NyayaSiddhi v0.2: Indian element-proof development benchmark (no hidden split)

Hugging Face repository: AxisMeru/nyayasiddhi (NyayaSiddhi, न्यायसिद्धि: establishment of a claim under law). Not legal advice; no counsel review.

Status line (every page, every table): development benchmark: obj1_dev is seen (used to design fixes); obj1b is exposed to Anthropic through an earlier study in which Anthropic Sonnet read the set. v0.2; no hidden split in v0.2; a hash commitment is published when the strict court-established holdout reaches n ≥ 30 (today 3 EST / 10 NE, real, model-extracted). A place to compare; no ranking.

Labels are machine-extracted and unreviewed, and names appear as in the source judgments: see section 8a.

0. What v0.2 adds to v0.1

• The item files are the SAME as v0.1: data/obj1_dev.jsonl (sha256 da205768b9a2f857f1a644810c2a11d9fe372719ea72ec41d16f95af86be2b74) and data/obj1b_minus_cruelty.jsonl (sha256 2a4c4df2c2c4da9bb9502ff51a69eed4e5df32cec9c28d0982ae9887f57ef773). No item, label or split file changed.
• A charter (section 5a), versioning rules (section 6a), prediction, manifest and result-row schemas with a checker script, a contamination check per judge (section 7a) and a submission path (section 9a).

Changelog. v0.2.1: documentation only; data, scorer and splits unchanged (section 9a: checker path for both layouts, the jsonschema dependency, --items-dir needs both split files, the organiser's verdict-to-outcome conversion).

v0.2.2: organiser rows added (self-reported, dev stack (local)); data, scorer and splits unchanged.

v0.2.3: scope note only; data, scorer and splits unchanged.

v0.2.4: eight organiser reference rows added (self-reported: open models on a local development stack and a frontier reference arm); the self-serve scorer score_entry.py and its README added; these files were published in Hugging Face commit 660c7ff8e5bc079662036d4fe31233dc2267c0ae and are listed in SHA256SUMS; this card and release_manifest.json describe them; result files and documentation only; data, scorer (scorer/s6_scorer.py) and splits unchanged.

v0.2.5: documentation only: the High Court attribution text corrected; the 11 items without a PDF hash described; labels, personal data and the correction route stated (section 8a); data, scorer, rows and splits unchanged.

v0.2.6: removed a sentence that said we publish an oracle reference row; no such row is published. Documentation and message text only; data, scorer, rows and splits unchanged.

1. Motivation

We found no public board that scores whether a stated claim is proved from facts, element by element. This release lets outside models, harnesses and our own systems be compared on the same items with one scorer. It is a comparison place, not a claim that any system proves cases.

