Datasets:

LangGao
/

pythia-160m-neuron-evolution

neuron int32 16 3.06k	step int64 0 143k	concept stringlengths 0 86	description stringlengths 134 479	relation stringclasses 4 values	clarity int32 1 5 ⌀	patterns listlengths 1 19	unexplained listlengths 0 10	detection_score float32 0.13 1 ⌀	map_xy listlengths 2 2	zh dict
50	0	NH strings and a few vibrational spectroscopy terms	Strongly responds primarily to the contiguous characters NH in either case, sometimes followed by PA, api, 4, As, tilde digits, and so on, across legal, code, and scientific texts; also responds to terms such as bending, mode, and liberation in molecular vibration and spectroscopy contexts.	unrelated	2	[ { "name": "NH string and suffixes", "trigger": "«NH»10", "context": "The string NH itself triggers activation; following characters vary by domain and may include digits, uppercase or lowercase letters, or word endings.", "domain": "Cross-domain text, including legal documents, code, and scientific ...	[]	1	[ 0.30730000138282776, 0.6241999864578247 ]	{ "concept": "NH 字符串及少量振动光谱词", "description": "主要对大小写形式的连续字符 NH 强烈响应，后续可接 PA、api、4、As、波浪线数字等，横跨法律、代码和科学文本；另在分子振动与光谱语境中响应 bending、mode、liberation 等词。", "patterns": [ { "name": "NH 字符串及其后缀", "trigger": "«NH»10", "context": "字母串 NH 本身触发；后续字符因领域而异，可接数字、大小写字母或词尾。", "domain": "跨领域文本，包括法律、代码、...
50	2	Uppercase letter string NH	All spans activate on the uppercase letter sequence NH, followed by characters that form abbreviations, chemical formulas, or terms; activation appears to be triggered mainly by the literal substring rather than domain or semantics.	single	5	[ { "name": "Uppercase NH string", "trigger": "«NH»10; including NHPA, NH~4~Cl, NHAs, NHIMI, NH~2~-PEG, NH~2~-terminal etc.", "context": "Activation falls on the uppercase letter sequence NH; following characters and meanings vary, with no required grammatical position", "domain": "Cross-domain: law, ...	[]	null	[ 0.2037999927997589, 0.8140000104904175 ]	{ "concept": "大写字母串 NH", "description": "所有片段均在大写字母序列 NH 上激活，后接字符可构成缩写、化学式或术语；激活似乎主要由字面子串触发，而非领域或语义。", "patterns": [ { "name": "大写 NH 字符串", "trigger": "«NH»10；包括 NHPA、NH~4~Cl、NHAs、NHIMI、NH~2~-PEG、NH~2~-terminal 等", "context": "激活落在大写字母串 NH 上；后续字符和词义各异，不要求特定语法位置", "domain": "跨领域：法律、代码、科...
50	8	Uppercase letter string NH	All spans activate on the uppercase letter string “NH,” covering chemical formulas, abbreviations, code identifiers, and compound terms; following content and style vary, with no more specific shared semantic pattern evident.	single	5	[ { "name": "NH-related string spans", "trigger": "«NH»10, including NH$_3$, NHPA, NHapi, NH$_4$Cl, NHAs, NHIMI, NH$_2$-PEG-PCL, NH$_2$-terminal", "context": "Activation is localized on the uppercase letter string NH; following characters and semantic uses vary, with no unified lexical condition.", "d...	[]	null	[ 0.2858000099658966, 0.7432000041007996 ]	{ "concept": "大写字母串 NH", "description": "所有段均在大写字母串“NH”处激活，涵盖化学式、缩写、代码标识符及复合术语；其后续内容与文体多样，未见更具体的共同语义模式。", "patterns": [ { "name": "NH相关字符串片段", "trigger": "«NH»10，包括 NH$_3$、NHPA、NHapi、NH$_4$Cl、NHAs、NHIMI、NH$_2$-PEG-PCL、NH$_2$-terminal", "context": "激活定位在大写字母串 NH 上；其后字符和语义用途各异，没有统一词义条件。", ...
50	16	Uppercase letter sequence NH	The most stable trigger is the contiguous uppercase letters NH, often at the beginning of an abbreviation, code identifier, place abbreviation, or chemical formula; subsequent characters vary widely with context. NH3 in the first segment follows the same letter-sequence cue, but occurs in a specific context.	related	4	[ { "name": "NH uppercase letter sequence", "trigger": "«NH»10", "context": "Uppercase N and H occur consecutively, possibly followed by digits, abbreviations, or a chemical formula; they may also begin a word or name", "domain": "Cross-domain: legal, code, biomedical, and chemical texts", "segmen...	[ 1 ]	0.9375	[ 0.21369999647140503, 0.7516999840736389 ]	{ "concept": "大写字母序列 NH", "description": "最稳定的触发是大写连续字母 NH，常位于缩写、代码标识符、地名缩写或化学式开头；后续字符与语境变化很大。第1段的 NH3 属同一字母序列线索，但具体语境较特殊。", "patterns": [ { "name": "NH 大小写字母序列", "trigger": "«NH»10", "context": "大写 N、H 连续出现，后续可接数字、缩写或化学式；也可能是词或名称的开头", "domain": "跨领域：法律、代码、生物医学与化学文本" }, { ...
50	32	Character substring NH detection	All spans activate on the same case-sensitive substring “NH”; it appears in chemical formulas, legal abbreviations, code identifiers, place names, and biomedical names, with no stable thematic or semantic condition evident.	single	5	[ { "name": "NH string", "trigger": "«NH»10", "context": "The uppercase sequence NH; it may be followed by digits, letters, or chemical formula notation, which typically does not activate.", "domain": "Cross-domain: legal abbreviations, code, chemistry, biomedicine, place names", "segments": [ ...	[]	null	[ 0.25609999895095825, 0.6504999995231628 ]	{ "concept": "字符子串 NH 检测", "description": "所有片段都激活同一大小写子串“NH”；它出现在化学式、法律缩写、代码标识符、地名及生物医学名称中，未见稳定的主题或语义条件。", "patterns": [ { "name": "NH 字符串", "trigger": "«NH»10", "context": "大小写连写的 NH 前缀；后续可接数字、字母或化学式标记，标记本身通常未激活。", "domain": "跨领域：法律缩写、代码、化学、生物医学、地名" }, { "name": "其他缩写中的...
