
countryside | meadow | mid | [
3
] | null | null |
countryside | pastoral | mid | [
3
] | null | null |
countryside | insatiable | far | [
4,
5
] | null | null |
countryside | neumann | far | [
4,
5
] | null | null |
countryside | perjury | far | [
4,
5
] | null | null |
countryside | daddies | far | [
4,
5
] | null | null |
countryside | autocrat | far | [
4,
5
] | null | null |
desert | wasteland | near | [
1,
2
] | null | null |
desert | barren | near | [
1,
2
] | null | null |
desert | dune | mid | [
3
] | null | null |
desert | drought | mid | [
3
] | null | null |
desert | kubota | far | [
4,
5
] | null | null |
desert | synthetic | far | [
4,
5
] | null | null |
desert | umar | far | [
4,
5
] | null | null |
desert | eternally | far | [
4,
5
] | null | null |
desert | stabs | far | [
4,
5
] | null | null |
dessert | candy | near | [
1,
2
] | null | null |
dessert | confection | near | [
1,
2
] | null | null |
dessert | sweetness | mid | [
3
] | null | null |
dessert | treat | mid | [
3
] | null | null |
dessert | schiff | far | [
4,
5
] | null | null |
dessert | wracked | far | [
4,
5
] | null | null |
dessert | nun | far | [
4,
5
] | null | null |
dessert | thinking | far | [
4,
5
] | null | null |
dessert | pertussis | far | [
4,
5
] | null | null |
duration | interval | near | [
1,
2
] | null | null |
duration | moment | near | [
1,
2
] | null | null |
duration | schedule | mid | [
3
] | null | null |
duration | era | mid | [
3
] | null | null |
duration | underlings | far | [
4,
5
] | null | null |
duration | muss | far | [
4,
5
] | null | null |
duration | fielded | far | [
4,
5
] | null | null |
duration | spartans | far | [
4,
5
] | null | null |
duration | interfaces | far | [
4,
5
] | null | null |
enjoyment | amusement | near | [
1,
2
] | null | null |
enjoyment | delight | near | [
1,
2
] | null | null |
enjoyment | cheer | mid | [
3
] | null | null |
enjoyment | playful | mid | [
3
] | null | null |
enjoyment | redress | far | [
4,
5
] | null | null |
enjoyment | misunderstand | far | [
4,
5
] | null | null |
enjoyment | brampton | far | [
4,
5
] | null | null |
enjoyment | attested | far | [
4,
5
] | null | null |
enjoyment | grandfathers | far | [
4,
5
] | null | null |
freedom | liberty | near | [
1,
2
] | null | null |
freedom | independence | near | [
1,
2
] | null | null |
freedom | rights | mid | [
3
] | null | null |
freedom | release | mid | [
3
] | null | null |
freedom | tablecloth | far | [
4,
5
] | null | null |
freedom | juliette | far | [
4,
5
] | null | null |
freedom | patching | far | [
4,
5
] | null | null |
freedom | pimlico | far | [
4,
5
] | null | null |
freedom | granny | far | [
4,
5
] | null | null |
flour | batter | mid | [
3
] | null | null |
flour | protein | mid | [
3
] | null | null |
flour | outstanding | far | [
4,
5
] | null | null |
flour | depot | far | [
4,
5
] | null | null |
flour | jason | far | [
4,
5
] | null | null |
flour | commissions | far | [
4,
5
] | null | null |
flour | font | far | [
4,
5
] | null | null |
justice | fairness | near | [
1,
2
] | null | null |
justice | equity | near | [
1,
2
] | null | null |
justice | envelope | mid | [
3
] | null | null |
justice | adjudicator | mid | [
3
] | null | null |
Hinglish Relatedness Benchmark
A practical, reproducible, AI-authored romanized-Hinglish (Latin-script Hindi-English) semantic-relatedness benchmark: 34 secret words across 7 domains (nature, emotion, abstract, object, place, food, kinship), each with tiered near/mid/far associate words scored for expected relatedness band, plus a parallel English-language version and cross-lingual translation-consistency pairs.
This is not a peer-reviewed gold standard. It is a small, honestly-sourced, methodologically transparent benchmark built to fill a real gap: no public romanized-Hinglish relatedness benchmark exists to compare against. Report weaknesses and residuals when you use it — that's the standard this card holds itself to.