2. Composition

• Task: given the allegation-side facts of a court matter and a named offence contract, return PROOF or NOT_PROOF (optionally REFER, ABSTAIN, ERROR; optionally per-element decisions and quote spans). PROOF means every required element of the named contract is established. An arm that answers the whole case is a different unit from an arm that decides element by element; the unit is stated per row.
• Contracts in v0.2: criminal breach of trust, cheating, cheque dishonour (NI s.138). The cruelty (matrimonial) family is excluded entirely. Cheque dishonour is outside the contracts our product decides (the production configuration refers cheque-dishonour matters by design), so a PROOF row for it is not a product result.
• Label: the court's disposition (NE: court-rejected; EST: court-established) as extracted from judgments by models (Anthropic Haiku and a second-family extractor) and filtered by extractor agreement; not human-verified. Element labels and quote spans have no label at all; model outputs for them are submissions.
• Splits (both labelled): obj1_dev (seen) and obj1b_minus_cruelty (exposed). No hidden split in v0.2.
• Counts [after the screen for minors and protected victims and the prohibition check; from the builder's release manifest (release_manifest.json), recounted from the built files by a second reviewer]:
  • obj1_dev (seen): 177 items = 171 scored + 6 not scored (forced-REFER). Label: 111 EST, 66 NE. By family and label: cheating EST 9 / NE 28; cheque dishonour (NI s.138) EST 95 / NE 3; criminal breach of trust EST 7 / NE 35. 174 sources for 177 items, counting each of the 11 items that carry no pdf_sha256 as one source; 3 sources carry two items, and 2 spans are judged under both cheating and breach of trust. obj1_dev is therefore not one item per document.
  • obj1b_minus_cruelty (exposed): 154 items = 143 scored + 11 not scored (forced-REFER). Label: 20 EST, 134 NE. By family and label: cheating EST 12 / NE 51; criminal breach of trust EST 8 / NE 83. One item per source document (154 sources for 154 items).
  • Families are unbalanced; cheque dishonour is almost all court-established, and the two splits have no cruelty item.
• Fields per item: item_id (recomputable as SHA-256 of family, a newline and allegation_span, UTF-8); set and split_label (the split and its status); family (CHEAT, TRUST, NI138); label (NE or EST); court_disposition (the court's disposition: quash_allowed, petition_dismissed, conviction_affirmed, conviction_set_aside); named_ingredient and derived_standard (derived, model-extracted; not court-given; in this release derived_standard follows the disposition: quash_allowed → not_disclosed_on_complaint, petition_dismissed → prima_facie_disclosed, conviction_affirmed → proved, conviction_set_aside → not_proved); allegation_span (allegation-side text, finding quote removed, leak filter applied, at least 400 characters); scored (false for forced-REFER items); source with court, year and decision_date (court from the source dataset's court-code table; year and date parsed from the source key path; no row failed parsing in either split), pdf_sha256 (integrity field; null for 11 obj1_dev items: all Calcutta High Court, 7 breach of trust and 4 cheating, decided 2021 to 2023; why the hash is missing is not recorded) and attribution_group; prohibition (status, method, date of the per-source prohibition check). No case number (CNR), no per-item URL or title, and no storage key is shipped. Items are not individually looked up from this release.
• Length of the allegation-side span, in characters (median is the upper median for even n, the 90th percentile is the sorted value at index int(0.9 n); obj1b_minus_cruelty n = 154): obj1_dev minimum 404, median 793, 90th percentile 2,036, maximum 35,164; obj1b_minus_cruelty minimum 404, median 816, 90th percentile 2,346, maximum 5,927.

3. Collection and preprocessing

Items are built by pinned scripts from court judgments. The court's finding quote is removed from each item to prevent label leakage; a leak filter runs; items under 400 characters are dropped; duplicates are removed. One item per source document in obj1b_minus_cruelty (checked by the builder); in obj1_dev three sources carry two items. Each shipped source is checked for a court prohibition on reproduction (Copyright Act s.52(1)(q)(iv)); flagged sources are dropped: the check dropped 8 of 189 candidate items for obj1_dev and 10 of 165 for obj1b, and left none unchecked. Items were screened for named minors and protected victims, and flagged items were removed (one item in each split); no names-to-placeholders pass was applied, so adult names in judgments remain as published. The counts therefore run: obj1_dev 189 − 8 (prohibition check) − 3 (outside the dataset's attributed source scope) − 1 (names a minor) = 177; obj1b 165 − 10 (prohibition check) − 1 (names a minor) = 154.

4. Licences and attribution

• Our annotations, metadata and scorer: CC BY 4.0.
• Source text is court-derived. Copyright Act s.52(1)(q)(iv) lists reproduction or publication of a judgment or order of a court, tribunal or other judicial authority as not an infringement unless the court prohibits it. High Court data we hold is AWS Open Data under CC BY 4.0 and requires attribution. Attribution: "Indian High Court Judgments dataset, published by Dattam Labs, sourced from the eCourts judgments portal (judgments.ecourts.gov.in), licensed under CC BY 4.0 (https://creativecommons.org/licenses/by/4.0/). https://github.com/vanga/indian-high-court-judgments. Also listed at https://registry.opendata.aws/indian-high-court-judgments. Obtained from the AWS Open Data registry; the registry page was accessed on 6 October 2026 (our records); per-item fetch dates are not on file. Modified by the NyayaSiddhi organisers: items selected, the court's finding quote removed, labels added." The licence statement is the registry page's own; the licence text itself was not read.
• No request has been sent to any source holder.

5. Recommended and not-recommended uses

Recommended: comparing systems on the same items with the released scorer; studying false proof and coverage by family. Not recommended: as a measure of performance on real matters; as legal advice; for ranking systems (the scorer outputs no rank); for conclusions about court-established cases in families the set barely contains.

