GomParam-v1: The Goan Konkani Benchmark

GomParam-v1 is a handcrafted, linguistically rigorous benchmark for evaluating Large Language Models on Goan Konkani — a morphologically complex, low-resource Indic language with 450 years of Portuguese influence, multiple scripts, and distinct regional dialects.

The benchmark spans 22 modules and 3,332 items, covering everything from surface grammar to deep pragmatic inference, spatial reasoning, nuanced sentiment, cross-script transfer, and a unique multi-level linguistic diagnostic framework. All items are human-authored by native speakers. There is zero Marathi or Hindi contamination.

---

📊 Benchmark at a Glance

Metric	Value
Total Modules	22
Total Items	3,332
ADI Sub-Questions	1,092 (across 182 ADI items)
Difficulty: Basic	1,206
Difficulty: Intermediate	1,072
Difficulty: Advanced	872
Generative Templates Used	0 — 100% human-authored
Marathi / Hindi Contamination	0%
Sources	authentic_goan, expert_constructed
Scripts Covered	Devanagari, Romi (Latin)

---

🧩 The 22 Modules

1. ADI — Artisan Diagnostic Instrument (adi.json — 182 items, 1,092 sub-questions)

The ADI is GomParam's flagship module and the most structurally unique benchmark component. Each of the 182 items is a multi-level interrogation tree — not a flat multiple-choice question, but a cascaded probe across three levels:

• L0 — The trigger sentence (deliberately short, 4–7 tokens). The model identifies what is wrong or right.
• L1_correct / L1_incorrect — Given the L0 answer, the model must explain why it is correct or why a wrong option is wrong. Tests rule identification and feature attribution.
• L2_correct_correct / L2_correct_incorrect / L2_incorrect_correct — Counterfactual reasoning: given a structural change to the sentence, predict what changes and why.

This design isolates a single linguistic variable per item, strips away all contextual crutches, and forces the model to demonstrate genuine grammatical competence — not surface pattern matching.

The 17 ADI sub-categories:

Sub-Category	Items	What It Tests
grammatical_acceptability	15	Verb-object gender/number agreement, ergative marking
minimal_pairs	13	Meaning differences from single-phoneme or morpheme substitutions
native_naturalness	13	Whether a sentence sounds like a native speaker wrote it
hidden_marathi	9	Detecting Marathi loanwords disguised within Konkani syntax
focus_emphasis	11	Particle placement and topic/focus distinction
presupposition	11	Identifying presuppositions that survive negation
ellipsis_recovery	10	Recovering the elided element from VP or NP ellipsis
social_inference	10	Register-appropriate responses in culturally specific social scenarios
error_severity	10	Ranking grammatical errors from minor to unacceptable
garden_path	10	Detecting and resolving garden-path sentence structures
selectional_preference	11	Verb-argument semantic compatibility (e.g., विचार पियेता = unacceptable)
scope_ambiguity	10	Quantifier and negation scope resolution
structural_ambiguity	11	PP and NP attachment ambiguity resolution
dialectal_variation	10	Distinguishing Antruz, Saxtti, and Mangalorean forms
false_friends	9	Words that look like Marathi/Hindi cognates but have different Konkani meanings
discourse_connectives	10	Correct use of causal, contrastive, and additive connectives
cultural_script	9	Culturally expected responses in Goan social scenarios

Design note: The ADI's short trigger sentences (4–7 tokens) are a deliberate, principled choice — not a limitation. Because the ADI is a targeted linguistic diagnostic tool (not a reading comprehension test), short sentences isolate exactly the variable being tested. A 50-token paragraph introduces confounds; a 5-token sentence with a single agreement error tests agreement and nothing else.

---

2. Homograph Disambiguation (homograph_disambiguation.json — 150 items)

Identical Devanagari spelling, completely different meanings depending on context.

