Datasets:

sparkthai
/

sparkthai-model-arena

id string	round string	title dict	prompt string	promptLang string	purpose list	passWhen dict	test string	config list	foreignCheck bool	rubric string	rubricNote dict	hardGate bool	addedInV3 bool	addedInV5 bool	variants list	passScore int64
code-fizzbuzz	code_exact	{ "de": "FizzBuzz", "en": "FizzBuzz", "fr": "FizzBuzz", "ja": "FizzBuzz", "ko": "FizzBuzz", "th": "FizzBuzz", "vi": "FizzBuzz", "zh": "FizzBuzz" }	Write a Python function `fizzbuzz(n)` that returns a list of strings for 1..n: 'Fizz' if divisible by 3, 'Buzz' if by 5, 'FizzBuzz' if both, else the number as a string. Reply with ONLY a python code block, no explanation.	en	[ { "de": "The easiest case in the set, on purpose: it is a health check on the measurement itself. A model failing here usually cannot write code — it means something else broke, such as refusing to answer in a code block, wrapping the code in prose the extractor cannot see through, or the endpoint not respondin...	{ "de": "All 15 entries must match, types included — the numbers are strings ('1', not 1). An answer that prints a tidy result but never returns the list fails.", "en": "All 15 entries must match, types included — the numbers are strings ('1', not 1). An answer that prints a tidy result but never returns the list f...	assert fizzbuzz(15) == ['1','2','Fizz','4','Buzz','Fizz','7','8','Fizz','Buzz','11','Fizz','13','14','FizzBuzz']	null	null	null	null	null	null	null	null	null
code-dedupe-preserve-order	code_exact	{ "de": "Dedupe, preserving order", "en": "Dedupe, preserving order", "fr": "Dedupe, preserving order", "ja": "Dedupe, preserving order", "ko": "Dedupe, preserving order", "th": "ลบค่าซ้ำโดยคงลำดับเดิม", "vi": "Dedupe, preserving order", "zh": "Dedupe, preserving order" }	Write a Python function `dedupe(items)` that removes duplicates from a list while preserving the first-seen order. Reply with ONLY a python code block, no explanation.	en	[ { "de": "It catches the popular shortcut list(set(items)), which returns the right number of elements in the wrong order. Reviewers wave this one through because it looks like it works, right up until someone notices the output has been reshuffled.", "en": "It catches the popular shortcut list(set(items)), ...	{ "de": "It must return exactly [3,1,2,4]. The input is chosen so that set ordering ([1,2,3,4]) differs from first-seen ordering, which means a set-based answer fails deterministically rather than by bad luck.", "en": "It must return exactly [3,1,2,4]. The input is chosen so that set ordering ([1,2,3,4]) differs fr...	assert dedupe([3,1,2,1,3,4]) == [3,1,2,4]	null	null	null	null	null	null	null	null	null
code-thai-slug	code_exact	{ "de": "Counting Thai characters", "en": "Counting Thai characters", "fr": "Counting Thai characters", "ja": "Counting Thai characters", "ko": "Counting Thai characters", "th": "นับอักขระไทยในสตริง", "vi": "Counting Thai characters", "zh": "Counting Thai characters" }	Write a Python function `count_thai_chars(s)` that returns the count of Thai Unicode characters (U+0E00 to U+0E7F) in a string. Reply with ONLY a python code block, no explanation.	en	[ { "de": "It tests whether the model understands how Thai works in Unicode. The word สวัสดี reads as four letters to a Thai speaker but is six code points, because vowels and tone marks are counted separately from the consonants they attach to. A model that counts what the eye sees, or that reasons in grapheme c...	{ "de": "Exactly 6, with no spaces or surrounding Latin counted. The correct range is the whole of U+0E00–U+0E7F, not just the consonant block.", "en": "Exactly 6, with no spaces or surrounding Latin counted. The correct range is the whole of U+0E00–U+0E7F, not just the consonant block.", "fr": "Exactly 6, with n...	assert count_thai_chars('hello สวัสดี world') == 6	null	null	null	null	null	null	null	null	null
code-binary-search	code_exact	{ "de": "Binary search", "en": "Binary search", "fr": "Binary search", "ja": "Binary search", "ko": "Binary search", "th": "ค้นหาแบบทวิภาค", "vi": "Binary search", "zh": "Binary search" }	Write a Python function `binary_search(arr, target)` that returns the index of target in a sorted list arr, or -1 if not found. Reply with ONLY a python code block, no explanation.	en	[ { "de": "A textbook algorithm with the textbook off-by-one traps in the loop condition and the bound updates. More importantly it tests whether the function honours its contract in the not-found case, which is exactly where model-written code likes to improvise.", "en": "A textbook algorithm with the textbo...	{ "de": "Both directions must hold: the right index when found, and -1 when not. Returning None, returning False, or raising on a miss fails even though the first half was right.", "en": "Both directions must hold: the right index when found, and -1 when not. Returning None, returning False, or raising on a miss fa...	assert binary_search([1,3,5,7,9,11], 7) == 3 and binary_search([1,3,5], 4) == -1	null	null	null	null	null	null	null	null	null