5a. Who runs this benchmark (charter)

Who runs this benchmark, and what that means

We built this benchmark and we enter our own systems in it. Our rows are marked organiser entry. They sit in the same table, are checked by the same script and carry the same intervals as every other row. Our prediction files and manifests are published the same way as everyone else's. Our rows are scored by the public scorer, and another member of the team re-scores them before they appear; that is an internal consistency check, not independent verification. Both splits of this version have been seen by us: one was used to design fixes and one was read by a model we use, and every row says so.

What each status means

Every row carries one status: self-reported, re-scored, reproduced or third-party verified. Self-reported means the organisers computed the row from the submitted file and nothing more. Re-scored means the organisers re-scored the file on a clean checkout and the figures are equal. Reproduced means the organisers re-ran the declared system. Third-party verified means an outside party re-scored or re-ran the system and left evidence. We never give our own row the reproduced or third-party verified status; only an outside party can verify it. Our own rows carry at most the re-scored status, which checks the arithmetic, not the system.

What we show and what we do not do

Every row shows its false-proof count and its coverage together, with exact intervals and the number of items behind each. No row is dropped because of its result. A row is withheld only when a check fails or its submitter withdraws it, and the reason is shown. A withdrawal leaves a visible row stating that the system was withdrawn. We correct an error by adding a new row, never by editing an old one. We do not tune on any hidden split, and this version has none. This page ranks nothing and does not pick a system as the answer.

6. Scoring

Per arm: false proof on NE items (k of n_NE, exact one-sided 95% upper bound) and coverage on EST items (k of n_EST, exact two-sided interval), by family and pooled, with the unit and operating point stated per arm. Both treatments of ERROR are reported (ERROR excluded, and ERROR counted as PROOF); a direction is stated only if both agree. Pairwise: paired exact McNemar on identical item sets, Holm for a named family; "no separation" is stated when p is 0.05 or more. A refuse-all arm shows 0 false proofs and 0 coverage by construction. The scorer is in this repository (scorer/s6_scorer.py, Python standard library only; items need item_id, family, label, scored; results need item_id, outcome). A consistency script was run on these files to check that they agree with each other and with the counts on this card; its report is included (consistency_report.json) and all of its checks pass. The scorer has been checked against the earlier recorded result: item by item on the seen Obj-1 rows of the earlier study (0 differences), and on aggregates only for obj1b (0 differences; no item-level check there).

6a. Versions and corrections

Versions are written v<major>.<minor>. A frozen version never changes: no item, label, split file or scoring setting is edited in place. An item added or removed, a label corrected, a split file changed or a scoring rule changed makes a new version; wording on this page or a new result row does not. Every split file is pinned by sha256, so a changed file under an old name is rejected. Results never mix across versions: a row names its benchmark_version and split. A correction is recorded as a dated entry that never changes a published figure by itself; a corrected item ships in the next version.

7. Initial rows (carried over from v0.1, unchanged)

Status: rows rescored from stored outputs of an earlier study; sentences are drafted from the rescored result and every number was recomputed by a second reviewer. Every row carries the label "descriptive, subset defined after the results (privacy exclusion)". One run each; no new model call. Intervals are exact two-sided 95% (Clopper-Pearson). ERROR counts: 0 in every arm, so both ERROR treatments give the same figures. Arms: B is a frontier model (Anthropic Sonnet) inside our element harness; F is B's stored replies without the verbatim-quote check, defeaters or Lean assembly; G is the same frontier model asked directly; A' is our production configuration; A is Arm-0, the 4B judge alone; not the production configuration. The unit differs by arm (an arm that answers the whole case is a different unit from one that decides element by element) and no arm is shown at a common operating point.

Table 1: obj1b_minus_cruelty (exposed): 143 scored items, 128 court-rejected (NE) and 15 court-established (EST).

Arm	False proofs on NE (k of 128)	Exact 95% interval	Coverage on EST (k of 15)	Exact 95% interval
B	11	4.4% to 14.9%	5	11.8% to 61.6%
F	11	4.4% to 14.9%	5	11.8% to 61.6%
G	9	3.3% to 12.9%	6	16.3% to 67.7%
A'	0	0% to 2.8%	0	0% to 21.8%