What's in it
| File | Records | Purpose |
|---|---|---|
dim1_hien.json |
302 | Headline metric: romanized-Hindi secret -> near/mid/far tiered associate words, English-side associates included, translation-tagged where applicable |
dim1_en.json |
303 | Parallel English-language version of the same 34 secrets/tiers |
dim2_pairs.json |
34 | Secondary metric: cross-lingual (hi-en secret, EN translation) consistency pairs |
public_benchmark_split.json |
— | Domain-stratified dev/test split (17/17), seed=42, for checkpoint selection |
Each dim1_* record: {secret, word, tier (near/mid/far), expected_bands, subtype, note}.
Each dim2_pairs record: {hien_secret, en_gloss, domain}.
Provenance — read this before citing "expert-annotated"
This benchmark is majority AI-authored content, not independently human-annotated, not externally-sourced commercial data, and not a crowd-sourced or academically-validated dataset:
- Near/mid tier word associations (the semantically load-bearing part): hand-authored as literal code by an AI coding assistant (Claude, Anthropic) during this project's development. Spot-checked by the project owner at specific review gates (e.g. a homograph-collision audit on the earlier 25-secret core set), not independently re-annotated word-by-word by a human linguist.
- Far-tier filler words: mechanically sampled via a hash-seeded RNG over the project's own vocabulary — not curated at all, by design; their only intended property is "not near/mid," verified after the fact.
- Safety-scrub and semantic-proximity review: an AI coding-assistant subagent ran an automated string-collision scrub and a semantic-proximity scan (embedding-rank based) against all live game content, fixed 17 real string collisions and 20 semantic near-misses, all disclosed with before/after word tables in the full Phase 1 report (see Reproducibility).
Do not describe this dataset as "human-annotated" or "expert-annotated" in the sense of an independent linguist's judgment. It is AI-authored, methodologically rigorous (documented tiering discipline, verified vocabulary membership, two independent scrub passes), and transparently disclosed as such.
Attribution
- Base romanized-Hindi vocabulary and transliteration resources build on AI4Bharat's Aksharantar and IndicCorp v2 (CC0), used earlier in this project's data pipeline.
- The reference fine-tuned model this benchmark evaluates, gauravgandhi2411/hinglish-relatedness-sbert, is built on L3Cube's indic-sentence-bert-nli. Neither L3Cube nor AI4Bharat authored or reviewed this specific benchmark's word pairs.
License: CC-BY-4.0. Attribution to this project (Warmer / gauravgandhi2411) requested.
Headline metric: dim1 (semantic-ranking quality)
Spearman rank correlation between expected closeness tier (near/mid/far) and a model's actual embedding-rank, pooled per secret. Dev/test split (17/17, domain-stratified, seed=42) was built to fix a known problem in this project's earlier, smaller held-out set (dev-test Pearson correlation only ~0.12, i.e. dev performance barely predicted test performance, making checkpoint selection untrustworthy). Measured on this split: dev<->test Pearson r = 0.920 (n=8 models) — dev now materially predicts test.
Reference model (gauravgandhi2411/hinglish-relatedness-sbert) dim1 Spearman: 0.657
[95% CI 0.569, 0.736] full 34-secret set (0.671 dev-only, 0.642 test-only —
CI-overlapping, stable across the split).
Secondary metric: dim2 (cross-lingual consistency) — honest cluster-dependence
Dim2 measures whether a secret's correct English translation lands in the expected relatedness band. This metric is noticeably less stable across domains than dim1, and degrades in specific, identifiable clusters rather than uniformly.
Methodology notes (found during Phase 1 review, both corrected before v1.1 — see Changelog):
- 2 of the 34 secrets (
paneer,khichdi) are loanwords spelled identically in both languages — their EN gloss and hi-en secret are the literal same string, so they trivially self-match at rank 1 ("win" band). The pass/fail definition here treats only bands 1-2 ("scorching"/"hot") as consistent, deliberately excluding band 0 to prevent literal string duplicates from gaming the metric in the general case — but for a genuinely-identical loanword pair this excludes the ONE correct outcome, scoring a trivial self-match as a "failure." Not a meaningful test of cross-lingual understanding either way (a lookup table scores them identically); excluded from the domain rates below, which reflect only the 32 informative (non-degenerate) pairs. besan's EN gloss was originally "gram" — in this model's embedding space "gram" reads overwhelmingly as the Indian-English "gram panchayat" (village council) sense, not "gram flour," so besan/gram scored a spurious failure driven by gloss ambiguity, not a real relatedness gap. Corrected tobesan→**"flour"** (verified: unambiguous neighborhood — dough, wheat, grain, bread — and besan now correctly lands in the "hot" band against it). This is a benchmark-labeling fix, not a model fix.
