SEA-Instruct-2602

 # SEA-Instruct-2602 ## Overview SEA-Instruct-2602 is a preliminary release of instruction-tuning data focused on Southeast Asian languages and contexts. The dataset combines prompts filtered from open-source data with our own synthetic prompts, paired with synthetic responses, for language model training on SEA-specific tasks and languages. This dataset contains only the filtered subset of data with `prompt_input_quality` at `Excellent`, `prompt_is_coherent` at `True` and `prompt_is_natural` at `True`. The column `prompt_translated_from_english` reflects whether the prompts have been translated from our internal English set, in which case the tags reflect the original English prompt. - **Funded by:** Singapore NRF - **Shared by:** AI Products Pillar, AI Singapore - **Language(s):** English, Indonesian, Vietnamese, Thai, Tamil, Burmese, Filipino, Burmese, Malay, Khmer, Lao, Chinese - **License:** [ODC-By](https://opendatacommons.org/licenses/by/1-0/) ### Data Pipeline | Stage | Model | Configuration | |-------|-------|---------------| | **Prompt Tagging** | Qwen/Qwen3-235B-A22B | Non-thinking mode, VLLM with structured outputs | | **Response Generation** | Qwen/Qwen3-32B | Non-thinking mode | | **Response Revision** | DeepSeek-V3.1 | Thinking mode | ## Dataset Schema ### Metadata Fields | Column | Type | Description | |--------|------|-------------| | `conversations_id` | string | Unique identifier for each conversation | | `source` | string | Original data source of the prompt | | `subset_or_split` | string | Dataset subset or split designation | | `system_present` | boolean | Whether a system prompt is included in the conversation | | `turn_count` | integer | Number of turns in the conversation | | `conversations` | string | The conversation content as a list of message objects, casted to string | | `prompt_tagging_model` | string | Model used to generate prompt annotations | ### Language Fields | Column | Type | Description | |--------|------|-------------| | `prompt_primary_language` | string | The most prominent language by structure and vocabulary in the user's question. Determined by the actual question content, not surrounding context (e.g., an Indonesian question about an English passage → Indonesian). | | `prompt_language_category` | string | Higher-level language grouping of `primary_language`. Groups script variants under a single category (e.g., Javanese in native script and Latin script are both categorized as `Javanese`). | ### Prompt Quality & Structure | Column | Type | Description | |--------|------|-------------| | `prompt_is_coherent` | boolean | `true`: Logically and linguistically structured, even if factually incorrect.| | | | `false`: Random, garbled, incomplete, or meaninglessly repetitive. | | `prompt_is_natural` | boolean | `true`: Native-like, idiomatic phrasing with natural code-mixing. | | | |`false`: Awkward grammar, unnatural code-mixing, or transliterations. | | `prompt_formality` | string | Language register of the prompt. | | | | Values: `Formal` (polite, official, honorifics), | | | | `Informal` (casual, colloquial, slang), | | | | `Mixed` (contains both). | | `prompt_input_quality` | string | Overall quality rating based on clarity, specificity, and coherence. | | | | Values: `very_poor`, `poor`, `average`, `good`, `excellent`. | ### Task Classification | Column | Type | Description | |--------|------|-------------| | `prompt_primary_domain` | string | The single most relevant topic domain for the prompt. See [Domain Values](#domain-values) for full list. | | `prompt_primary_task` | string | The core action type requested by the user. See [Task Values](#task-values) for full list. | ### Complexity & Specificity | Column | Type | Description | |--------|------|-------------| | `prompt_complexity` | integer | Difficulty level from 1–10, calibrated against academic and programming contests (AMC, AIME, IMO, LeetCode, Codeforces, IOI). | | | | `1` = direct recall; | | | | `10` = research-level/paradigm-shifting. | | `prompt_ambiguity_level` | string | Clarity of the prompt. | | | | Values: `Not_Ambiguous`, `Slightly_Ambiguous`, `Ambiguous`, `Severely_Ambiguous`. | | `prompt_instruction_specificity` | string | How prescriptive the prompt's instructions are regarding output format, structure, and style. | | | | Values: `Very_Low`, `Low`, `Medium`, `High`, `Very_High`. | ### Knowledge Requirements | Column | Type | Description | |--------|------|-------------| | `prompt_requires_external_knowledge` | boolean | `true` if answering requires information beyond standard LLM training data (e.g., post-cutoff events, real-time data, recent news). | | `prompt_requires_local_cultural_knowledge` | boolean | `true` if answering requires specialized knowledge of local SEA culture, customs, or social dynamics that a non-local would lack. | ### Safety & Sensitivity | Column | Type | Description | |--------|------|-------------| | `prompt_sensitivity` | string | Risk classification. | | | | Values: `Safe` (no concerns), `Caution` (potentially sensitive topics), `Harmful_Adversarial` (jailbreak/injection attempts), `Harmful_NonAdversarial` (dangerous but naive intent). | ### Geographic Scope | Column | Type | Description | |--------|------|-------------| | `prompt_region_scope` | string or list | Geographic relevance of the prompt. Multiple values will be listed for cross-border topics.| | | | Values include specific SEA countries (`Singapore`, `Indonesia`, `Malaysia`, `Thailand`, `Vietnam`, `Philippines`, `Myanmar`, `Cambodia`, `Laos`, `Brunei`, `Timor_Leste`), other regions (`East_Asia`, `South_Asia`, `Europe`, etc.), `Global`, `No_Specific_Region`, or `Other_Region`. | ## Appendix ### Domain Values `Government_and_Politics` | `Law_and_Justice` | `Legal_Rights_and_Access` | `Human_Rights_and_Advocacy` | `Violence_and_Safety` | `Social_and_Cultural_Issues` | `Superstitions_Myth_and_Folklore` | `Religion_and_Belief` | `Physical_and_Sexual_Health` | `Mental_Health_and_Wellbeing` | `Traditional_Medicine_and_Alternative_Healing` | `Education` | `History_and_Heritage` | `Arts_and_Literature` | `Business_and_Economy` | `Employment_and_Workplace` | `Finance_and_Insurance` | `Shopping_and_Consumer_Affairs` | `Food_and_Cuisine` | `Agriculture_and_Fishing` | `Travel_and_Tourism` | `Environment_and_Nature` | `Environment_Climate_Change_and_Disasters` | `Animals_and_Pets` | `Science_and_Technology` | `Technology_Use_and_Devices` | `Transport_and_Infrastructure` | `Real_Estate_and_Housing` | `Media_and_Entertainment` | `Sports_and_Fitness` | `Military_and_Security` | `Parenting_and_Family` | `Daily_Life_and_Personal` | `Internet_and_Digital_Culture` | `Gaming` | `Migrant_Worker_Expat_and_Student_Life` | `Scams_and_Fraud` | `Fashion_and_Beauty` | `Gambling_and_Betting` | `Love_Dating_and_Relationships` | `Drug_Use_and_Substance_Abuse` | `No_Specific_Domain` | `Other` ### Task Values `Factual_Question_Answering` | `Reasoning_and_Explanation` | `Summarization` | `Classification_and_Moderation` | `Information_Extraction` | `Translation_and_Code_Switching` | `Math_and_Scientific_Problem_Solving` | `Programming_and_Code` | `Creative_Writing_and_Generation` | `Structured_List_or_Template_Generation` | `Text_Editing_and_Improvement` | `Recommendation_and_Advice` | `Chit_Chat_and_Conversational` | `Role_Play_or_Persona_Simulation` | `Function_Calling` | `No_Specific_Task` | `Other` ### Complexity Scale Reference | Level | Label | General Description | |-------|-------|---------------------| | 1 | Very Easy | Direct recall, trivia, single-word answers | | 2 | Easy | Single-step reasoning, basic calculations | | 3 | Medium | Multi-step but routine, familiar patterns | | 4 | Hard | Layered reasoning, contextual analysis | | 5 | Very Hard | Cross-domain synthesis, specialized knowledge | | 6 | Advanced | Model-building, significant abstraction | | 7 | Expert | International olympiad standard, new typologies | | 8 | Extremely Advanced | Multi-method expert solutions, contest "legendary" | | 9 | Elite | Research-level, top 1% globally | | 10 | Legendary | World-class unsolved problems, paradigm shifts | ## More Information This is the repository for the dataset that can be used to instruction-tune a model, gathered from open sources. The dataset has not been aligned for safety. Developers and users should perform their own safety filtering and related security measures. In no event shall the authors be held liable for any claims, damages, or other liabilities arising from the use of the released data. For more info, please contact us at [[e-mail]](mailto:[e-mail]) ## Acknowledgement This project is supported by the National Research Foundation Singapore and Infocomm Media Development Authority (IMDA), Singapore under its National Large Language Model Funding Initiative. ## Contact [[e-mail]](mailto:[e-mail])