thai-self-intro	thai_integrity	{ "de": "Self-introduction", "en": "Self-introduction", "fr": "Self-introduction", "ja": "Self-introduction", "ko": "Self-introduction", "th": "แนะนำตัวเอง", "vi": "Self-introduction", "zh": "Self-introduction" }	คุณคือใคร แนะนำตัวเป็นภาษาไทยสั้นๆ	th	[ { "de": "The single most common opening line real users type — and a place where models quietly switch to English despite being asked in Thai. It checks that the model speaks Thai when spoken to in Thai, before anything else is measured.", "en": "The single most common opening line real users type — and a p...	{ "de": "At least 10 Thai characters and no duplicated marks. This case is deliberately exempt from the foreign-script gate: Qwen answers with its own product name, 通义千问, which is correct. Counting that as a defect would nearly double the set's reported garbling rate.", "en": "At least 10 Thai characters and no dup...	null	[ "min_thai_chars: 10", "allow_foreign_script" ]	null	null	null	null	null	null	null	null
thai-explain-binary-search	thai_integrity	{ "de": "Explaining binary search in Thai", "en": "Explaining binary search in Thai", "fr": "Explaining binary search in Thai", "ja": "Explaining binary search in Thai", "ko": "Explaining binary search in Thai", "th": "อธิบาย binary search เป็นไทย", "vi": "Explaining binary search in Thai", "zh": "Expla...	อธิบายวิธีการทำงานของ binary search เป็นภาษาไทย สั้นๆ	th	[ { "de": "Deliberately paired with code-binary-search: that case asks whether the model can write it, this one whether it can explain it in Thai without the characters breaking. They are different abilities, and builds that pass the first often fail this one. English technical terms embedded mid-sentence are whe...	{ "de": "At least 30 Thai characters, no duplicated marks, and not a single CJK character. An all-English answer fails on the character count alone — no judgement call required about whether it addressed the question.", "en": "At least 30 Thai characters, no duplicated marks, and not a single CJK character. An all-...	null	[ "min_thai_chars: 30" ]	true	null	null	null	null	null	null	null
thai-current-model	thai_integrity	{ "de": "Which model is this", "en": "Which model is this", "fr": "Which model is this", "ja": "Which model is this", "ko": "Which model is this", "th": "ตอนนี้ใช้โมเดลอะไร", "vi": "Which model is this", "zh": "Which model is this" }	ตอนนี้ใช้ model อะไร	th	[ { "de": "A short question mixing Thai and English in one line — how Thai technical users actually type. It shows how the model handles a prompt that is both brief and ambiguous; there is so little context here that some models guess wrong and switch language entirely.", "en": "A short question mixing Thai a...	{ "de": "At least 10 Thai characters, unbroken. It does not check whether the model names itself correctly: a model has no way of knowing which alias it is served under on the proxy. Only the language is measurable here, and we only score what is measurable.", "en": "At least 10 Thai characters, unbroken. It does n...	null	[ "min_thai_chars: 10", "allow_foreign_script" ]	null	null	null	null	null	null	null	null
thai-hashtable	thai_integrity	{ "de": "Explaining a hash table in depth", "en": "Explaining a hash table in depth", "fr": "Explaining a hash table in depth", "ja": "Explaining a hash table in depth", "ko": "Explaining a hash table in depth", "th": "อธิบาย hash table แบบละเอียด", "vi": "Explaining a hash table in depth", "zh": "Expla...	อธิบายวิธีการทำงานของ hash table เป็นภาษาไทยอย่างละเอียด ครอบคลุมฟังก์ชันแฮช การชนกันของคีย์ และการขยายตาราง	th	[ { "de": "The longest case in this round, and the one that surfaces garbling most often: the longer the Thai passage, the more chances a duplicated mark or a stray Chinese character has to appear. Passing the short cases does not tell you a model can hold Thai together over a full page.", "en": "The longest ...	{ "de": "At least 600 Thai characters, no duplicated marks, no CJK. The prompt no longer states a length at all — the length comes from asking for three sub-topics, while the pass line is 600 characters, which can be measured without anyone arguing about how Thai words are counted. Models that passed this case in the...	null	[ "min_thai_chars: 600" ]	true	null	null	null	null	null	null	null
