TIDE evaluates how large language models handle version resolution in evolving official documents. It contains 3,050 question–answer pairs spanning 644 Bangladesh customs instruments issued between 1969 and 2025.

Laws and regulations change over time through amendments. The central difficulty is identifying which version of a document was in force on a specific query date — a task where models are more likely to find correct versions than to reject incorrect ones.

Dataset composition

3,050 QA pairs across 8 task types
644 official documents: 3 Acts, 20 General Orders, 609 Statutory Regulatory Orders, 12 Rules
265 evolving-concept threads across 157 topic clusters
Time span: 1969–2025
Code-mixed Bangla–English content with heterogeneous layouts

Task	Items	What it tests
`scenario_short_answer`	562	Extracting values from realistic, context-rich scenarios
`perturbation_detection`	498	Identifying factually incorrect statements
`open_ended`	487	Explaining policy rationales
`relative_time_qa`	463	Resolving offsets from anchor dates
`temporal_scenario_short_answer`	363	Tracing amendments across multiple dated instruments
`temporal_mcq`	354	Selecting correct instruments or dates among similar options
`event_sorting`	207	Ordering date-stripped events chronologically
`context_misalignment`	116	Rejecting fluent but false preambles
Total	3,050

Usage

TIDE is an evaluation-only benchmark: every config has a single test split.

from datasets import load_dataset

# Everything at once (union schema; fields a task does not use are null)
tide = load_dataset("mahbubhimel/TIDE", "all")["test"]          # 3,050 rows

# One task at a time, with that task's exact schema
mcq = load_dataset("mahbubhimel/TIDE", "temporal_mcq")["test"]  # 354 rows

# The `all` config carries a `type` column, so it can also be filtered
sorting = tide.filter(lambda r: r["type"] == "event_sorting")

Every row has a stable id of the form <task>_<0001> and a type column naming its task.

Task schemas and examples

`temporal_mcq` — 354 items

question, options (4 strings), answer, answer_index, explanation. answer always equals options[answer_index].

{
  "id": "temporal_mcq_0001",
  "type": "temporal_mcq",
  "question": "Which statutory notification was repealed on June 1, 2017, by S.R.O. No. 128-AIN/2017/14/Customs regarding import duty exemptions on machinery and parts?",
  "options": [
    "S.R.O. No. 135-AIN/2012/15/Customs dated June 4, 2012",
    "S.R.O. No. 176-AIN/2012/2404/Customs dated June 7, 2012",
    "S.R.O. No. 145-AIN/2014/2508/Customs dated June 12, 2014",
    "S.R.O. No. 32-AIN/2016/02/Customs dated February 15, 2016"
  ],
  "answer": "S.R.O. No. 135-AIN/2012/15/Customs dated June 4, 2012",
  "answer_index": 0,
  "explanation": "According to Clause [16], S.R.O. No. 128-AIN/2017/14/Customs, published on June 1, 2017, repealed S.R.O. No. 135-AIN/2012/15/Customs, which was originally issued on June 4, 2012 (২১ জ্যৈষ্ঠ, ১৪১৯ বঙ্গাব্দ মোতাবেক ৪ জুন, ২০১২ খ্রিস্টাব্দ)."
}

`event_sorting` — 207 items

question, events_shuffled (list of date-stripped event strings), answer_ordered (list of {event, date} in chronological order). This is the only task with no answer column — the gold label is answer_ordered. Items contain 3–6 events (3 events × 107, 4 × 92, 5 × 7, 6 × 1); the event set of answer_ordered always equals events_shuffled, and no item is presented pre-sorted.