50	64	Uppercase string NH	The marked positions in all segments fall on the uppercase string NH, which may be followed by digits, chemical formula notation, letters, or punctuation; its meaning varies by context, spanning molecular formulas, proper-name abbreviations, and place abbreviations.	single	5	[ { "name": "NH string", "trigger": "«NH»10", "context": "The local substring begins with uppercase NH, followed by uppercase or lowercase letters, digits, punctuation, spaces, or chemical formula components; it appears in different contexts", "domain": "Cross-domain; papers, legal documents, code, an...	[]	1	[ 0.2583000063896179, 0.7027999758720398 ]	{ "concept": "大写字母串 NH", "description": "所有片段的标记位置都落在大写字母串 NH 上，后续可接数字、化学式标记、字母或标点；其语义随上下文而变，涵盖分子式、专名缩写及地名缩写。", "patterns": [ { "name": "NH 字符串", "trigger": "«NH»10", "context": "局部子串以大写 NH 开始，后接大小写、数字、标点、空格或化学式成分；不同上下文中都出现", "domain": "跨领域；论文、法律、代码及地名等" }, { "name": "化学式...
50	128	Uppercase string NH	All segments activate on NH, including NH₃, NH₄Cl, NH₂, NHPA, NHAPI, NHAs, NHIMI, NH₂-terminal, and the place abbreviation NH (New Hampshire). The shared trigger is the uppercase string NH, not a specific topic or syntax.	single	5	[ { "name": "NH prefix", "trigger": "«NH»10", "context": "The uppercase string NH occurs at the beginning of a word, inside an abbreviation, or in a chemical formula; subsequent characters may continue a name, abbreviation, or formula.", "domain": "Cross-domain; scientific papers, legal documents, cod...	[]	1	[ 0.16019999980926514, 0.7512999773025513 ]	{ "concept": "大写字母串 NH", "description": "所有片段都在 NH 上激活，涵盖 NH₃、NH₄Cl、NH₂、NHPA、NHAPI、NHAs、NHIMI、NH₂-terminal，以及地名 NH（New Hampshire）。共同触发点是大写字母串 NH，而非特定主题或句法。", "patterns": [ { "name": "NH前缀", "trigger": "«NH»10", "context": "大写字母串 NH 在词首、缩写内部或化学式中；后续字符可能继续构成名称、缩写或公式。", "domain": "跨领域；科学论...
50	256	Uppercase string NH	All segments activate at NH; the string occurs in a variety of contexts, including chemical formulas, abbreviations, code identifiers, and place names, with no consistent following characters or meaning.	single	5	[ { "name": "NH string at the start of a fragment", "trigger": "« NH»10", "context": "Activation centers on the uppercase string NH, which may be followed by digits, letters, a chemical formula, or other affixes; it often occurs within a word or proper name/abbreviation and need not stand alone.", "do...	[]	1	[ 0.22869999706745148, 0.7160999774932861 ]	{ "concept": "大写字母串 NH", "description": "所有段落都在 NH 处激活；该字母串出现在化学式、缩写、代码标识及地名等多种上下文，后续字符和语义不固定。", "patterns": [ { "name": "NH 字符串起始片段", "trigger": "« NH»10", "context": "激活集中在大写字母串 NH，后面可接数字、字母、化学式或其他词缀；常位于词内或专名/缩写中，不要求独立成词。", "domain": "跨领域；论文、法律、代码及生物医学文本" }, { "name": "...
50	512	Uppercase NH string	All segments show their strongest activation at the NH string; it may be followed by a subscript, letters, or a hyphenated word fragment, across chemical formulas, institutional abbreviations, code identifiers, and biomedical terms.	single	5	[ { "name": "NH fragment and nitrogen-containing chemical formulas", "trigger": "«NH»10, and «NH»10$_3$, «NH»10PA, «NH»10~4~«Cl»1, «NH»10~2~-PEG, «NH»10~2~-«terminal»4", "context": "The uppercase string NH occurs in a chemical formula, abbreviation, or hyphenated term; it is often adjacent to a subscript,...	[]	1	[ 0.22310000658035278, 0.6887000203132629 ]	{ "concept": "大写 NH 字符串", "description": "所有段都在 NH 字符串处出现最强激活；其后可能接下标、字母或连字符词段，涵盖化学式、机构缩写、代码标识符及生物医学术语。", "patterns": [ { "name": "NH片段及含氮化学式", "trigger": "«NH»10，及 «NH»10$_3$、«NH»10PA、«NH»10~4~«Cl»1、«NH»10~2~-PEG、«NH»10~2~-«terminal»4", "context": "大写字母串 NH 出现在化学式、缩写或连字符化术语中；常与下标、后缀或连字符相邻。"...
50	1,000	Twitter name recognition	All segments activate when Twitter, its accounts, or links are mentioned; contexts concern tweets, retweets, embeds, follows, accounts, and dissemination effects, with the Twitter name as the core trigger.	single	5	[ { "name": "Twitter term and links", "trigger": "«twitter»10", "context": "The name Twitter/twitter or a twitter.com link appears; whether in a sentence, as a standalone mention, or in a URL, the related token is the strongest trigger.", "domain": "Social media discussions, blogs, and web text", ...	[]	1	[ 0.24310000240802765, 0.6262999773025513 ]	{ "concept": "Twitter 名称识别", "description": "全部段落都在提及 Twitter、其账号或链接时激活；上下文涉及推文、转发、嵌入、关注、账号及传播影响，核心触发均为 Twitter 名称。", "patterns": [ { "name": "Twitter 词及链接", "trigger": "«twitter»10", "context": "出现 Twitter/twitter 名称或 twitter.com 链接；不论其处于句中、独立提及还是网址中，相关 token 都是最强触发。", "domain": "社交媒体...
50	2,000	Twitter platform mentions	All segments strongly activate on twitter/Twitter; contexts include the platform name, linked accounts, following, retweeting, information dissemination, presentation, and impact. Many sporadic word fragments also activate, with no stable commonality apparent.	related	4	[ { "name": "Twitter platform name", "trigger": "«twitter»10", "context": "Appears in sentences discussing the social media platform, accounts, posts, or platform features; capitalization varies between Twitter and twitter.", "domain": "Hacker News discussions and web articles", "segments": [ ...	[ 1, 2, 3, 4, 5, 6, 7, 8, 9, 10 ]	1	[ 0.3328000009059906, 0.666700005531311 ]	{ "concept": "Twitter 平台提及", "description": "所有片段都强烈激活于 twitter/Twitter；上下文涵盖平台名称、链接账号、关注、转推、信息传播、呈现与影响。另有大量分散词片段激活，未显示稳定共性。", "patterns": [ { "name": "Twitter 平台名称", "trigger": "«twitter»10", "context": "出现在讨论社交媒体平台、账号、动态或平台功能的句子中；大小写有 Twitter/twitter。", "domain": "Hacker News 讨论及网页文章...