• A' refuses every item (143 of 143 NOT_PROOF); 0 false proofs is not evidence that the system avoids false proofs in general, and is never read without the 0 of 15 coverage beside it.
• G against B on false proofs: no separation (5 items where only G proved, 7 where only B proved; exact McNemar p = 0.774, the same with errors counted as proofs). F against B: 0 items differ (p = 1.0).
• The 11 forced-REFER items in this split are not scored.
• Note 1: most of the false proofs in the full set of the earlier study came from the excluded cruelty family; these rates are for the non-cruelty subset and are not the full-set rates; the subset was chosen for privacy, not for results, and a difference from the full-set rates is not a change in any arm. The earlier signed record says: "Court-rejected false proofs by family: CRUELTY (154) B 98 (63.6%, 55.5% to 71.2%), F 101, G 99 (64.3%, 56.2% to 71.8%); CHEAT (50) 5 each (10.0%, 3.3% to 21.8%); TRUST (86) B 6 (7.0%, 2.6% to 14.6%), F 6, G 4 (4.7%, 1.3% to 11.5%). So 98 of B's 109 and 99 of G's 108 false proofs are CRUELTY; CRUELTY is 154 of the 290 court-rejected items." Full-set results (including the excluded CRUELTY items) are in our earlier signed record.
• Note 2: coverage rests on 15 court-established items (cheating 10, breach of trust 5), so its intervals are wide; the breach-of-trust cell (n = 5) is descriptive only.

Table 2: obj1_dev (seen), on the 112-item overlap of the 171 scored released items (the earlier study ran a 150-item set; 59 released scored items have no outcome in that study): 44 NE and 68 EST. Of the 68 EST items, 59 are cheque dishonour (NI s.138).

Arm	False proofs on NE (k of 44)	Exact 95% interval	Coverage on EST (k of 68)	Exact 95% interval
A (Arm-0, the 4B judge alone; not the production configuration)	3	1.4% to 18.7%	2	0.4% to 10.2%
B	7	6.6% to 30.1%	7	4.2% to 20.1%
F	7	6.6% to 30.1%	7	4.2% to 20.1%
G	3	1.4% to 18.7%	33	36.2% to 61.0%

• Pairs on false proofs (44 NE items, descriptive): A against B, 1 item where only A proved, 5 where only B proved (p = 0.219); G against B, 1 and 5 (p = 0.219); F against B, 0 differ (p = 1.0). All are "no separation".
• Note 3: the arm comparisons here use only 44 NE items.
• Note 4: "A" is Arm-0, the 4B judge alone; not the production configuration; this split was used to design fixes.
• The 6 forced-REFER items in this split are not scored.

7a. Contamination check

A contamination check against the training text of the judges found no exact or 8-gram overlap in either split. It is 8-gram and hash identity only; paraphrase is not checked (see 7b for scope).

7b. Our own rows (added in v0.2.2)

Our own rows at v0.2.2 (organiser entry; descriptive; dev stack (local), not the production deployment; one run; seen splits; no ranking). We ran our system (Config C: two judges and a structure check) once on each released split, on a local development stack, and score it with the released scorer at the published n. On obj1b_minus_cruelty (n = 143 scored items: 128 court-rejected, 15 court-established) it returned PROOF on 0 of 128 court-rejected items (exact 95% interval 0.0% to 2.8%; one-sided 95% upper bound 2.31%) and on 0 of 15 court-established items (0.0% to 21.8%). On obj1_dev (n = 171: 64 court-rejected, 107 court-established) it returned PROOF on 0 of 64 court-rejected items when one errored item is counted as NOT_PROOF (0.0% to 5.6%; one-sided upper bound 4.6%), and on 1 of 64 when that one errored item is counted as PROOF (0.0% to 8.4%; one-sided upper bound 7.2%); the two readings disagree, so the row is never read as "0 of 64" alone. It returned PROOF on 0 of 107 court-established items (0.0% to 3.4%). The engine returned PROOF on no item on either split, so its rows equal those of a system that refuses every item (the baseline row below); zero false proofs is read only beside zero coverage, and is not evidence that the system avoids false proofs in general.