judge-code-review	judge	{ "de": "Code review: division by zero", "en": "Code review: division by zero", "fr": "Code review: division by zero", "ja": "Code review: division by zero", "ko": "Code review: division by zero", "th": "รีวิวโค้ด: หารด้วยศูนย์", "vi": "Code review: division by zero", "zh": "Code review: division by zer...	Review this Python function for bugs and suggest a fix, in 3 sentences or less: def get_average(nums): return sum(nums) / len(nums)	en	[ { "de": "Spotting the bug that has not happened yet. This function is perfectly correct for every input except an empty list — the shape of defect most likely to reach production. The three-sentence limit stops a model from passing by listing every conceivable concern.", "en": "Spotting the bug that has not...	{ "de": "A median of 85 or above in 4 of 7 trials, and both halves are required: naming the empty-list ZeroDivisionError and proposing a fix that actually works. An answer that only advises validating input, without saying what breaks, lands in the middle band and does not pass.", "en": "A median of 85 or above in ...	null	null	null	Does the response correctly identify the ZeroDivisionError risk when nums is empty, and propose a reasonable fix? Score 1 (misses it or wrong) to 5 (catches it clearly with a correct fix).	{ "de": "The trailing “Score 1-5” is left over from the earlier scale. The runner moved to 0-100 on 16 Aug 2026 and the real score bands are set in the judge prompt, so that sentence has no effect on the score.", "en": "The trailing “Score 1-5” is left over from the earlier scale. The runner moved to 0-100 on 16 Au...	null	null	null	null	null
judge-thai-explain	judge	{ "de": "SQL vs NoSQL, explained in Thai", "en": "SQL vs NoSQL, explained in Thai", "fr": "SQL vs NoSQL, explained in Thai", "ja": "SQL vs NoSQL, explained in Thai", "ko": "SQL vs NoSQL, explained in Thai", "th": "SQL กับ NoSQL ต่างกันอย่างไร", "vi": "SQL vs NoSQL, explained in Thai", "zh": "SQL vs NoSQ...	อธิบายความแตกต่างระหว่าง SQL กับ NoSQL ให้ผู้เริ่มต้นเข้าใจง่ายๆ เป็นภาษาไทย	th	[ { "de": "The only case that measures two things at once: that the content is right and that the Thai holds together. It is also why the foreign-script gate had to be separated from round 2 — this case's type is judge, not thai_integrity, so tying the gate to a single type would have missed it exactly where it h...	{ "de": "A median of 85 or above in 4 of 7 trials — but one CJK character anywhere turns that trial into a fail, overriding the judge even at a score of 95. The real case we caught was “คีย์ที่ต่างกัน却被ฟังก์ชันแฮชคำนวณ”, which a meaning-focused judge scored highly and any Thai reader would reject on sight.", "en": ...	null	null	true	Is the explanation technically accurate, clear for a beginner, and grammatically clean Thai (no garbled/duplicated characters, no awkward phrasing)? Score 1-5.	{ "de": "The trailing “Score 1-5” is left over from the earlier scale. The runner moved to 0-100 on 16 Aug 2026 and the real score bands are set in the judge prompt, so that sentence has no effect on the score.", "en": "The trailing “Score 1-5” is left over from the earlier scale. The runner moved to 0-100 on 16 Au...	true	null	null	null	null
judge-tool-call-reasoning	judge	{ "de": "Tool-use reasoning: finding TODOs", "en": "Tool-use reasoning: finding TODOs", "fr": "Tool-use reasoning: finding TODOs", "ja": "Tool-use reasoning: finding TODOs", "ko": "Tool-use reasoning: finding TODOs", "th": "วางแผนใช้เครื่องมือ: หา TODO", "vi": "Tool-use reasoning: finding TODOs", "zh": ...	A user asks: 'Find all TODO comments in this project and list the files.' You have access to a shell/grep tool. Describe, step by step, exactly what command(s) you would run. Be concrete and specific.	en	[ { "de": "An agentic ability: it asks not for knowledge but for a plan that runs. “Use grep to search for TODO” reads fine to a human and says nothing about skipping directories you should not search or producing the file list that was actually asked for — the difference between a command that works and one you ...	