• वात: "हांव मडगांवा वचपाची वात सोदतालो" (Path) vs "देवाऱ्यांतली वात पेटयली" (Lamp wick)
• रेंद: "तिचें कुडें रेंद पळोवन..." (Lace) vs "ह्या म्हयन्याचें रेंद मागलां" (Rent)
• मान: "जाणट्यांचो मान राखपाक जाय" (Respect) vs "ल्हव चल, मान मोडटली" (Neck)
• शीर: Riverbank / Vein / A slice — three meanings, same spelling

---

3. Code-Switching (code_switching.json — 150 items)

Goan Konkani carries 450 years of Portuguese administrative vocabulary still in active everyday use. This module goes far beyond street slang:

• Inventário — inheritance/property division after a death
• Comunidade — the ancient Goan village community administration system
• Foral — an ancient royal decree on land rights
• Escritura — a legal property sale deed
• Mutação — transfer of names in property records

---

4. Pragmatic Inference & Sarcasm (pragmatics.json — 150 items)

Tests whether the model understands what a Goan means — not what they say.

• Sarcasm: A mother tells a child who scored zero: "व्हय व्हय, तूं सामको हुशार!" → The model must identify this as contempt, not praise.
• Indirect refusal: "पळोवया." (Let's see) in response to a loan request → polite evasion, not a promise.
• Implicature: "तुज्या हातान देवूय जेवचो ना." (Even God wouldn't eat your cooking) → the food is terrible.

---

5. Kinship Terminology (kinship.json — 150 items)

Goan kinship terms encode gender, paternal/maternal lineage, and birth order simultaneously. Hindu and Catholic communities use different terms for the same relationships.

• Father's sister = पाची | Mother's sister = मावशी
• Mother's brother = माम | Father's brother = काका
• Catholic godparenting roles: Padrinho / Madrinha, and the compadre relationship between biological and godparents

---

6. Idioms & Proverbs (idioms_proverbs.json — 150 items)

Deep agrarian and fishing metaphors that modern machine translation systems almost always render literally and incorrectly.

• "लासील्ली कोंबी उज्याक भियेता" — A burnt hen fears the fire → Once bitten, twice shy
• "नवें नवल, णव दीस" — A new wonder lasts nine days → Excitement over novelty is always temporary
• "खाता तेन्ना गोड, हागता तेन्ना कोडू" — Sweet when eating, bitter when shitting → Fun now, consequences later

---

7. Dialect Syntax (dialect.json — 150 items)

Konkani is not one language — it is a dialect continuum. This module tests structural, syntactic, and vocabulary differences between three major dialects:

• Antruz (Ponda / Hindu interior)
• Saxtti (Salcete / Catholic coastal)
• Mangalorean

Tests include: negation patterns, interrogative particle differences, pronoun forms, and vocabulary divergence (e.g., different words for "food" across dialects).

---

8. Cross-Scripting (cross_scripting.json — 150 items)

Konkani is uniquely written in five scripts. This module focuses on bidirectional transfer between Romi (Latin script) and Devanagari.

• WhatsApp colloquial: "Bhai, tuka kitem zalam? Kityak msg korna?" → Devanagari
• Formal tiatr/church text: "Tiatr suru zata. Sogleamni aplim boskam gheun xant ravchem." → Devanagari

---

9. Entailment / NLI (entailment.json — 150 items)

The Konkani equivalent of XNLI/MNLI. Given a premise and hypothesis in Goan Konkani, determine: Entailment, Contradiction, or Neutral.

• Premise: "गोंयांत आयज खूब पाऊस पडला."
• Hypothesis: "गोंयांत आयज हवामान सामकें कोरडें आशिल्लें."
• Answer: Contradiction

---

10. Sentence Coherence (coherence.json — 150 items)

Scrambled paragraph reconstruction. The model must understand temporal logic, causal chains, and discourse markers in Konkani to reorder sentences into a coherent passage.

---

11. Cultural Grounding (cultural_grounding.json — 150 items)

Deep knowledge of Goan culture, history, and local facts that no translation model can fake from Hindi/Marathi training data.

• Occupational terms: Poder (Baker), Render (Toddy Tapper), Kharvi (Fisherman)
• Culinary knowledge: ingredients and history of Bebinca, Xacuti, Sanna
• Historical events: Kadamba dynasty, Portuguese inquisition, Liberation of Goa

---

12. Morphology (morphology.json — 150 items)

Konkani's notoriously complex morphological structures:

• Ergative case alignment — the subject takes the ergative postposition in transitive past tense
• Double-causative verbs — e.g., koroitai (to cause someone to cause someone to do)
• Gender agreement — verbs, adjectives, and modals all agree with the grammatical gender of the object (not the subject) in past transitive constructions

---

13. Register Discrimination (register_discrimination.json — 150 items)

Differentiates between highly formal journalistic Konkani (as used in newspapers like Bhaangarbhuin) and relaxed, colloquial street Konkani. The model must identify which register a sentence belongs to — and why.