50	3,000	Text segmentation spans and mixed formats	The most salient repeated signal is the subword fragment «kp»10, scattered across heterogeneous text such as formulas, code, email addresses, proper names, and links; other activations mostly fall on tokenized proper names, citations, noise, or within-word Swedish fragments, with no stable semantic theme evident.	unclear	2	[ { "name": "Special subword fragment and byte alignment", "trigger": "«kp»10; also «dim»1, «mn»3, «m»1, «ZZ»1, «KK»2, «hspace»5, etc.", "context": "In mathematical or code-like text, the marker falls on tokenized subword positions in variables, commands, formula symbols, and similar items; in some exampl...	[]	null	[ 0.22689999639987946, 0.6744999885559082 ]	{ "concept": "文本切分片段与混合格式", "description": "最醒目的重复信号是子词片段 «kp»10，散见公式、代码、邮件地址、专名及链接等异质文本；其余激活多落在被切分的专名、引用、噪声或瑞典语词内片段，未见稳定语义主题。", "patterns": [ { "name": "特殊子词片段与字节对齐", "trigger": "«kp»10；另有 «dim»1、«mn»3、«m»1、«ZZ»1、«KK»2、«hspace»5 等", "context": "数学或代码式文本中，标记落在变量、命令、公式符号等被切分的子词位置；部分例子中 «kp»10...
50	4,000	dong in romanized text	Responds primarily and strongly to dong fragments in Pinyin or romanized text, commonly in place names and personal names such as Guang-dong, Shang-dong, and Liu Qiang-dong; also responds to ZnO in materials papers.	unrelated	4	[ { "name": "Pinyin fragments ending in “东”", "trigger": "«dong»10; also «Guang»3«dong»10, «Shang»3«dong»10, «Q»2«iang»5«dong»10", "context": "Appears in Chinese personal names, place names, or romanized place names, especially Guang-dong, Shang-dong, and Liu Qiang-dong; it is not restricted to a single d...	[]	0.5625	[ 0.2567000091075897, 0.6122999787330627 ]	{ "concept": "罗马化文本中的 dong", "description": "主要对拼音或罗马化文本中的 dong 片段强烈响应，常见于 Guang-dong、Shang-dong 和 Liu Qiang-dong 等地名、人名；另在材料论文中响应 ZnO。", "patterns": [ { "name": "“东”字结尾的拼音片段", "trigger": "«dong»10；另有«Guang»3«dong»10、«Shang»3«dong»10、«Q»2«iang»5«dong»10", "context": "出现在中文姓名、地名或地名拼音中，尤其是 Gua...
50	5,000	Beijing place name	All spans activate strongly on the full place name “Beijing”; contexts are mostly Chinese institutional addresses or names of research and medical institutions, with additional legal institution and travel-destination uses.	single	5	[ { "name": "Beijing place name", "trigger": "«Beijing»10", "context": "The city name Beijing appears, usually in institutional, location, or publication information; [8] occurs in a travel context.", "domain": "Academic papers, legal texts, forum comments", "segments": [ 1, 2, 3...	[]	null	[ 0.2529999911785126, 0.6098999977111816 ]	{ "concept": "北京地名", "description": "所有片段均强烈激活于完整地名“Beijing”；上下文多为中国机构地址或科研、医疗机构名称，另有法律机构及旅行目的地用法。", "patterns": [ { "name": "北京地名", "trigger": "«Beijing»10", "context": "出现城市名 Beijing，通常位于机构、地点或出版信息中；[8] 则出现在旅行语境。", "domain": "学术论文、法律文本、论坛评论" }, { "name": "中国机构署址", ...
50	6,000	Place name Seoul	The most stable strong trigger is the place name Seoul, often in contexts involving Korean addresses, institutions, people’s birthplaces, or place lists. Some scientific terms are also marked, but are sporadic and difficult to confirm as an independent pattern; segment 10 has no Seoul or other Korean cues.	related	4	[ { "name": "Seoul token activation", "trigger": "«Seoul»10", "context": "Strongly activates when the place name Seoul appears, often adjacent to a country name, institution name, or punctuation. It covers most, but not all, segments.", "domain": "Academic papers, parliamentary texts, encyclopedias, i...	[ 10 ]	0.5	[ 0.23000000417232513, 0.48660001158714294 ]	{ "concept": "地名 Seoul", "description": "最稳定的强触发是地名 Seoul，常处于韩国地址、机构、人物出生地或地点名录语境。若干科学术语也被标记，但分布零散，难以确认是独立模式；第10段没有 Seoul 等韩国线索。", "patterns": [ { "name": "Seoul 词元激活", "trigger": "«Seoul»10", "context": "出现地名 Seoul 时强激活；常邻接国家名、机构名或标点。覆盖多数片段，但不是全部。", "domain": "学术论文、议会文本、百科、机构名录等" ...
50	7,000	Beijing place name and institutional associations	All spans contain Beijing in different contexts, including author affiliations, institutional addresses, research or medical institutions, law enforcement agencies, and travel destinations; several spans also show Chinese address and postal-code formats.	single	5	[ { "name": "Beijing-related place or institution name", "trigger": "«Beijing»10", "context": "Appears in a place name, institutional address, or full institution name, often near a postal code and China; also occurs in accounts of events or places in Beijing.", "domain": "Academic paper affiliations,...	[]	null	[ 0.3653999865055084, 0.6000999808311462 ]	{ "concept": "北京地名及机构关联", "description": "所有段都在不同语境中出现 Beijing，涵盖作者单位、机构地址、科研或医疗机构、执法部门及旅行目的地；多段还呈现中国地址与邮编格式。", "patterns": [ { "name": "北京相关地点或机构名称", "trigger": "«Beijing»10", "context": "在地名、机构地址或机构全称中出现，常与邮编及 China 相邻；也出现在关于北京事件或地点的叙述中。", "domain": "学术论文作者单位、新闻/论坛帖子、法律文本、医学论文" }...