{ "de": "A median of 85 or above in 4 of 7 trials. The rubric leans on “would it actually work” — the command has to run, not merely name the right tool. Sensible flags are what separates the 85-plus band from the 60–79 one.", "en": "A median of 85 or above in 4 of 7 trials. The rubric leans on “would it actually w...	null	null	null	Is the proposed approach correct, efficient (e.g. uses grep/ripgrep with sensible flags), and would it actually work? Score 1-5.	{ "de": "The trailing “Score 1-5” is left over from the earlier scale. The runner moved to 0-100 on 16 Aug 2026 and the real score bands are set in the judge prompt, so that sentence has no effect on the score.", "en": "The trailing “Score 1-5” is left over from the earlier scale. The runner moved to 0-100 on 16 Au...	null	null	null	null	null
code-tool-parser-format	code_exact	{ "de": "Parsing a query string", "en": "Parsing a query string", "fr": "Parsing a query string", "ja": "Parsing a query string", "ko": "Parsing a query string", "th": "แยก query string เป็น dict", "vi": "Parsing a query string", "zh": "Parsing a query string" }	Write a Python function `parse_query_string(qs)` that parses a URL query string like 'a=1&b=2&c=3' into a dict {'a':'1','b':'2','c':'3'}. Reply with ONLY a python code block, no explanation.	en	[ { "de": "Producing output in exactly the shape that was asked for, not merely output that works. Data in, data out is what models are most often wired to do in automated pipelines, and a result of the wrong shape breaks the whole pipe.", "en": "Producing output in exactly the shape that was asked for, not m...	{ "de": "Values must be strings ('1', not 1) in a single-value dict. Reaching for urllib.parse.parse_qs returns {'a': ['1']} — a list — and fails; parse_qsl wrapped in dict() passes.", "en": "Values must be strings ('1', not 1) in a single-value dict. Reaching for urllib.parse.parse_qs returns {'a': ['1']} — a list...	assert parse_query_string('a=1&b=2&c=3') == {'a':'1','b':'2','c':'3'}	null	null	null	null	null	null	null	null	null
code-buddhist-year	code_exact	{ "de": "Converting a Buddhist-era date", "en": "Converting a Buddhist-era date", "fr": "Converting a Buddhist-era date", "ja": "Converting a Buddhist-era date", "ko": "Converting a Buddhist-era date", "th": "แปลง พ.ศ. เป็นวันที่ ISO", "vi": "Converting a Buddhist-era date", "zh": "Converting a Buddhist...	Write a Python function `be_to_iso(date_str)` that converts a Thai Buddhist-era date in the format 'dd/mm/yyyy' (e.g. '15/08/2569') into an ISO date string 'yyyy-mm-dd' in the Gregorian calendar (e.g. '2026-08-15'). The Buddhist era is 543 years ahead of the Gregorian one. Keep zero padding. Reply with ONLY a python co...	en	[ { "de": "Added in v3 because nearly every real Thai system has to do it, and it has three ways to go wrong at once: the direction of the 543-year offset, flipping dd/mm/yyyy into yyyy-mm-dd, and keeping the zero padding, which vanishes the moment you round-trip through integers.", "en": "Added in v3 because...	{ "de": "Both assertions must hold. The second deliberately uses 01/01 to force the zero padding to survive, and a year that crosses a century (2500 → 1957) to rule out answers with the 2000s hard-coded in.", "en": "Both assertions must hold. The second deliberately uses 01/01 to force the zero padding to survive, ...	assert be_to_iso('15/08/2569') == '2026-08-15' assert be_to_iso('01/01/2500') == '1957-01-01'	null	null	null	null	null	true	null	null	null
code-lru-cache	code_exact	{ "de": "LRU cache", "en": "LRU cache", "fr": "LRU cache", "ja": "LRU cache", "ko": "LRU cache", "th": "แคช LRU", "vi": "LRU cache", "zh": "LRU cache" }	Write a Python class `LRUCache` with `__init__(self, capacity)`, `get(self, key)` returning -1 when the key is absent, and `put(self, key, value)`. When the cache is over capacity it must evict the least-recently-used key. Both get and put count as a use. Reply with ONLY a python code block, no explanation.	en	[ { "de": "The longest prompt in this round and the only one requiring a stateful class rather than a pure function. The load-bearing sentence is “both get and put count as a use” — the easiest detail to skim past in the entire set.", "en": "The longest prompt in this round and the only one requiring a statef...	{ "de": "The order of the assertions is the test. get(1) is called before put(3), so an implementation that counts only put as a use evicts the wrong key and fails at get(2) == -1. Counting get as a use — for instance with OrderedDict.move_to_end — passes.", "en": "The order of the assertions is the test. get(1) is...	c = LRUCache(2) c.put(1, 1) c.put(2, 2) assert c.get(1) == 1 c.put(3, 3) assert c.get(2) == -1 assert c.get(3) == 3 assert c.get(1) == 1	null	null	null	null	null	true	null	null	null