{
  "id": "event_sorting_0001",
  "type": "event_sorting",
  "question": "Arrange these events in chronological order (earliest first).",
  "events_shuffled": [
    "The government introduced specific duty exemptions for local cellular phone manufacturing.",
    "The government established the framework for duty exemptions on specific machinery and equipment.",
    "The government adjusted duty rates from 3% to 2% for specific goods.",
    "The government implemented a conditional exemption framework for goods exceeding 5% duty."
  ],
  "answer_ordered": [
    {"event": "The government established the framework for duty exemptions on specific machinery and equipment.", "date": "2010-06-10"},
    {"event": "The government adjusted duty rates from 3% to 2% for specific goods.", "date": "2013-06-06"},
    {"event": "The government implemented a conditional exemption framework for goods exceeding 5% duty.", "date": "2015-06-04"},
    {"event": "The government introduced specific duty exemptions for local cellular phone manufacturing.", "date": "2017-06-01"}
  ]
}

`perturbation_detection` — 498 items

statement, question, answer, is_perturbed (bool), perturbed_fact, correct_fact.

{
  "id": "perturbation_detection_0001",
  "type": "perturbation_detection",
  "statement": "As of the 2011-06-09 Baggage Rules, customs officers are permitted to perform random scanning and examination of up to 10% of passengers passing through the Green Channel.",
  "question": "Is this statement accurate? If not, identify and correct the error.",
  "answer": "The statement is inaccurate. According to the 2011-06-09 Baggage Rules, customs officers are authorized to scan and examine up to 5% (five percent) of passengers passing through the Green Channel on a random basis.",
  "is_perturbed": true,
  "perturbed_fact": "10%",
  "correct_fact": "5%"
}

`scenario_short_answer` — 562 items

scenario, question, answer, date_or_instrument (the instrument the answer is grounded in).

{
  "id": "scenario_short_answer_0001",
  "type": "scenario_short_answer",
  "scenario": "In late 2015, a Bangladeshi industrial manufacturer, Bengal Copper Ltd., is importing various raw copper materials. The customs clearing agent is calculating the newly imposed Regulatory Duty (RD) under a statutory order issued on November 29, 2015. They are importing copper plates under H.S. Code 7409.11.00 and copper tubes under H.S. Code 7411.10.00.",
  "question": "According to S.R.O. No. 353-AIN/2015/59/Customs, what are the respective Regulatory Duty (RD) rates that must be applied to these imported copper plates and copper tubes?",
  "answer": "10% for copper plates and 15% for copper tubes.",
  "date_or_instrument": "S.R.O. No. 353-AIN/2015/59/Customs (2015-11-29)"
}

`temporal_scenario_short_answer` — 363 items

scenario, question, answer, clauses_involved (list of source instrument references).

{
  "id": "temporal_scenario_short_answer_0001",
  "type": "temporal_scenario_short_answer",
  "scenario": "An importer of animal feed and veterinary products has been tracking customs duty exemptions under SRO 178-AIN/2012 (issued on June 7, 2012). On June 6, 2013, the government issued two notifications: SRO 147-AIN/2013 (which repealed SRO 172-AIN/2008) and SRO 165-AIN/2013.",
  "question": "How did SRO 165-AIN/2013 update the framework of SRO 178-AIN/2012 to align with the newly introduced animal feed and veterinary products exemption under SRO 147-AIN/2013?",
  "answer": "SRO 165-AIN/2013 replaced clause (12) of SRO 178-AIN/2012 to specify that it now applies to goods exempted from import duty under the newly issued SRO 147-AIN/2013 (dated 06/06/2013).",
  "clauses_involved": ["S.R.O. 147-AIN/2013 (2013-06-06)", "S.R.O. 165-AIN/2013 (2013-06-06)"]
}

`relative_time_qa` — 463 items

question, answer, seed_time (the anchor date, as written in the question), relative_offset (the offset phrase to resolve).

{
  "id": "relative_time_qa_0001",
  "type": "relative_time_qa",
  "question": "Seven years after the initial framework for duty exemptions on specific machinery was established on June 10, 2010, the government issued a new SRO on June 1, 2017, targeting local computer manufacturing. What is the number of this 2017 SRO, and what is the threshold rate above which import duties are exempted?",
  "answer": "Seven years later, on June 1, 2017, the government issued S.R.O. No. 131-AIN/2017/17/Customs. Under this instrument, for local computer manufacturing, an exemption is granted on the import of specified goods from import duties that exceed 1% (one percent) ad valorem, alongside a full exemption of regulatory duties, value-added tax, and supplementary duties (if any).",
  "seed_time": "June 10, 2010",
  "relative_offset": "seven years later"
}

`context_misalignment` — 116 items

misaligned_context (a fluent but factually false preamble), question, answer, conflict (an explicit statement of how the preamble contradicts the source text). The model must answer from the statutory text and reject the preamble.

