TIBSTC-CoT: Tibetan Chain-of-Thought Dataset

• TIBSTC-CoT is a large-scale Tibetan dataset featuring explicit chain-of-thought reasoning across five diverse domains.
• It utilizes a four-stage, multi-agent construction pipeline to ensure reproducibility, quality, and balanced domain representation.
• The dataset advances Tibetan NLP by providing interpretable, multi-step reasoning data that supports training robust low-resource language models.

TIBSTC-CoT refers to the "Tibetan Structured Chain-of-Thought" dataset, the first large-scale, multi-domain chain-of-thought (CoT) reasoning corpus constructed in the Tibetan language. Designed to tackle both linguistic resource scarcity and the need for explicit reasoning supervision in LLM training, TIBSTC-CoT enables the development and benchmarking of Tibetan-centric LLMs equipped for interpretable, step-wise reasoning. It also provides empirically grounded evidence for the benefits of CoT supervision in learning complex sequence-to-sequence mappings, particularly for low-resource and typologically distinct languages (Gao et al., 4 Aug 2025).

1. Definition and Scope

TIBSTC-CoT is a corpus of 40,121 Tibetan "question–CoT–answer" triples distributed across five domains: Science & Engineering (8,274), Life Sciences (7,546), Humanities & Social Sciences (8,621), Social Sciences (7,078), and a "Comprehensive" general knowledge domain (8,602). Each entry consists of:

• A Tibetan question, drafted via domain-specific instructions.
• A sequence of explicit intermediate reasoning steps (the CoT).
• The final answer.

The dataset serves two principal objectives:

• Publicly closing the resource gap for Tibetan NLP by offering a scalable, LLM-driven recipe for constructing instruction datasets.
• Enabling and benchmarking multi-step reasoning for Tibetan-centric LLMs, thereby advancing interpretability and factual reliability (Gao et al., 4 Aug 2025).

2. Dataset Construction Pipeline

TIBSTC-CoT employs a four-stage, multi-agent construction methodology designed for scalability and reproducibility:

1. Question Formulation: Claude-3.5-Sonnet generates Tibetan questions from a pool of five domain-balanced instruction templates, with sampling ensuring each domain achieves near-uniform representation.
2. CoT and Answer Generation: DeepSeek-R1 produces both the chain-of-thought (stepwise reasoning $r_1,\dots,r_k$) and the final answer for each question.
3. Automatic Assessment: Gemini-1.5-Flash evaluates each triple on five criteria (coherence, logical consistency, fluency, factuality, cultural appropriateness), retaining only those with scores $\geq 3.5$ (on a 1-5 scale).
4. Manual Verification: Expert annotators perform domain alignment, factual accuracy, and cultural sensitivity checks.

This process is iteratively deployed in 90 rounds, each yielding sets of 10 knowledge questions, 10 generative tasks, and 5 classification tasks per domain (Gao et al., 4 Aug 2025).

3. Domain Distribution and Data Statistics

The dataset is balanced across domains using a uniform sampling strategy ($p(d) = \frac{1}{5}$). Table 1 summarizes the distribution:

Domain	Entry Count	Generative	Know. Q&A	Classif.
Science & Engineering (SE)	8,274	3,308	3,275	1,691
Life Sciences (LS)	7,546	3,062	3,011	1,473
Humanities & Soc. Sci. (HSS)	8,621	3,453	3,426	1,742
Social Sciences (SS)	7,078	2,832	2,810	1,436
Comprehensive (Com)	8,602	3,440	3,468	1,694

Note: No fixed train/val/test split is prescribed; an 80/10/10 partition is suggested as a plausible configuration.

4. Chain-of-Thought Reasoning Patterns

While TIBSTC-CoT does not enumerate an explicit reasoning taxonomy, the corpus encompasses a variety of reasoning types observed in the generated CoTs:

• Arithmetic reasoning (stepwise calculations).
• Deductive/logical reasoning (general-to-specific rules).
• Commonsense/abductive reasoning (plausible inference).
• Factual recall (sequential knowledge retrieval).
• Classification reasoning (category elimination).

Each triple can be formalized as:

$\mathrm{CoT}(q) = (r_1, r_2, \dots, r_k, a)$

where $r_i$ are Tibetan textual steps, and $a$ is the final answer. Examples include explicit breakdowns of multi-operational math problems, categorization tasks, and logical deductions (Gao et al., 4 Aug 2025).

5. Model Training and Evaluation on TIBSTC-CoT

The development of the Sunshine-Thinking LLM family exemplifies the use of TIBSTC-CoT as a resource for Tibetan-centric reasoning models. The training protocol is standard supervised fine-tuning (SFT) with cross-entropy loss:

$$\mathcal{L} = -\sum_{t=1}^T \log P(x_t \mid x_{<t}, q)$$

using AdamW optimizer ($\beta_1=0.9$, $\beta_2=0.999$, $\epsilon=10^{-8}$) and mixed-precision training. Two notable models are:

• Sunshine-Thinking-1.7B: Qwen-3-1.7B backbone, 2 epochs, single L20 GPU.
• Sunshine-Thinking-8B: Qwen-3-8B backbone, 3 epochs, three H20 GPUs.

Performance is benchmarked primarily against the TLUE (Tibetan Language Understanding Evaluation) suite:

• DA (Direct Answer): Measures final answer accuracy.
• CAA (Concern-All Answer): Assesses overall alignment with expected reasoning.

Sunshine-Thinking-8B achieves SOTA results in DA/CAA compared to Claude-3.7-Sonnet and GPT-4.1, e.g., achieving 36.25% China-DA, 39.00% SS-DA, 41.00% HSS-DA, and 40.71% STEM-DA. Significance is confirmed by paired-bootstrap testing ($\geq 3.5$0). On FLORES-200 translation tasks (chrF++), the model remains competitive with baseline LLMs despite being smaller in scale (Gao et al., 4 Aug 2025).

6. Significance and Broader Implications

TIBSTC-CoT constitutes the first public, large-scale CoT dataset for a low-resource language, establishing a reproducible paradigm for resource creation in typologically diverse settings. Its explicit chain-of-thought annotation introduces a framework for both:

• Improving LLM reasoning interpretability and reliability, critical for education, QA, and dialogue applications in Tibetan.
• Addressing broader language equity in NLP by bridging the digital resource gap for underrepresented languages.

A plausible implication is that similar CoT-centric dataset construction pipelines could be adapted for other low-resource languages, enabling inclusive AI development and catalyzing multilingual, domain-spanning research.

All data and checkpoints are released at https://github.com/Vicentvankor/sun-shine, facilitating further experimentation and extension (Gao et al., 4 Aug 2025).