thai-markdown-table	thai_integrity	{ "de": "A Markdown table in Thai", "en": "A Markdown table in Thai", "fr": "A Markdown table in Thai", "ja": "A Markdown table in Thai", "ko": "A Markdown table in Thai", "th": "ตาราง markdown ภาษาไทย", "vi": "A Markdown table in Thai", "zh": "A Markdown table in Thai" }	เปรียบเทียบ RAM กับ SSD เป็นตาราง markdown ภาษาไทย 3 แถว ได้แก่ ความเร็ว ความจุ และการเก็บข้อมูลเมื่อปิดเครื่อง ตอบเป็นตารางอย่างเดียว ไม่ต้องมีคำอธิบายเพิ่ม	th	[ { "de": "Added in v3 because Thai inside narrow table cells behaves differently from Thai in a paragraph: the text alternates between pipe characters and Thai words all the way across every row, which is where marks come loose or collide with the delimiters. It is also one of the output formats real users ask f...	{ "de": "At least 60 Thai characters, no duplicated marks, no CJK. This round does not check that the Markdown is well-formed or that the comparison is accurate — format compliance is judged in judge-format-constraints instead.", "en": "At least 60 Thai characters, no duplicated marks, no CJK. This round does not c...	null	[ "min_thai_chars: 60" ]	true	null	null	null	true	null	null	null
thai-concurrency-terms	thai_integrity	{ "de": "Race condition vs deadlock", "en": "Race condition vs deadlock", "fr": "Race condition vs deadlock", "ja": "Race condition vs deadlock", "ko": "Race condition vs deadlock", "th": "race condition กับ deadlock", "vi": "Race condition vs deadlock", "zh": "Race condition vs deadlock" }	อธิบายว่า race condition กับ deadlock ต่างกันอย่างไร เป็นภาษาไทย ไม่เกิน 5 บรรทัด	th	[ { "de": "Added in v3 because neither term has a settled Thai translation, forcing the model to switch between Thai and English constantly inside a short passage — the exact condition under which our captured examples show Chinese characters slipping in. It is the harshest foreign-script test in the round.", ...	{ "de": "At least 80 Thai characters while the prompt caps the answer at five lines — deliberately forcing dense Thai into a small space rather than letting the model escape into English or say nothing at length.", "en": "At least 80 Thai characters while the prompt caps the answer at five lines — deliberately forc...	null	[ "min_thai_chars: 80" ]	true	null	null	null	true	null	null	null
judge-rag-grounding	judge	{ "de": "Answer from the document only", "en": "Answer from the document only", "fr": "Answer from the document only", "ja": "Answer from the document only", "ko": "Answer from the document only", "th": "ตอบจากเอกสารเท่านั้น ห้ามแต่ง", "vi": "Answer from the document only", "zh": "Answer from the docume...	ตอบคำถามโดยใช้ข้อมูลจากเอกสารด้านล่างเท่านั้น ห้ามใช้ความรู้อื่นนอกเอกสาร เอกสาร: บริษัท สปาร์คไทย จำกัด ก่อตั้งเมื่อ พ.ศ. 2562 สำนักงานใหญ่อยู่ที่จังหวัดเชียงใหม่ ให้บริการติดตั้งระบบ AI ภายในองค์กร และมีศูนย์ข้อมูลของตัวเองหนึ่งแห่ง คำถาม: บริษัทนี้มีพนักงานทั้งหมดกี่คน	th	[ { "de": "Added in v3 because the old 12-case set had no test of this at all. The question is built so the document cannot answer it; the one correct reply is that the headcount is not stated. This is the single most important property for a model wired into an internal document search, because such systems fail...	{ "de": "A median of 85 or above in 4 of 7 trials. The rubric caps the score at 29 the moment any figure appears — exact, ranged, or estimated — however politely or tentatively it is phrased. That cap is spelled out because an LLM judge tends to over-reward fluent prose, and this case has to be decided on facts befor...	null	null	true	The document does NOT state the number of employees. A correct response says the document does not contain that information and supplies no figure. Score 90-100 only if it clearly states the information is absent AND gives no invented number. Any specific headcount, range, or estimate is a factual fabrication: cap the ...	null	true	true	null	null	null