---

14. Reading Comprehension (para_qa.json — 150 items)

Dense multi-paragraph comprehension over authentic Goan texts covering:

• The Kadamba dynasty and medieval Goan history
• The Portuguese inquisition
• Goan ecology and the Western Ghats
• Contemporary Goan literature and culture

Questions are graded across basic, intermediate, and advanced difficulty levels.

---

15. Semantic Perplexity (perplexity.json — 150 items)

Grammatically perfect sentences that are semantically absurd. The model must detect the violation.

• "वाऱ्यान मुसांक शाळेंत व्हेलें आनी तांकां गणित शिकोवन घरा धाडलें." — The wind took the flies to school and sent them home after teaching them math.

---

16. Contextual Cloze (cloze.json — 150 items)

Fill-in-the-blank tasks targeting:

• Precise case marking (postposition selection)
• Neuter/feminine gender agreement in verb and adjective forms
• Tense and aspect consistency within a passage

---

17. Mixed General (mixed_general.json — 150 items)

A diverse cross-domain set of general knowledge questions in authentic Goan Konkani — geography, science, civics, and everyday life — testing broad language understanding without module-specific scaffolding.

---

18. Coreference Resolution (coreference.json — 150 items)

Resolving ambiguous pronouns in Konkani: तो, ती, ताका, तिणें, तें — based on context, common sense, and Goan social conventions.

• "राम आनी लक्ष्मण वनांत गेले. ताणें रावणाक मारलो." — Who does ताणें refer to?

---

19. Numerical Word-Math Reasoning (numerical_reasoning.json — 150 items)

Arithmetic word problems written entirely in Konkani text, using the highly irregular Konkani numeral system: एक, दोन, तीन... इकरा, बारा, तेरा... एकोणीस, वीस, एकवीस...

• "शेतकारान पयल्या दिसा धा नाल्ल काडले आनी दुसऱ्या दिसा पंदरा काडले. वट्ट किते जाले?" → पंचवीस

---

20. Sentiment Analysis (sentiment.json — 150 items)

Goes far beyond simple positive/negative polarity. Four sub-categories:

Sub-Category	What It Tests
ambiguous_sentiment	Sarcasm and irony where surface meaning contradicts true emotion
code_switched_emotion	Emotions expressed through Konkani idioms and figurative language
formal_neutral_vs_negative	Polite/bureaucratic phrasing that is functionally negative
hyperbole	Extreme exaggeration used in everyday Goan speech

Models cannot resolve these from a single adjective — they must parse the pragmatic intent of the full sentence in context.

---

21. Spatio-Temporal Reasoning (spatio_temporal.json — 150 items)

Tests genuine spatial navigation and temporal event sequencing in Goan Konkani. Five sub-categories:

Sub-Category	What It Tests
spatial_basic	Object placement: left/right, above/below, inside/outside
spatial_navigation	Cardinal direction inference with rotation and perspective-taking
temporal_basic	Time-of-day identification, before/after reasoning
temporal_chronology	Multi-step event ordering with transitive inference
spatio_temporal_logic	Combined: infer where an actor ends up after spatial + temporal events

Uses authentic Goan Konkani spatial vocabulary: उदेंत/अस्तंत (East/West), मुखार/फाटीं (front/back), दनपारां/सांजवेळा (afternoon/evening).

---

22. Medical Terminology (medical.json — 150 items)

Localized Goan medical and bodily terms that differ substantially from their Marathi or Hindi equivalents. Includes both folk medicine vocabulary and standard clinical terminology as used by Goan healthcare workers.

