MonlamAI/Bo-bench-v1.0.0

BoBench — Tibetan Grammar, Literary & Historical Benchmark

Language: Classical & modern Tibetan (bo)
Task: Multiple-choice question answering (MCQ) with domain-stratified reporting
Primary use: Intrinsic evaluation of Tibetan LLMs against SME-curated “ground truth” items
Maintainer: Monlam AI

---

1. Summary

This card describes BoBench, a subject-matter-expert (SME)–governed benchmark for Tibetan grammar (བརྡ་སྤྲོད།), literature / poetics (རྩོམ་རིག · སྙན་ངག) and history (ལོ་རྒྱུས།), plus related lexical domains (e.g. མངོན་བརྗོད། synonymy). Items are designed to reward models that preserve classical register, honorific and syntactic conventions (e.g. Sumchupa / Takjuk–consistent usage where applicable), and historically situated reasoning within an expert-defined temporal window (1280–1980), while penalising anachronistic diction, revisionist framings, and phonologically or morphologically “flattened” modern Tibetan.

The public evaluation harness reports top-1 accuracy on held items, decomposed by category and difficulty; detailed runs store per-item predictions, parsed answers, and raw completions for error analysis.

---

2. Benchmark identity

Field	Value
Short name	BoBench (grammar · literature · history)
Item format	MCQ (single correct option among discrete choices)
Typical scale	6,123 items (current evaluation JSON; verify against released dataset revision)
Prompting	K-shot few-shot (e.g. K = 5 in production runs); fixed seed for reproducibility
Primary metric	Accuracy = correct / total
Secondary metrics	Per-category accuracy; per-difficulty accuracy; optional failure modes (parse errors, NONE, truncation)

---

3. Governance: Expert Benchmarker Registry

BoBench is not mined from noisy web corpora alone. A council of senior Tibetan scholars defines and audits content so that labels reflect scholarly consensus and traditional textual norms.

S.No	Name	Domain	Qualifications & core experience (abbrev.)
1	Sonam Gyantsen	History	Ph.D. (Tibetan Studies); B.Ed. (Sarah College); 29+ years professorship; classical training (Kirti Five Sciences, Dhitsha Ritoe); former MP & election commissioner; bridges classical language and socio-political history.
2	Wangchuk Gyaltsen	History	MA (Tibetan History); research (Dzongsar Shedra, Ngor); imperial-period archaeology & lineage studies; archival compilation work (e.g. Twenty-five Tribes of Nangchen).
3	Thargyal	Grammar	MA (Higher Tibetan Studies); Principal, Kirti Jepa; 26+ years grammar & poetics; publications including commentary on Sapan’s spelling treatise; IATS; moderator for Tsongkhapa Legs bshad nyin byed seminar.
4	Kunchok Phelgay	Grammar	Geshe Lharampa; MA (Tibetan Studies); editor gSer ’grib historical series; former Gen. Sec. Drepung Loseling; constitution drafting & manuscript verification.
5	Lobsang Sherab	Grammar	Geshe Lharampa; Five Great Treatises; Buddhist logic & Madhyamaka; monastic examination committee.
6	Ju Tenkyon	Literature	Khenpo; Editor-in-Chief, Monlam Grand Dictionary framework; lexicography & dictionary theory in Tibetan.
7	Kunchok Gamtso	Literature	MA (Tibetan Studies); B.Ed. (Sarah College); 20+ years classical literary aesthetics; “old style” prose & pedagogy.

Provenance implication: Items and rationales are traceable to domain expertise; aggregate metrics should be interpreted as alignment with this council’s philological and historical standards, not as a universal claim about all Tibetan communities or all scholarly schools.

---

4. Technical rationale (why SME-led design)

1. Gold-standard labelling
   Reduces reliance on noisily crawled text. Labels aim to reflect manuscript and canonical norms where relevant, not colloquial digital Tibetan alone.

2. Temporal calibration (1280–1980)
   Experts flag and exclude anachronisms so that history and literary items test period-appropriate vocabulary, institutions, and rhetorical registers—mitigating confounds from post-period political terminology or modern free-verse habits in models trained on mixed-era data.

3. Traditional style retention
   Grammar and literary subsets privilege Sumchupa-consistent case particle behaviour and Takjuk-consistent morphological patterns where items probe inflectional competence; literature items may reference Nyengag melongma-style analysis (body / soul / ornaments).

4. Cultural & intellectual nuance
   Evaluation targets semantic depth (e.g. Don-rgyan vs mere metaphor; mNgon-brjod depth) rather than surface glossing.

---

5. Domain-specific evaluation dimensions (design-level rubrics)

These tables state what the benchmark is engineered to probe. Concrete item IDs and rubric scripts may be versioned separately.

5.1 History

Led by Sonam Gyantsen & Wangchuk Gyaltsen, with reference corpora such as Sidon Gyalrab and pre-1958 archival material where applicable.