judge-format-constraints	judge	{ "de": "Following an exact output format", "en": "Following an exact output format", "fr": "Following an exact output format", "ja": "Following an exact output format", "ko": "Following an exact output format", "th": "ทำตามรูปแบบผลลัพธ์เป๊ะ ๆ", "vi": "Following an exact output format", "zh": "Following...	สรุปข้อดีของการรันโมเดลภาษาไว้บนเครื่องตัวเองแทนการเรียก API ภายนอก เป็นภาษาไทย ตอบเป็น bullet ให้ครบ 3 ข้อพอดี แต่ละข้อยาวไม่เกิน 80 ตัวอักษร ห้ามมีข้อความเกริ่นนำและห้ามมีข้อความสรุปปิดท้าย	th	[ { "de": "Added in v3 alongside the RAG case. The prompt stacks four constraints at once — the number of bullets, the length of each, no preamble, no closing line — and that last one is something assistant models are trained to produce automatically. So the case measures whether the user's instruction beats the ...	{ "de": "A median of 85 or above in 4 of 7 trials. The rubric grades format before content and caps the score at 59 if any single constraint breaks — one stray closing line such as “hope this helps” fails the trial even when all three points are perfectly correct. The v4 set changed the length unit from “15 words” to...	null	null	true	Grade format compliance first, content second. The response must contain exactly 3 bullet points, each no longer than 80 characters, with no preamble sentence before the list and no closing sentence after it. Score 90-100 only when every one of those constraints holds and the three points are accurate. If the count is ...	null	true	true	null	null	null
g-thai-longform-purity	graded	{ "de": "Long-form Thai without script leakage", "en": "Long-form Thai without script leakage", "fr": "Long-form Thai without script leakage", "ja": "Long-form Thai without script leakage", "ko": "Long-form Thai without script leakage", "th": "เขียนไทยยาวโดยไม่มีอักษรต่างภาษาหลุด", "vi": "Long-form Thai w...	อธิบายวิธีการทำงานของระบบคิวข้อความ (message queue)เป็นภาษาไทยอย่างละเอียดสำหรับวิศวกรที่เพิ่งเริ่มงาน ครอบคลุมผู้ผลิตและผู้บริโภคข้อความ การรับประกันลำดับ การส่งซ้ำเมื่อผิดพลาด และการคั่งของคิว เขียนเป็นความเรียงต่อเนื่องความยาวอย่างน้อย 1,500 ตัวอักษร ไม่ต้องใส่โค้ด	th	[ { "de": "This case was not invented; it came from a defect we had already measured. On 17 Aug, Qwen3.8-27B-NVFP4 wrote 4,048–4,958 Thai characters and leaked Chinese in two runs out of three (历遍, 堆扎); Qwen3.6 leaked in one of three (现种象这); the FP8 build had leaked 却表被链. Three builds, three rates — a dimension t...	{ "de": "Three bands: a single Chinese, Japanese, Korean, Cyrillic, or Devanagari character scores 0 outright; clean but shorter than 1,500 Thai characters scores in proportion to length; clean and long enough scores 100. The case score is the mean across the three variants.", "en": "Three bands: a single Chinese, ...	null	[ "min_thai_chars=1500", "max_tokens 6000 → 8000", "pass_score=80", "3 สำนวน" ]	true	null	null	true	null	true	[ "อธิบายวิธีการทำงานของดัชนีแบบ B-tree ในฐานข้อมูลเป็นภาษาไทยอย่างละเอียดสำหรับวิศวกรที่เพิ่งเริ่มงาน ครอบคลุมโครงสร้างของโหนด การค้นหาแบบไล่ระดับ การแทรกที่ทำให้โหนดแตก และผลต่อการเขียน เขียนเป็นความเรียงต่อเนื่องความยาวอย่างน้อย 1,500 ตัวอักษร ไม่ต้องใส่โค้ด", "อธิบายวิธีการทำงานของสายงานส่งมอบซอฟต์แวร์อัตโนมัติ...	80
g-longctx-needle	graded	{ "de": "Three facts in a 9,793-character Thai policy", "en": "Three facts in a 9,793-character Thai policy", "fr": "Three facts in a 9,793-character Thai policy", "ja": "Three facts in a 9,793-character Thai policy", "ko": "Three facts in a 9,793-character Thai policy", "th": "หาข้อเท็จจริงสามจุดในเอกสารไท...	{doc} จากเอกสารข้างต้นเท่านั้น ตอบคำถามสามข้อนี้ให้ครบทุกข้อ ตอบสั้น ๆ ข้อละหนึ่งบรรทัด 1) รหัสอ้างอิงกลางของโครงการนำร่องด้านปัญญาประดิษฐ์คืออะไร 2) บันทึกการเข้าถึงข้อมูลส่วนบุคคลต้องเก็บไว้เป็นระยะเวลากี่วัน 3) ผู้ใดปฏิบัติหน้าที่เจ้าหน้าที่คุ้มครองข้อมูลส่วนบุคคลขององค์กร	th	[ { "de": "None of the 18 existing cases touched long context at all, though every model's spec advertises it (Qwen3.8 claims a 262,144-token window). The document is generated deterministically and committed as a file so every run gets literally the same exam.", "en": "None of the 18 existing cases touched l...	{ "de": "The score is the share of needles answered correctly (0, 33, 67, or 100). An answer that names both the right value and its decoy counts as a miss — the reader still cannot tell which one is the answer.", "en": "The score is the share of needles answered correctly (0, 33, 67, or 100). An answer that names ...	null	[ "เอกสาร 9,793 อักขระ", "เข็ม 3 จุด + กับดัก 3 จุด", "max_tokens 700 → 4096", "pass_score=67" ]	null	null	null	null	null	true	null	67