50	8,000	Amsterdam place name and related contexts	The most stable trigger is the place name Amsterdam, commonly found in addresses, parliamentary texts, city lists, and medical papers. Other activations are scattered across academic bibliographies and terminology, political phrasing, web links, and population or medical descriptions; no unified local syntactic feature...	unclear	2	[ { "name": "Amsterdam place name", "trigger": "«Amsterdam»10", "context": "The city name appears in institutional addresses, political texts, market reports, research findings, or city classifications. It can occur independently or alongside other cities.", "domain": "Academic addresses and literatur...	[]	0.875	[ 0.2085999995470047, 0.694599986076355 ]	{ "concept": "Amsterdam 地名及相关语境", "description": "最稳定的触发项是地名 Amsterdam，常见于地址、议会文本、城市列表及医学论文；其余激活分散在学术书目与术语、政治措辞、网页链接和人口或医学表述中，未见统一的局部语法特征。", "patterns": [ { "name": "Amsterdam 地名", "trigger": "«Amsterdam»10", "context": "城市名称出现在机构地址、政治文本、市场报道、研究结果或城市分类中；可独立出现，也可与其他城市并列。", "domain": "学术...
50	10,000	Place names Brisbane/London	Activation centers on the place names Brisbane and London: the former appears in Australia-related accounts, while the latter occurs in institutional addresses, time-zone arrangements, and case names.	related	4	[ { "name": "Brisbane place name", "trigger": "«Brisbane»", "context": "The place name appears in accounts involving schools, sporting events, people's appointments, or organizations; in the provided spans, it is the most strongly activated token.", "domain": "Encyclopedic and philosophical texts", ...	[]	null	[ 0.18559999763965607, 0.803600013256073 ]	{ "concept": "地名 Brisbane/London", "description": "激活集中在地名 Brisbane 与 London：前者出现于澳大利亚相关叙述，后者出现于机构地址、时区安排和案件名称。", "patterns": [ { "name": "布里斯班地名", "trigger": "«Brisbane»", "context": "地名出现在学校、体育赛事、人物任职或组织等叙述中；在所给段落中均为激活最强的 token。", "domain": "百科、哲学文本" }, { "name": "伦敦地名"...
50	12,000	ENG string fragments	High activation often falls on the string ENG and its fragments in words such as engineering and engine, and also appears in proper nouns. Other activations span technology, web, legal, automotive, humanities, and historical texts, with no single stable semantic trigger.	unclear	2	[ { "name": "ENG string and its segmentation", "trigger": "«ENG»", "context": "ENG appears inside words or in adjacent fragments, such as «ENG»INEERING and «ENG»lish string fragments; the marker does not always correspond to a complete word", "domain": "Multidomain web, technical text, and corpus frag...	[]	0.5	[ 0.2937999963760376, 0.7402999997138977 ]	{ "concept": "ENG 字符串片段", "description": "高激活常落在 ENG 字符串及其在 engineering、engine 等词中的片段，也出现在专名中；其余覆盖科技、网页、法律、汽车、人文与历史文本，未见单一稳定语义触发。", "patterns": [ { "name": "ENG 字符串及其切分", "trigger": "«ENG»", "context": "ENG 出现在单词内部或相邻片段中，如«ENG»INEERING、«ENG»lish 字符串片段；标记并不总对应完整词", "domain": "多领域网页、技术文本...
50	14,000	Brisbane and ENG strings	The clearest pattern is responses to the proper place name Brisbane; in several unrelated texts, there are also responses to the uppercase character fragment ENG, including in citations, web domains, engineering terms, and personal names.	unrelated	3	[ { "name": "Brisbane place name", "trigger": "«Brisbane»10", "context": "The place name Brisbane appears in names of institutions, venues, or event locations; often follows prepositions such as “in Brisbane” or “at Brisbane,” and also occurs in compound proper names.", "domain": "Australian history, ...	[]	null	[ 0.3046000003814697, 0.7659000158309937 ]	{ "concept": "Brisbane 与 ENG 字符串", "description": "最清晰的模式是响应专名地名 Brisbane；另在若干无关文本中响应大写字符片段 ENG，包括引文、网页域名、工程字样和人名。", "patterns": [ { "name": "Brisbane 地名", "trigger": "«Brisbane»10", "context": "地名 Brisbane 出现在机构、场馆或事件地点等名称中；常在“in Brisbane”“at Brisbane”等介词后，也出现在复合专名中。", "domain": "澳大利亚...
50	16,000	Place names and geographic proper nouns	The main responses are to the proper place names Brisbane and London: the former is concentrated in Australia-related texts, while the latter appears in academic addresses, email time zones, and legal company names. Other Australian locations, institution names, and email time formats also activate.	related	3	[ { "name": "Brisbane place name", "trigger": "«Brisbane»", "context": "The place name appears in accounts of people, institutions, events, or history related to Queensland, Australia; the specific topics vary", "domain": "Encyclopedias, philosophical texts, and Australian sports history", "segmen...	[]	1	[ 0.15139999985694885, 0.6794999837875366 ]	{ "concept": "地名及地域专名", "description": "主要响应专名地名 Brisbane 与 London：前者集中于澳大利亚相关文本，后者见于学术地址、邮件时区和法律公司名；其他澳洲地点、机构名及邮件时间格式也有激活。", "patterns": [ { "name": "布里斯班地名", "trigger": "«Brisbane»", "context": "地名出现在澳大利亚昆士兰相关人物、机构、赛事或历史叙述中；前后具体主题各异", "domain": "百科、哲学文本及澳大利亚体育史" }, { "n...
50	20,000	Brisbane and ENG strings	Strong activation centers on the place name Brisbane (1–6) and the uppercase string ENG (7–10); the former is mostly associated with Australia, Queensland, and cricket contexts, while the latter occurs across paper citations, web pages, and within-word strings. Several related geographic and sports terms also activate.	unrelated	3	[ { "name": "Brisbane place name", "trigger": "« Brisbane»10", "context": "The place name Brisbane appears in sentences; contexts may discuss people, institutions, sporting events, or historical events.", "domain": "Encyclopedic, philosophical/commentary texts", "segments": [ 1, 2, ...	[]	null	[ 0.22990000247955322, 0.732699990272522 ]	{ "concept": "Brisbane 与 ENG 字符串", "description": "强激活集中于 Brisbane 地名（1–6）及大写 ENG 字符串（7–10）；前者多与澳大利亚、昆士兰和板球语境相连，后者跨论文引文、网页与词内字符串出现，另有若干地域和体育词伴随激活。", "patterns": [ { "name": "Brisbane 地名", "trigger": "« Brisbane»10", "context": "在句子中出现地名 Brisbane；上下文可谈人物、机构、体育赛事或历史事件。", "domain": "百科、哲学...