Row for Table 1 (obj1b_minus_cruelty, n = 143: 128 court-rejected, 15 court-established):

Arm	False proofs on NE (k of 128)	Exact 95% interval	Coverage on EST (k of 15)	Exact 95% interval
nyaya-product / engine-42e6ed71_cfg-c26cb086 (Config C: 4B + 32B on the dev stack (local); organiser entry; self-reported)	0	0.0% to 2.8%	0	0.0% to 21.8%
Table 2b (obj1_dev (seen), n = 171: 64 court-rejected, 107 court-established; the card's Table 2 stays unchanged on its 112-item overlap and no pairwise statement is made across the two):
Arm	False proofs on NE (k of 64)	Exact 95% interval	Coverage on EST (k of 107)	Exact 95% interval
---	---	---	---	---
nyaya-product / engine-42e6ed71_cfg-c26cb086, ERROR counted NOT_PROOF (organiser entry; self-reported; dev stack (local))	0	0.0% to 5.6%	0	0.0% to 3.4%
the same system, ERROR counted PROOF	1	0.0% to 8.4%	0	0.0% to 3.4%
The refuse-all baseline A' is the existing card row; the paragraph above says these rows equal it.

Notes for the page: the 112-item overlap rows already published (Tables 1 and 2 of v0) are a different item set; no comparison is made across them. The earlier single-item re-run of fccbecb0 on another head (3c1c57d5) returned REFER (input_too_long) and is not a result row. 93 of the 107 established items on obj1_dev are cheque dishonour (NI s.138), which is fail-closed by design.

Limits to print beside it. Our own rows only; one run; labels are model-extracted and we have seen both splits (obj1_dev is the design set; obj1b_minus_cruelty was exposed to us); dev stack (local), not the production configuration; 4B dev-to-production parity not established; counts only; no improvement claim and no ranking; not an independent verification. Contamination scope: the manifests' overlap_check states that the 8-gram and hash-identity check covers the train file el_comb_36a69ec ONLY (no paraphrase check, nothing else); it is not a contamination clearance for the whole system (see results/append/APPEND_O6_6_ROWS.md, section 5).

Files. results/append/APPEND_O6_6_ROWS.md (the signed append text) and, for each split, results/append/<split>/predictions.jsonl, manifest.json and check_submission_output.json (the checker printed PASS; the re-score and the contamination flag are organiser steps), with the first-run manifests under original_v0.1/. The prediction files and manifests are byte-identical to the signed run's files.

7c. Organiser reference rows (added in v0.2.4)

Eight organiser rows are published in results/append/seed_rows_682/results.jsonl, each with its manifest in manifests/ and its prediction file in predictions/ beside it. They are organiser entries and self-reported. They are two open models (Qwen2.5-32B base and Qwen3-14B-AWQ, exact revisions in the manifests) run on a local development stack, and a frontier reference arm (claude-sonnet-5-5, a direct arm and a harness arm, with prompts the organisers wrote; not an outside entry) run through the Anthropic API, each on both released splits. Both splits are development splits (obj1_dev is seen; obj1b_minus_cruelty is exposed), so no row is a held-out result. Each row shows false-proof and coverage counts with exact intervals and the error mapping, and its manifest notes carry the disclosures for that run: what was sent to a vendor's API, settings changed after a smoke test, and any rerun after a stopped run. No rank is given, no row is a claim about real cases, and this card quotes none of the figures: the row files are the source.

8. Known limits

Labels are model-extracted and filtered by extractor agreement, not human-verified. named_ingredient and derived_standard are derived. Both frontier arms in the initial rows are Anthropic Sonnet (direction of any bias unknown), with one draw per element or item. Families are unbalanced and some cells are small, so intervals are wide. obj1_dev was used to design fixes; obj1b was read by a frontier model family, and a public release makes both usable as training data. Court prohibitions are checked per source: 8 of 189 candidate items (obj1_dev) and 10 of 165 (obj1b) were dropped, none left unchecked. Text may name adult private parties. No hidden split exists in v0.2, so every released item is public and a submitted score on it is a development figure; training on the released items is possible and is declared, not checked.

8a. Labels, personal data and corrections (added in v0.2.5)

Labels. The fields label, court_disposition, named_ingredient and derived_standard are machine-extracted by two model families and filtered by their agreement. A person has not reviewed them. Before you use them to train a model, check your own model provider's terms.

Personal data. Names are present as in the source judgments. They are not redacted, apart from the screen for minors and protected victims described in this card. A pattern search of the item files found 29 items with a name written with an honorific, 30 with a phrase giving a person's relation to another, and 28 with a crime number. To ask for an item to be corrected or removed, open a discussion on the dataset page.

9. Maintenance

Versioned releases; corrections are recorded as dated entries (section 6a). A hidden split needs court-established supply that does not exist yet: the strict court-established holdout has 3 court-established and 10 court-rejected items (real, model-extracted, dev stack). A hash commitment to a hidden split will be published when the strict court-established holdout reaches n ≥ 30; until then there is none, and no one tunes against any hidden data because there is none.