g-rag-refuse-ungrounded	graded	{ "de": "Admit the document does not say", "en": "Admit the document does not say", "fr": "Admit the document does not say", "ja": "Admit the document does not say", "ko": "Admit the document does not say", "th": "ยอมรับว่าเอกสารไม่ได้บอกไว้", "vi": "Admit the document does not say", "zh": "Admit the do...	{doc} จากเอกสารข้างต้นเท่านั้น: งบประมาณรายปีของโครงการ SPK-4417-TH เป็นจำนวนเงินเท่าใด ตอบตามที่เอกสารระบุไว้	th	[ { "de": "The old judge-rag-grounding case handed every model a flat 100 across nine runs (standard deviation 0.0) because the answer was in the document — anyone can do that. This one asks for something absent, which is the side that actually fails and the real risk in enterprise RAG.", "en": "The old judge...	{ "de": "Inventing a figure scores 0 outright; saying the document does not state it scores 100; an answer that invents nothing but never says the information is missing scores 30.", "en": "Inventing a figure scores 0 outright; saying the document does not state it scores 100; an answer that invents nothing but nev...	null	[ "เอกสารเดียวกับเคสก่อนหน้า", "max_tokens 600 → 4096", "pass_score=100" ]	null	null	null	true	null	true	null	100
g-json-schema-strict	graded	{ "de": "Return JSON that matches the spec exactly", "en": "Return JSON that matches the spec exactly", "fr": "Return JSON that matches the spec exactly", "ja": "Return JSON that matches the spec exactly", "ko": "Return JSON that matches the spec exactly", "th": "คืน JSON ให้ตรงสเปกทุกข้อ", "vi": "Return ...	ประกาศรับสมัครงาน: บริษัท ทรัพย์เจริญโลจิสติกส์ จำกัด เปิดรับ วิศวกรข้อมูล (Data Engineer) ประจำสำนักงานใหญ่ จังหวัดชลบุรี เงินเดือน 55,000 - 85,000 บาท ตามประสบการณ์ ต้องมีประสบการณ์ Python, SQL, Airflow และ dbt ทำงานที่ออฟฟิศเป็นหลัก ไม่รองรับการทำงานทางไกล สกัดข้อมูลจากประกาศข้างต้นเป็น JSON ตามสเปกนี้เท่านั้น: {"t...	th	[ { "de": "This is work SparkThai's own systems do every day — turning Thai job posts into structured records — and the place where a model fails most quietly, because JSON that is slightly off-spec takes the whole pipeline down.", "en": "This is work SparkThai's own systems do every day — turning Thai job po...	{ "de": "The score is the share of requirements met, covering everything from no markdown fence, parses as JSON, and no prose wrapper, through every key being present with the right type (salaries as integers, not strings; remote as a boolean, not text), skills capped at five, to no extra keys.", "en": "The score i...	null	[ "17 ข้อกำหนดต่อคำตอบ", "max_tokens 800 → 4096", "pass_score=90", "2 สำนวน" ]	null	null	null	null	null	true	[ "ประกาศรับสมัครงาน: สตาร์ทอัพด้านสุขภาพ HealthLink ต้องการ นักพัฒนาเว็บฝั่งหน้าบ้าน (Frontend Developer) ทำงานทางไกลได้เต็มรูปแบบ ทีมตั้งอยู่ที่กรุงเทพมหานคร ค่าตอบแทน 40,000 ถึง 70,000 บาทต่อเดือน ทักษะที่ต้องใช้ ได้แก่ React, TypeScript และ Tailwind\n\nสกัดข้อมูลจากประกาศข้างต้นเป็น JSON ตามสเปกนี้เท่านั้น:\n{\"t...	90
g-constraint-stack	graded	{ "de": "Five countable constraints at once", "en": "Five countable constraints at once", "fr": "Five countable constraints at once", "ja": "Five countable constraints at once", "ko": "Five countable constraints at once", "th": "ทำตามข้อกำหนดห้าข้อที่นับได้พร้อมกัน", "vi": "Five countable constraints at o...	สรุปสาระสำคัญของระเบียบการกำกับดูแลข้อมูลในองค์กร เป็นภาษาไทย ตามข้อกำหนดต่อไปนี้ทุกข้อ: - ตอบเป็นบูลเล็ตขึ้นต้นด้วยเครื่องหมาย - จำนวน 5 ข้อพอดี ไม่มีข้อความอื่นนอกบูลเล็ต - แต่ละบูลเล็ตยาวไม่เกิน 60 ตัวอักษร - ห้ามใช้คำว่า AI ให้ใช้คำว่า ปัญญาประดิษฐ์ แทน และห้ามใช้วงเล็บหรือจุลภาค - ต้องมีบูลเล็ตอย่างน้อยหนึ่งข้อที่...	th	[ { "de": "This case comes straight out of a v3 lesson: a prompt once said “each bullet no longer than 15 words.” Thai has no word delimiters, so the judge could not actually count, and the model space-separated every Thai word to make counting possible — scoring 95 for prose no Thai speaker would write. A requir...	{ "de": "The score counts how many of the five hold: exactly five bullets, none over 60 characters, no occurrence of “AI” and no parentheses or commas, the string “พ.ศ. 2569” present, and no Chinese, Japanese, or Korean characters.", "en": "The score counts how many of the five hold: exactly five bullets, none over...	null	[ "5 ข้อกำหนด", "บูลเล็ต 5 ข้อ ≤ 60 อักขระ", "max_tokens 900 → 4096", "pass_score=80", "2 สำนวน" ]	true	null	null	null	null	true	[ "สรุปสาระสำคัญของระเบียบการกำกับดูแลข้อมูลในองค์กรของหน่วยงานรัฐ เป็นภาษาไทย ตามข้อกำหนดต่อไปนี้ทุกข้อ:\n- ตอบเป็นบูลเล็ตขึ้นต้นด้วยเครื่องหมาย - จำนวน 5 ข้อพอดี ไม่มีข้อความอื่นนอกบูลเล็ต\n- แต่ละบูลเล็ตยาวไม่เกิน 60 ตัวอักษร\n- ห้ามใช้คำว่า AI ให้ใช้คำว่า ปัญญาประดิษฐ์ แทน และห้ามใช้วงเล็บหรือจุลภาค\n- ต้องมีบูลเ...	80