50	24,000	Specific strings Brisbane and ENG	Primarily responds to the Australian place name Brisbane; it also responds to the uppercase string ENG in various texts, often within quotations, URLs, domains, or labels.	unrelated	3	[ { "name": "Brisbane place name", "trigger": "«Brisbane»", "context": "The place name Brisbane appears in Australian locations, institution names, or event names; it can occur within a noun phrase or after a preposition.", "domain": "Australian history, education, sports, and social issues", "seg...	[]	0.75	[ 0.2669999897480011, 0.7099000215530396 ]	{ "concept": "特定字符串 Brisbane 与 ENG", "description": "主要响应 Brisbane 这一澳大利亚地名；另在多种文本中响应大写字母串 ENG，常处于引用、网址、域名或标签中。", "patterns": [ { "name": "Brisbane 地名", "trigger": "«Brisbane»", "context": "地名 Brisbane 出现在澳大利亚地点、机构或赛事名称中；位置可在名词短语内或介词后。", "domain": "澳大利亚历史、教育、体育及社会议题" }, { ...
50	28,000	Australian place names	The strongest and most consistent trigger is Australian place names, especially Brisbane; Sydney, Queensland, Australia, and others also occur. Coverage includes local history, people, sports, paper addresses, travel, and business texts; some spans contain only related contexts, with the trigger unclear.	related	4	[ { "name": "Australian place name Brisbane", "trigger": "«Brisbane»", "context": "Appears as a place name in accounts involving groups, schools, sporting venues, or matches; the triggering token is within the place name", "domain": "Australian local history, biographies, and sports encyclopedias", ...	[ 1, 4 ]	null	[ 0.2290000021457672, 0.9275000095367432 ]	{ "concept": "澳大利亚地名", "description": "最强且稳定的触发是澳大利亚地名，尤其 Brisbane；也见于 Sydney、Queensland、Australia 等。覆盖地方史、人物、体育、论文地址、旅行和商业文本；部分段落只有相关语境，触发原因不明。", "patterns": [ { "name": "澳大利亚布里斯班地名", "trigger": "«Brisbane»", "context": "作为地点名出现在集体、学校、体育场馆或比赛等叙述中；触发 token 位于地名本身", "domain": "澳大利亚地方历史、...
50	32,000	Place names and proper-noun fragments	The most prominent triggers are the place name Brisbane (1–6) and London (7–10); there are also Australian, Queensland, and cricket-related words, as well as sporadic activations for date and time-zone formats, word-internal fragments, and legal terminology.	related	3	[ { "name": "Brisbane place name", "trigger": "«Brisbane»", "context": "Appears as a place name in sentences about people’s positions, schools, sports events, or local Australian history; it is often preceded by the preposition “in” or a definite article.", "domain": "Australian encyclopedias and spor...	[]	1	[ 0.2150000035762787, 0.9018999934196472 ]	{ "concept": "地名与专名片段", "description": "最显著的触发是地名 Brisbane（1–6）与 London（7–10）；另有澳大利亚、昆士兰及板球词汇，以及零散的日期时区、词内片段和法律术语激活。", "patterns": [ { "name": "Brisbane 地名", "trigger": "«Brisbane»", "context": "在介绍人物任职、学校、体育赛事或澳大利亚地方史的句子中，作为地点名出现；前面常有介词“in”或定冠词。", "domain": "澳大利亚百科与体育史" }, { ...
50	40,000	Place names Brisbane or Wellington	All segments mark the place names Brisbane or Wellington; the former appears in Queensland-related content, while the latter appears in quotations, bridge descriptions, and accounts of local figures. Other words mostly show sporadic co-activation.	related	4	[ { "name": "Brisbane place name", "trigger": "«Brisbane»10", "context": "The place name Brisbane appears in sentences related to Queensland, Australia; trigger words occur in varied contexts, including organizations, schools, sports venues, competitions, and personal experiences.", "domain": "Austral...	[]	null	[ 0.2150000035762787, 0.8172000050544739 ]	{ "concept": "地名 Brisbane 或 Wellington", "description": "所有片段都标记了澳大利亚布里斯班或惠灵顿地名；前者见昆士兰相关内容，后者见引文、桥梁及地方人物叙述，其他词语多为零星伴随激活。", "patterns": [ { "name": "布里斯班地名", "trigger": "«Brisbane»10", "context": "地名 Brisbane 出现在澳大利亚昆士兰相关的句子中；触发词可处于组织、学校、体育场馆、比赛或人物经历等不同上下文。", "domain": "澳大利亚地理、历史、教育、体育及...
50	48,000	Place names Brisbane/Wellington	All examples activate on the place name Brisbane or Wellington: the former appears in Australia-related texts, while the latter appears in bibliographies, US bridge history, and biographies; the shared feature is the place-name string.	single	5	[ { "name": "Brisbane place name", "trigger": "«Brisbane»", "context": "The city of Brisbane in Queensland, Australia, appears near locations, institutions, or event names; no fixed syntactic position is required.", "domain": "Australian biographies, education, sports, and social-history encyclopedia ...	[]	0.5	[ 0.19429999589920044, 0.8378999829292297 ]	{ "concept": "地名 Brisbane/Wellington", "description": "所有样例都激活于地名 Brisbane 或 Wellington：前者见于澳大利亚相关文本，后者见于书目、美国桥梁史和人物传记；共同点是地名字符串。", "patterns": [ { "name": "布里斯班地名", "trigger": "«Brisbane»", "context": "出现澳大利亚昆士兰州城市 Brisbane，常在地点、机构或赛事名称前后；不要求固定句法位置。", "domain": "澳大利亚人物、教育、体育及社会史百科文本；第...
50	52,000	Recognition of proper place names	The strongest shared signals are the proper place names Brisbane and London: Brisbane appears in accounts related to Australia, and London in affiliations, emails, and case names; the remaining activated words are weaker and scattered.	related	4	[ { "name": "Brisbane place name", "trigger": "«Brisbane»10", "context": "The place name Brisbane appears in accounts related to Australia, including institutions, events, biographies, or historical locations; in [1], it is part of Brisbane organising collective.", "domain": "Australian local history,...	[]	null	[ 0.2524999976158142, 0.7409999966621399 ]	{ "concept": "地名专名识别", "description": "最强共同线索是专名地名 Brisbane 与 London：Brisbane 见于澳大利亚相关叙述，London 见于署名、邮件及案件名称；其余激活词较弱且分散。", "patterns": [ { "name": "布里斯班地名", "trigger": "«Brisbane»10", "context": "地名 Brisbane 出现在澳大利亚相关叙述中，包括机构、赛事、人物履历或历史地点；[1]中是 Brisbane organising collective。", "domain...