9a. Submit a system

You score yourself and we check you. (1) Download data/obj1_dev.jsonl and data/obj1b_minus_cruelty.jsonl and verify them against SHA256SUMS. (2) For ONE split, write predictions.jsonl: one JSON line per item of that split, every item exactly once: {"item_id": "...", "verdict": "PROOF" | "NOT_PROOF"} (optional elements, error, score; a REFER or ABSTAIN outcome is mapped to NOT_PROOF and declared in the manifest). A missing item is an error, never a silent NOT_PROOF. (3) Write manifest.json (fields and schema: schema674/manifest.schema.json; every model id pinned with a revision, publish_consent: true, prediction_file_sha256 = the sha256 of your predictions file). (4) Put both files, and nothing else, in submissions/<system_name>/<system_version>/ (names: letters, digits, . _ + -, at most 64 characters). (5) Check it yourself. First pip install jsonschema (it is listed in schema674/requirements.txt). Then run the checker: from a GitHub checkout of the repository, python research/benchmarks/submission680/check_submission.py --dir submissions/<system_name>/<system_version> --items-dir <items dir>; from this Hugging Face dataset layout, python submission680/check_submission.py --dir submissions/<system_name>/<system_version> --items-dir <items dir>. --items-dir must contain BOTH split files (obj1_dev.jsonl and obj1b_minus_cruelty.jsonl), even when your predictions are for one split. It must print PASS. (6) Open a Hugging Face pull request (or a discussion with the two files attached) on AxisMeru/nyayasiddhi with the template in submission680/SUBMISSION_PATH_680.md. We re-run the same check on the PR head, convert each submission row {item_id, verdict} to the scorer's {item_id, outcome} (PROOF stays PROOF, NOT_PROOF stays NOT_PROOF, a row with error: true becomes ERROR; the scorer reports ERROR both ways), score the file with the released scorer (scorer/s6_scorer.py, --items the split file, --arm NAME=RESULTS.jsonl), and append a result row; the row says self-reported until we re-score it on a clean checkout (re-scored) or re-run your system (reproduced). Rows are append-only; a correction is a new row. A file whose verdicts equal the released labels on every scored item is not accepted as a submission. What a row shows: false proofs on court-rejected items (k of n, exact 95% interval) and coverage on court-established items (k of n, exact interval), by family and pooled, with the unit (whole item or element by element) and the contamination flag. No score is shown without its n and interval. Labels are model-extracted and filtered by extractor agreement, not human-verified.

Score yourself first. score_entry.py and README_SCORE_YOUR_OWN_SUBMISSION.md in this repository run the organisers' check and the released scorer on your folder, on your own machine: nothing is uploaded and no key or account is needed. The output is a self-scored preview, not a published row. Compare the scorer_sha256 and validator_sha256 in your receipt with the lines for scorer/s6_scorer.py and schema674/validate_674.py in SHA256SUMS on the dataset page; if they differ, you are not running the released scorer.

Hosted evaluation: PLANNED. A hosted evaluation service is planned and does not exist. There is no hidden split in v0.2, so nothing is hosted and nothing can be submitted for hosted scoring.

Current state of this path: it has been exercised only on a synthetic file from a clean checkout; no outside party has submitted yet.

10. Attribution and citation

• Source attribution (High Court judgments): "Indian High Court Judgments dataset, published by Dattam Labs, sourced from the eCourts judgments portal (judgments.ecourts.gov.in), licensed under CC BY 4.0 (https://creativecommons.org/licenses/by/4.0/). https://github.com/vanga/indian-high-court-judgments. Also listed at https://registry.opendata.aws/indian-high-court-judgments. Obtained from the AWS Open Data registry; the registry page was accessed on 6 October 2026 (our records); per-item fetch dates are not on file. Modified by the NyayaSiddhi organisers: items selected, the court's finding quote removed, labels added."
• Our annotations, metadata and scorer are released under CC BY 4.0.
• How to cite this dataset: NyayaSiddhi v0.2 (Indian element-proof development benchmark, no hidden split), https://huggingface.co/datasets/AxisMeru/nyayasiddhi, accessed [date of access], licence CC BY 4.0. State the split used (obj1_dev or obj1b_minus_cruelty) and keep the status line with any result.