| Domain | Fail rate (32 informative pairs, n=3-5/domain) |
|---|---|
| emotion | 2/5 (40%) |
| nature | 2/5 (40%) |
| kinship | 1/4 (25%) |
| place | 1/5 (20%) |
| object | 1/5 (20%) |
| food | 0/3 (0%) |
| abstract | 0/5 (0%) |
After both corrections, the real residual is emotion and nature specifically (40% each) — not a broad "food/nature/emotion" pattern. Food's apparent weakness was entirely a labeling artifact (2 degenerate pairs + 1 bad gloss); once fixed it's a clean 0/3. Object, place, and kinship each have exactly one genuine failure (chaabi/key, dehat/countryside, bhanja/nephew respectively) — not zero, but not clustered either. At n=3-5 secrets per domain this remains a directional signal, not a statistically tight one — a single item flipping changes a domain's rate by 20-25 points. Reported honestly rather than smoothed into a single aggregate pass-rate (0.706 [0.559, 0.853] overall), which would hide this structure.
Root cause of the emotion/nature residual (checked directly, not assumed): for
junoon, registan, and partially jheel/udaasi, the failing word's own nearest
embedding-neighbors are dominated by unrelated English words sharing a coincidental
subword fragment — e.g. junoon neighbors afternoon/noon/june (shares the "noon"/
"jun" subword), registan neighbors register/registry/registration (shares
"regist-"). These are conventionally, correctly spelled Hindi romanizations — this is
not a transliteration error, it's the base model's English-dominated subword
tokenizer fragmenting short/rare romanized tokens into pieces that collide with common
English morphemes. This is a genuine model/tokenizer characteristic, not something
further benchmark editing can fix.
Downstream mitigation exists, deliberately not reflected in the numbers above. The
reference model's deploying application (Warmer) now ships a build-time, non-fine-tuning
generator mitigation for exactly these words (Devanagari-embedding substitution for
junoon/registan/jheel/udaasi specifically, not a general fix) — see the model
card's "Production wrapper" section. This benchmark intentionally continues to measure
the raw, unwrapped model so it stays a fair, stable ruler for comparing models —
applying a generator-side patch here would make this benchmark measure a moving,
non-reproducible target instead of the model itself. Anyone comparing models against
this benchmark should do the same: evaluate raw model output, not an application-level
wrapper.
Scrub / safety notice (for anyone extending this benchmark)
Every secret and associate word was checked against this project's live game content (daily-schedule secrets, all themed-pack secrets, aliases) two ways: string/morphological collision, and embedding-rank semantic proximity. Zero blocking collisions of either kind remain in the published files. Full methodology, the calibration analysis behind the semantic-proximity check's severity thresholds, and every word substitution made (37 total across both scrub passes) are documented in the full Phase 1 report (Reproducibility below).
Reproducibility
Full methodology, all intermediate numbers, the complete word-substitution history, and the
scrub calibration analysis: public_benchmark_phase1_report.md in the source repository
(gauravgandhi2411/mindmeld, generator/evals/finetune/data/). Split seed 42 throughout.
Changelog
- v1.1 (this version): corrected
besan's EN gloss from "gram" (ambiguous — reads as "village"/gram-panchayat in this corpus) to "flour" (verified unambiguous), and its near-tier associate word from "legume" to "wheat" (better embedding-verified fit). This changed the dim2 domain-dependence table (food's fail rate corrected from an apparent 33% to 0%, once combined with the v1.0 degenerate-pair exclusion) and improved dev<->test Pearson r from 0.894 to 0.920. All 8 model checkpoints re-resolved against the corrected word content; both scrubs re-run clean. No other secret's content changed in this version. - v1.0: initial publish. 34 secrets, 7 domains, dim1 headline + dim2 secondary metric, 20 targeted word substitutions for semantic-proximity optics (see report section 2c) on top of 17 string-collision fixes (section 2). Briefly carried an uncorrected domain table overstating food's fail rate at 60% (the degenerate-pair artifact, not yet identified) — corrected same-day, before v1.1's gloss fix.
- Downloads last month
- 34