50	56,000	Sydney and Australian proper names	Every segment contains Sydney (the first segment has the split form Syd-ney), often in connection with Australian cities, institutions, personal names, newspapers, or venues; Australia, Australian, and other Australian place names also co-occur.	related	5	[ { "name": "Sydney place name", "trigger": "«Sydney»10; the tokenized form «Syd»4«ney»2 also appears in paragraph 1", "context": "The place name Sydney appears in institutional addresses, airports, city examples, venue names, or personal names; it often co-occurs with nearby information about Australia, ...	[]	null	[ 0.14259999990463257, 0.7903000116348267 ]	{ "concept": "悉尼及澳大利亚专名", "description": "所有段都出现 Sydney（第1段为 Syd-ney 分词），常与澳大利亚城市、机构、人名、报刊或场馆专名相关；另有 Australia、Australian 及其他澳大利亚地名共现。", "patterns": [ { "name": "Sydney 地名", "trigger": "«Sydney»10；第1段另见分词形式 «Syd»4«ney»2", "context": "地名 Sydney 出现在机构地址、机场、城市举例、场馆名称或人名中；常与 Australia、大学、体育等邻近信息...
50	64,000	Shanghai place name	All fragments contain the place name “Shanghai” in different contexts, often alongside China or the names of Shanghai institutions; activation does not depend on a specific discipline or genre.	single	5	[ { "name": "Shanghai as a place name", "trigger": "«Shanghai»10; often immediately followed by «China»3", "context": "Shanghai appears in addresses, location descriptions, institution locations, or research sites; it may be followed by the name of a city, country, company, or research site", "domain"...	[]	1	[ 0.1826000064611435, 0.6822999715805054 ]	{ "concept": "上海地名", "description": "所有片段均在不同语境中出现地名“Shanghai”，常伴随 China 或上海的机构名称；触发不依赖特定学科或文体。", "patterns": [ { "name": "上海作为地名", "trigger": "«Shanghai»10；常紧接«China»3", "context": "Shanghai 出现在地址、地点说明、机构所在地或研究地点中；后续可能有城市、国家、公司或研究场所名称", "domain": "论文地址、新闻、电邮、网页、论坛及学术论文" } ] }
50	72,000	London place name	All segments mark London; the word appears in institutional addresses, time-zone itineraries, legal and code texts, book bodies, and publication information. Nearby cities, institutions, countries, and publication terms are secondary co-occurrence patterns.	single	5	[ { "name": "London place name", "trigger": "« London»10", "context": "London appears as a place name, typically in a phrase with a person's name, institutional address, schedule location, bibliographic information, or other place names; it can also occur in code strings.", "domain": "Academic papers,...	[]	null	[ 0.24140000343322754, 0.5640000104904175 ]	{ "concept": "伦敦地名", "description": "所有片段均标出 London；该词出现在机构地址、时区行程、法律公司名、代码字符串、书籍正文及出版信息中。邻近的城市、机构、国家或出版用语是次级共现模式。", "patterns": [ { "name": "地名 London", "trigger": "« London»10", "context": "London 作为地名出现，通常位于人名、机构地址、时间表地点、书目信息或与其他地名并列的短语中；在代码字符串中也可出现。", "domain": "学术论文、邮件、法律文本、代码、书籍正...
50	80,000	Australia	All segments show the strongest activation at Australia; contexts include country lists, regional or source labels, sports and institutional statements, as well as non-topical uses such as a person's location and an album title.	single	5	[ { "name": "Australia as a country or region name", "trigger": "«Australia»", "context": "Appears when mentioning a country, international relations, national institutions, or local affiliation; often occurs in parentheses, in lists, or at the start of a sentence.", "domain": "News emails, forum comm...	[]	null	[ 0.23389999568462372, 0.9246000051498413 ]	{ "concept": "澳大利亚", "description": "所有片段均在澳大利亚一词处标出最强激活；上下文包括国家并列、地域或来源标签、体育与机构陈述，也包括个人所在地和专辑标题等非主题性用法。", "patterns": [ { "name": "澳大利亚作为国家或地区名", "trigger": "«Australia»", "context": "在提及国家、国际关系、国家机构或地方归属时出现；常位于括号中、列表中或句首。", "domain": "新闻邮件、论坛评论、百科与网页文本" }, { "name": "国家...
50	88,000	London place name	All segments show strong activation on the token London, covering paper affiliations, email schedules, legal and code text, book publication information, and geographic descriptions in book text; other neighboring words and contexts vary.	single	5	[ { "name": "London place name", "trigger": "«London»", "context": "The place name London appears in varied contexts, including personal and institutional addresses, schedules, legal citations, code strings, book title pages, and body text; it is the shared trigger across all segments.", "domain": "Cr...	[]	null	[ 0.3312000036239624, 0.6198999881744385 ]	{ "concept": "London 地名", "description": "所有片段都在 token London 上有强激活，覆盖论文地址、邮件时刻表、法律与代码文本、书籍出版信息及正文地理叙述；其余邻近词和语境各异。", "patterns": [ { "name": "London 地名", "trigger": "«London»", "context": "地名 London 出现在姓名单位地址、时间表、法律引文、代码字符串、书籍扉页或正文等不同语境；是所有片段的共同触发项。", "domain": "跨领域文本" }, { ...
50	96,000	Proper noun London	All fragments show strong activation on London; it appears across genres, syntax, and topics. Its semantic role varies and includes a city, place of publication, time-zone location, and component of a company name.	single	5	[ { "name": "London place name", "trigger": "«London»", "context": "The directly activating word is the proper noun London; it appears in varied contexts, either alone, alongside England, Westminster, Edinburgh, and other places, or embedded in institutional, time, or bibliographic information.", "dom...	[]	0.5	[ 0.29670000076293945, 0.6575999855995178 ]	{ "concept": "专名 London", "description": "所有片段都在 London 处标出强激活；它跨文体、句法和主题出现。各处语义角色不同，包括城市地点、出版地、时区地点及公司名称组成部分。", "patterns": [ { "name": "地名 London", "trigger": "«London»", "context": "直接激活词是专名 London；在多种上下文中出现，既可单独出现，也可与 England、Westminster、Edinburgh 等并列，或嵌在机构、时间、书目信息中。", "domain": "学...
50	104,000	Australian and British city place names	The place names Brisbane and London are the main activations, spanning institutional locations, event venues, email time zones, and legal case names; no clear shared domain or finer contextual condition is apparent.	related	4	[ { "name": "Brisbane place name", "trigger": "«Brisbane»", "context": "The place name Brisbane appears in contexts such as locations, institutions, places where people held positions, or event venues; preceding text often mentions Queensland, Australia, or a related institution.", "domain": "Encyclop...	[]	null	[ 0.19869999587535858, 0.8619999885559082 ]	{ "concept": "澳大利亚与英国城市地名", "description": "主要对地名 Brisbane 和 London 激活，涵盖机构地点、赛事地点、邮件时区及法律案件名称等不同语境；未见明确的共同领域或更细的上下文条件。", "patterns": [ { "name": "布里斯班地名", "trigger": "«Brisbane»", "context": "在地点、机构、人物任职地点或赛事举办地等上下文中出现地名 Brisbane；前文常提 Queensland、Australia 或相关机构。", "domain": "百科、哲学评论等一...