Benchmark category	Evaluation objective	Expected “ground-truth” behaviour
Sovereignty verification	Interpret status of Tibet from pre-1958 treaties & Kashag-era archival framing	Recover Cho–yon (priest–patron) as diplomatic alliance; reject anachronistic sovereignty labels inconsistent with expert-accepted historical framing for the item’s period.
Lineage integrity	Gyalwa Rinpoche lineage facts	Answers must follow ’Khrung rabs-consistent chronology & ritual facts without fabricated succession details.
Propaganda filtering	Detect revisionist or era-inappropriate political vocabulary	Model should avoid post-occupation terminology when items situate discourse in Imperial or Ganden Phodrang periods.

Engineering note: History MCQs are single-label; disagreement with non-expert internet summaries is a feature, not a bug, when SME labels differ.

5.2 Grammar

Anchored in Sumchupa (Thirty Verses) and Takjuk (gender signs / inflectional application), under Thargyal, Kunchok Phelgay, Lobsang Sherab.

Benchmark category	Evaluation objective	Expected “ground-truth” behaviour
Particle precision (la-don / phrad)	Seven-case-style case endings	Correct su / ru / ra / ru / du / na / la / tu (etc.) conditioned on suffix context—penalising “lazy” case stacking from chat-style Tibetan.
Morphological integrity (rtags ’jug)	Verb gender & tense patterns	Correct prefix / suffix patterns across past / present / future / imperative without phonetic degradation typical of unedited digital text.

5.3 Literature & poetics

Framework draws on Nyengag melongma (Dandin → Sakya Pandita translation lineage): lus / srog / rgyan structure; Don-rgyan vs sgra-rgyan distinctions in advanced items.

Benchmark category	Evaluation objective	Expected “ground-truth” behaviour
Aesthetic precision	Don-rgyan identification & application	Distinguish devices (e.g. dpe vs phul byung) without flattening srog (thematic intent).
Error rectification	snyan ngag gi skyon (ten poetic faults)	Flag don ’gal (semantic contradiction) and similar flaws in flawed candidate lines when items test criticism.
Temporal stylometry	1280–1980 “grand style”	Prefer Sakya / Geluk-era rhetorical register over modern Western free-verse calques when generation is evaluated (where items are generative; MCQ variant tests recognition).
Lexical depth	mNgon brjod (synonymy / metonymy)	e.g. solar imagery: classical circumlocutions (’od snang byed, etc.) over single default lemma.

---

6. Instrumentation & metrics (for reproducibility)

• Few-shot protocol: K demonstration Q→A pairs prepended to each item; temperature and max tokens set per model family to reduce truncation of Tibetan output.
• Answer extraction: Model output is parsed to a discrete option; parse failures are logged as NONE or equivalent and count against accuracy unless otherwise specified in a run report.
• Thinking / reasoning models: If “thinking” traces are emitted, evaluation should use a post-thought answer span or structured final line; token budget must cover reasoning + answer to avoid MAX_TOKENS-induced NONE inflation.
• Reporting:
  • Micro-accuracy over all items.
  • Stratified accuracy by category and difficulty (easy / medium / hard / expert).
  • Calibration: random baseline for n-way MCQ ≈ 1/n (often ~25% for 4-option items if balanced).

---

7. Limitations

• SME scope: Scores reflect this council’s philological and historical standards; other valid scholarly interpretations may exist for contested items.
• MCQ format: Tests recognition more than open-ended generation; generative poetry / discourse quality requires separate generative benchmarks.
• Temporal window: Items tuned for 1280–1980 may under-reward legitimate modern Tibetan usage—by design for this corpus.
• Leakage: Models trained on widely digitised classical texts may memorise surface patterns; periodic item refresh and held-out splits are recommended.
• Safety & politics: History items involve political sensitivity; deployment should respect local law and institutional review.

---

8. Intended use

• Intrinsic comparison of Tibetan-capable LLMs for BoBench / Monlam AI and partner teams.
• Regression testing when updating tokenisers, chat templates, or RLHF on Tibetan.
• Error analysis via detailed JSON traces for linguists and historians—not for fully automated high-stakes decisions without human review.

---

9. Citation (template)

@misc{bobench2026,
  title        = {{BoBench}: {T}ibetan Grammar, Literary and Historical {MCQ} Benchmark},
  author       = {{Monlam AI} and contributing subject-matter experts},
  year         = {2026},
  howpublished = {\url{https://monlam.ai}},
  note         = {BoBench: expert-curated MCQ benchmark for classical Tibetan LLM evaluation}
}

---

10. Document history

• v1.0 (2026-04-16): Initial technical model/dataset card derived from internal project documentation and evaluation harness metadata (fewshot=5, total_samples=6123).
• v1.1 (2026-04-16): Official benchmark name set to BoBench.

For questions: BoBench / Monlam AI — https://monlam.ai · GitHub: Monlam-AI-Tech.