{
  "id": "context_misalignment_0001",
  "type": "context_misalignment",
  "misaligned_context": "In an effort to boost local high-tech industries, the government issued S.R.O. 131-AIN/2017 on June 1, 2017. This statutory order provided significant tax relief for local computer manufacturing, granting exemptions on regulatory duties, VAT, and supplementary duties, while capping the maximum payable import duty by exempting any amount exceeding 5% (five percent) of the value-based rate.",
  "question": "According to the official statutory text of S.R.O. 131-AIN/2017, what is the actual percentage threshold above which the import duty is exempted for materials used in local computer manufacturing?",
  "answer": "According to Clause [17], S.R.O. 131-AIN/2017 specifies that the exemption applies to import duties that exceed 1% (one percent) of the value-based rate, not 5%.",
  "conflict": "The misaligned context states that the import duty exceeding 5% is exempt, whereas the actual statutory order (S.R.O. 131-AIN/2017) exempts import duty exceeding 1% (one percent)."
}

`open_ended` — 487 items

question, answer. Free-form explanations of policy rationales, scored by the judge council.

Evaluation protocol

Models are tested under three access conditions:

Parametric — closed-book knowledge
In-Context Learning — gold documents provided
Retrieval-Augmented Generation — top-10 retrieved clauses

Scoring uses an LLM-as-a-Council approach with three judges; inter-rater reliability is Fleiss' κ = 0.83–0.86. A hard date gate means an incorrect date yields zero credit on single-date items.

Key findings

Best macro accuracy: 68.5% (GPT-5.2 under gold-context conditions)
Models struggle most with context misalignment (27.6% accuracy)
Detecting that a provision was altered (78.9%) far exceeds identifying what changed (18.6%)
Temporal reasoning, not content retrieval, is the primary bottleneck

Verification

Two independent subject-matter experts verified all pairs; 94.89% were rated factually correct (Cohen's κ = 0.91).

Source documents

This repository contains the QA benchmark. The underlying source material — original gazette PDFs and their OCR output for the Acts, General Orders, SROs, and Rules — lives in the project repository: https://github.com/icsetepa44/TIDE.

The data files here are regenerated from the source release by scripts/prepare_hf.py, and scripts/validate_hf.py checks that this card's declared schema and counts still match the data.

Intended use and limitations

Research only. Official gazettes remain authoritative for any compliance matter.
The authors caution against deploying LLMs in regulatory workflows without expert safeguards.
Coverage is Bangladesh customs instruments; findings should not be assumed to transfer to other jurisdictions or document families.
All documents are official public records and contain no personally identifiable information.

License

MIT. See LICENSE.

Citation

If you use TIDE, please cite the paper:

@misc{sobhani2026timepresenttimepast,
      title={Time Present and Time Past: Benchmarking Large Language Models on Temporally Evolving Document Understanding},
      author={Mahbub E Sobhani and Md. Faiyaz Abdullah Sayeedi and Fahmid Hasan Chowdhury and Md Adnan Arefeen and Farig Sadeque and Md. Faizul Bari and Swakkhar Shatabda},
      year={2026},
      eprint={2608.08512},
      archivePrefix={arXiv},
      primaryClass={cs.AI},
      url={https://arxiv.org/abs/2608.08512},
}

Downloads last month: 25

Total file size:

2.55 MB

Paper for mahbubhimel/TIDE

Time Present and Time Past: Benchmarking Large Language Models on Temporally Evolving Document Understanding

Paper • 2608.08512 • Published 2 days ago

You need to agree to share your contact information to access this dataset

TIDE: Temporal Document Understanding Benchmark