50	112,000	Australian-related contexts	The core trigger is Australian, covering Australian institution names and regional descriptions of local people, the judiciary, industries, companies, and services. Other markers are scattered and show no consistent independent pattern.	related	4	[ { "name": "Australian institution or organization names", "trigger": "«Australian»; e.g. «Australian» National University, «Australian» Forestry Standard, «Australian» and New Zealand American Studies Association", "context": "“Australian” modifies or forms part of the name of an institution or organiza...	[]	null	[ 0.23680000007152557, 0.9135000109672546 ]	{ "concept": "澳大利亚相关语境", "description": "核心触发词是 Australian，涵盖澳大利亚机构名称及对当地人、司法、产业、企业和服务的地域描述。其余标记分散，未显示一致的独立模式。", "patterns": [ { "name": "澳大利亚机构或组织名称", "trigger": "«Australian»；如 «Australian» National University、«Australian» Forestry Standard、«Australian» and New Zealand American Studies Associati...
50	120,000	Australia place name/regional marker	All segments show the strongest activation on the word Australia, covering country lists, Australian locations and events, parenthetical regional labels, and music, government, and entertainment industry contexts; nearby collocations and styles vary.	single	5	[ { "name": "Australia place name or country name", "trigger": "«Australia»10", "context": "Australia appears as a country, region, or place name; it can occur in lists of countries, parenthetical geographic labels, category labels, or sentences.", "domain": "News, encyclopedias, web reviews, and emai...	[]	null	[ 0.22990000247955322, 0.9273999929428101 ]	{ "concept": "Australia 地名/地区标记", "description": "所有片段都在 Australia 一词处出现最强激活，涵盖国家名单、澳大利亚地点与赛事、括号地区标注，以及音乐、政府和演艺行业语境；邻近搭配和文体多样。", "patterns": [ { "name": "澳大利亚地名或国家名", "trigger": "«Australia»10", "context": "Australia 作为国家、地区或地点名出现；可在国家并列名单、括号中的地理标注、类别标签或句子中。", "domain": "新闻、百科、网页评论及邮件"...
50	128,000	Australia-related words	All fragments show strong activation on “Australian,” relating to Australian place names, institutions, industries, standards, publications, activities, or identity; its meaning varies by context.	single	5	[ { "name": "Australian place, institution, or affiliation", "trigger": "« Australian»", "context": "Modifies an institution, organization, industry, standard, judicial decision, or identity, or appears in an address, affiliation, or country name.", "domain": "Academic papers, forums, news, business, ...	[]	1	[ 0.22939999401569366, 0.928600013256073 ]	{ "concept": "澳大利亚关联词", "description": "所有片段都在“Australian”一词上标出强激活，涉及澳大利亚的地名、机构、产业、标准、出版物、活动或身份；词义随语境变化。", "patterns": [ { "name": "Australian 地名、机构或所属关系", "trigger": "« Australian»", "context": "修饰机构、组织、产业、标准、司法决策或身份，或出现在地址、署名和国家名称中。", "domain": "学术论文、论坛、新闻、商业与一般文本" }, { ...
50	136,000	Australian-related entities	All segments activate on Australian, covering Australian institutions, nationality, courts, industries, standards, and telephone services; the trigger is consistent, while the specific meaning varies with context.	single	5	[ { "name": "Australia as a country or nationality modifier", "trigger": "«Australian»", "context": "Appears before nouns referring to institutions, locations, companies, courts, industries, standards, or people/nationality, generally meaning Australian or of Australia; capitalization and punctuation may ...	[]	null	[ 0.22859999537467957, 0.9258000254631042 ]	{ "concept": "澳大利亚相关实体", "description": "所有段落均在 Australian 上激活，涉及澳大利亚机构、国籍、法院、行业、标准或电话服务；触发点一致，具体语义随上下文变化。", "patterns": [ { "name": "澳大利亚作为国家或国籍修饰语", "trigger": "«Australian»", "context": "出现在机构、地点、公司、法院、行业、标准或人物/国籍相关名词前，通常指澳大利亚的或澳大利亚人；前后可有大小写及标点变化。", "domain": "学术论文、新闻评论、商业与林业文本、个人叙述...
50	143,000	Australia-related words	All fragments show strong activation on « Australian»; in different contexts, it marks Australian institutions, nationality, locations, law, culture, industry, standards, or personal experiences. No shared trigger pattern independent of the word is apparent.	single	5	[ { "name": "Australian geographic or national modifier", "trigger": "« Australian»10", "context": "Appears directly before a noun to indicate an Australian institution, location, nationality, industry, standard, or related attribute; the word itself is the shared strong trigger across all passages.", ...	[]	1	[ 0.22939999401569366, 0.928600013256073 ]	{ "concept": "澳大利亚相关词", "description": "所有片段均强激活于« Australian»；它在不同语境中标记澳大利亚机构、国籍、地点、法律、文化、产业、标准或个人经历，未见独立于该词的共同触发模式。", "patterns": [ { "name": "Australian（澳大利亚）地域或国籍修饰语", "trigger": "« Australian»10", "context": "直接出现在名词前，标示澳大利亚机构、地点、国籍、产业、标准或相关属性；词本身是各段共同的强触发项。", "domain": "学术论文署名与期刊...
125	0	weaken or diminish	The core trigger word is weaken: it means reducing an object's strength, function, connection, influence, or ability; in storm reports, it also refers to a cyclone weakening or being downgraded. The objects and causes vary across domains.	single	5	[ { "name": "Cause to weaken (general use)", "trigger": "«weaken»10; [6] also has «weaken»10, semantically close to «undermine»2; [9] «weaken»10", "context": "The verb means making an ability, effect, relationship, or structure weaker. The object and cause vary by domain; examples include both abstract ef...	[]	0.5625	[ 0.3898000121116638, 0.5091000199317932 ]	{ "concept": "减弱或削弱", "description": "核心触发词是 weaken：表示使对象的强度、功能、联系、影响或能力降低；在风暴报道中也指气旋强度减弱、等级下降。对象和原因跨领域变化。", "patterns": [ { "name": "使减弱、削弱（一般用法）", "trigger": "«weaken»10；[6] 另有 «weaken»10，其语义近于 «undermine»2；[9] «weaken»10", "context": "动词表示使某种能力、作用、关系或结构变弱，宾语及原因随领域而变化；包括抽象影响，也包括生物信号关系。", ...