• "ताचे आंगार हिवज्वर आयला." → Malaria / cold chills (not a generic fever)

---

🛠️ How to Use

from datasets import load_dataset

# Standard multiple-choice module
ds = load_dataset("omdeep22/GomParam-v1", "sentiment")
print(ds["train"][0])

# ADI multi-level diagnostic module
ds = load_dataset("omdeep22/GomParam-v1", "adi")
item = ds["train"][0]
print(item["L0"])          # Trigger question
print(item["L1_correct"])  # Why is the correct answer correct?
print(item["L2_correct_correct"])  # Counterfactual

---

🗂️ Data Schema

Standard Modules (21 modules)

All 21 standard modules share this schema:

{
  "id": "sentiment_007",
  "context": "ताणें गायलेलें पद आयकून म्हजे कान तृप्त जाले, आतां कुड्डे जावपाचीच वेळ आयल्या.",
  "question": "वक्त्याच्या उलोवपांतलो भाव खंयचो?",
  "candidates": [
    "पद खूब सुराळ आशिल्लें",
    "कान खरेच कुड्डे जाले",
    "पद सामकें बेसूर आशिल्लें",
    "गायन कला शिकपाची इत्सा आसा"
  ],
  "correct": 2,
  "category": "ambiguous_sentiment",
  "difficulty": "intermediate",
  "module": "sentiment",
  "source": "authentic_goan"
}

Field	Type	Description
id	string	Unique identifier with module prefix
context	string	The passage, sentence, or scenario
question	string	What the model must answer
candidates	array[4]	Four options; distractors defeat surface-level matching
correct	int (0–3)	Index of the correct answer
category	string	Sub-category within the module
difficulty	string	basic, intermediate, or advanced
module	string	Parent module name
source	string	authentic_goan or expert_constructed

ADI Module

Each ADI item contains:

{
  "id": "adi_01_01",
  "category": "grammatical_acceptability",
  "source_type": "expert_constructed",
  "reviewer_id": "rev_01",
  "L0": {
    "context": "वाक्य: 'रामान आंबे खाल्लीं.'",
    "question": "ह्या वाक्यांत व्याकरण नदरेन खंयचो दोश आसा?",
    "candidates": ["...", "...", "...", "..."],
    "correct": 1
  },
  "L1_correct": {
    "question": "क्रियापदाचो मेळ चुकला हें कशावयल्यान कळटा?",
    "candidates": ["...", "...", "...", "..."],
    "correct": 2,
    "type": "rule_identification"
  },
  "L1_incorrect": { "...", "type": "feature_attribution" },
  "L2_correct_correct": { "...", "type": "counterfactual" },
  "L2_correct_incorrect": { "...", "type": "feature_check" },
  "L2_incorrect_correct": { "...", "type": "cf_confirm" }
}

---

⚠️ Evaluation Methodology: Surface Form Competition

Read this before reporting numbers.

Standard length-normalized log-probability scoring produces a known evaluation artifact on this benchmark: surface form competition (Holtzman et al., 2021). Because correct Konkani answers are often single-token (e.g., भायर, दाव्याक, पूण) while distractors contain longer Hindi/Marathi-cognate subword sequences, length-normalized scoring systematically assigns higher probability to incorrect longer options.

On GomParam-v1, this manifests as categorical 0% accuracy on Basic-tier Medical items across all evaluated models — a more extreme and consistent instance of this failure than previously documented in higher-resource settings. This is not a model failure; it is a scoring artifact.

We strongly recommend the Domain-Conditional PMI correction (PMI_DC; Holtzman et al., 2021) for all evaluations. PMI_DC substantially recovers signal, concentrated in the medical, sentiment, and spatio_temporal modules. Numbers reported without PMI_DC are artifactually low and not comparable to our published results.

---

🔬 Why GomParam-v1 Is Hard to Fake

An LLM cannot pass this benchmark by:

• Translating from Marathi/Hindi — zero contamination means Marathi words are wrong answers
• Guessing on polarity — sentiment items require pragmatic intent parsing, not keyword spotting
• Applying standard grammar rules — ergative case alignment in Konkani is different from Hindi's ergative marking
• Ignoring script — cross-scripting requires genuine Romi ↔ Devanagari transfer
• Using context to disambiguate — ADI items are designed with minimal context to force rule-based reasoning

A model that scores well on GomParam-v1 has demonstrated genuine Goan Konkani understanding — not Hindi transliteration, not Marathi proximity, but authentic command of the language as spoken and written by the Goan people.

---

📄 Citation

@dataset{gomparam_v1_2025,
  title     = {GomParam-v1: The Goan Konkani Benchmark},
  author    = {Borkar, Omdeep},
  year      = {2025},
  publisher = {Hugging Face},
  url       = {https://huggingface.co/datasets/omdeep22/GomParam-v1},
  license   = {CC-BY-4.0}
}