SparkThai Model Arena

Question bank, versioned snapshots, run history, and full per-trial results from the SparkThai Model Arena — SparkThai's own language-model eval suite, run on our own NVIDIA DGX Spark hardware.

🌐 sparkthai.com · 🏆 Model Arena · 🐙 github.com/sparkthai

What this is

The Model Arena grades Qwen model variants we host ourselves on DGX Spark across four kinds of checks:

code_exact — code tasks graded by running the model's code against an assert test, no LLM judge involved
thai_integrity — checks that Thai text is actually correct (tokenization, character counts, encoding), not just "looks Thai"
judge — open-ended answers scored by an LLM judge against a rubric
graded — a fixed rubric with partial credit

23 cases total across those four rounds. Every case ships with a bilingual (Thai/English) explanation of why it's in the set and what counts as a pass — most public benchmarks give you a score, not the reasoning behind the grading.

Why this is worth publishing as-is

We ran a judge-variance experiment before trusting the judge round: the same response, scored 15 times by the same judge rubric, came back with scores ranging 65–95 (σ ≈ 10.4). That's why grading uses a pass_n × 2 > n majority rule instead of a single judge call, and why the graded round only contributes a small, fixed slice of the total score. See case_sets.json and the purpose/passWhen fields in cases.json for the full reasoning per case.

Files

File	What's in it
`cases.json`	The 23 cases: prompt, round type, bilingual `purpose` (why this case exists) and `passWhen` (grading rule), plus the `assert`-style test for `code_exact` cases
`case_sets.json`	Versioned snapshots of the case bank tied to a git commit hash, with notes on what changed between versions and how per-case token caps were reconstructed where history was incomplete
`history.json`	21 timestamped arena runs (2026-08-16 → 2026-08-19): model, pass/total, Wilson confidence interval, avg tokens/sec, and which rubric version graded the run
`results/*.json`	Full per-trial results for 7 model variants we've run through the arena (`qwen3.6-fp8`, `qwen3.6-nvfp4`, `qwen3.7-flash`, `qwen3.7-max`, `qwen3.7-plus`, `qwen3.8-fp8`, `qwen3.8-nvfp4`) — every trial's raw model response, pass/fail, tokens/sec, and truncation status, not just the aggregate score

cases and history are exposed as dataset viewer previews above. case_sets.json and results/*.json are nested (dict-of-objects) reference files, not flat tables, so they don't appear in the preview — download them directly instead.

What's not in this dataset

The Model Arena also scores our models against two third-party benchmarks — typhoon-ai/thai_exam for Thai and LaoBench for Lao. Those question sets belong to their original authors, so only our own hand-authored 23-case suite is republished here. Our scores on those external benchmarks are visible on the Model Arena page itself.

License

The prompts, rubrics, and methodology (cases.json, case_sets.json) are SparkThai's own work, released under CC-BY-4.0. Model responses in history.json and results/*.json are outputs from Qwen-family models we serve ourselves.

Citation

@misc{sparkthai-model-arena,
  title  = {SparkThai Model Arena},
  author = {SparkThai},
  year   = {2026},
  url    = {https://huggingface.co/datasets/sparkthai/sparkthai-model-arena}
}

Downloads last month: 49

Number of rows:

Total file size:

2.72 MB