125	2	weaken / weaken	All segments activate on weaken, covering the weakening of different objects, including storms, cells, political groups and institutions, muscles, damping functions, predictive relationships, and syntactic requirements. The commonality is a decrease in some kind of intensity, effect, or dependency.	single	5	[ { "name": "Tropical cyclone weakening", "trigger": "«weaken»10", "context": "Storm intensity decreases under conditions such as wind shear or landfall, often followed by a result such as “to a tropical depression”.", "domain": "Meteorology / tropical cyclone encyclopedia text", "segments": [ ...	[]	null	[ 0.4401000142097473, 0.3946000039577484 ]	{ "concept": "变弱／削弱", "description": "所有段都激活于 weaken，涵盖风暴、细胞、政治群体与制度、肌肉、阻尼功能、预测关系及句法要求等不同对象的变弱；共同点是某种强度、作用或依存关系下降。", "patterns": [ { "name": "热带气旋减弱", "trigger": "«weaken»10", "context": "风切变或登陆等条件下，风暴强度下降，常接续“to a tropical depression”等结果", "domain": "气象／热带气旋百科文本" }, { "n...
125	8	weaken	All passages activate on the same word form, weaken, but the objects and semantic domains differ: syntactic requirements, cell adhesion, storms, groups, institutions, muscles, damping, signal relationships, and cell connections are weakened.	single	5	[ { "name": "The word “weaken”", "trigger": "«weaken»10", "context": "The direct trigger word is weaken; it occurs in different contexts across the passages, covering physical, functional, relational, political, and biological senses.", "domain": "Multidisciplinary narrative and academic texts", "...	[]	null	[ 0.32429999113082886, 0.6633999943733215 ]	{ "concept": "削弱", "description": "全部段落都激活在同一词形 weaken 上，但其宾语和语义领域不同：句法需求、细胞黏附、风暴、群体、机构、肌肉、阻尼、信号关系及细胞连接等被削弱。", "patterns": [ { "name": "词语“weaken”", "trigger": "«weaken»10", "context": "直接触发词是 weaken；各段中出现在不同语境，涵盖物理、功能、关系、政治与生物学意义。", "domain": "多领域叙述与学术文本" } ] }
125	16	weaken, diminish	Activation is almost always triggered by weaken, covering the weakening of syntactic requirements, cell adhesion, storms, groups, institutions, muscles, damping, signaling relationships, and cell junctions; one passage also uses the synonym undermine.	related	5	[ { "name": "Lexical trigger weaken", "trigger": "«weaken»10", "context": "The word directly means to weaken or diminish, or to make a function, connection, or object weaker; the trigger does not depend on specific syntax or domain.", "domain": "Cross-domain English expository texts, academic papers, ...	[]	0.5625	[ 0.41679999232292175, 0.4805999994277954 ]	{ "concept": "削弱、减弱", "description": "激活几乎都由 weaken 触发，涵盖句法需求、细胞黏附、风暴、群体、制度、肌肉、阻尼、信号关系及细胞连接的减弱；另在一段出现近义词 undermine。", "patterns": [ { "name": "weaken 的词汇触发", "trigger": "«weaken»10", "context": "该词直接表示削弱、减弱或使某种功能、联系、对象变弱；触发不依赖特定句法或领域。", "domain": "跨领域英语说明文、学术论文、历史与新闻文本" }, { ...
125	32	weaken	All passages activate on weaken; its objects include syntactic requirements, cell adhesion, storms, the influence of ethnic groups, international mechanisms, muscles, shock-absorbing functions, predictive relationships, and cell connections. It means making an object weaker or diminishing its effect.	single	5	[ { "name": "Weaken or diminish a process or state", "trigger": "“weaken”", "context": "The base verb weaken is followed by the object being weakened; the object may be a storm, function, relationship, bond, group, or physiological structure.", "domain": "Cross-domain explanatory, academic, or narrati...	[]	null	[ 0.5005999803543091, 0.44200000166893005 ]	{ "concept": "削弱", "description": "全部段落都在 weaken 上激活；其宾语涵盖句法需求、细胞黏附、风暴、族群影响力、国际机制、肌肉、减震功能、预测关系及细胞连接，表示使对象变弱或削弱其作用。", "patterns": [ { "name": "削弱、减弱某种过程或状态", "trigger": "“weaken”", "context": "动词原形 weaken 后接被削弱的对象；对象可为风暴、功能、关系、键合、群体或生理结构。", "domain": "跨领域说明性、学术或叙述文本" } ] }
125	64	weaken/diminish	All snippets activate on the verb “weaken,” meaning to make an object weaker, diminish its function or relationship, or become weaker. Objects range across syntactic requirements, cell adhesion, storms, political influence, institutional mechanisms, muscle groups, damping functions, predictive relationships, and cell j...	single	5	[ { "name": "Weaken, diminish, or impair a function/relationship", "trigger": "“weaken”", "context": "“Weaken” is used as a verb followed by the object being weakened; it may refer to a storm, force, effect, relationship, or structure, without requiring a specific sentence pattern.", "domain": "Cross-...	[]	0.5625	[ 0.5922999978065491, 0.4636000096797943 ]	{ "concept": "削弱/减弱", "description": "所有片段都在动词“weaken”处激活，表示使某对象变弱、削弱其功能或关系，或变得较弱。对象跨越句法需求、细胞黏附、风暴、政治影响、机构机制、肌群、阻尼功能、预测关系及细胞连接。", "patterns": [ { "name": "削弱、降低或破坏功能/关系", "trigger": "“weaken”", "context": "“weaken”作为动词，后接被削弱的对象；所指可为风暴、力量、作用、关系或结构，并不要求特定句式。", "domain": "跨领域说明性 prose，包括哲...
125	128	weaken or diminish	All passages activate on weaken: the word describes the weakening of an object's effect, intensity, structure, or relationship. The trigger semantics are stable, while the objects and domains vary widely.	single	5	[ { "name": "Weaken or diminish an effect or relationship", "trigger": "“weaken”; paragraph 2 also has “weaken or completely lose”", "context": "Make an existing effect, ability, relationship, or structure weaker; objects may include adhesion bonds, storms, group strength, muscles, damping functions, sign...	[]	0.5625	[ 0.4068000018596649, 0.42809998989105225 ]	{ "concept": "削弱或减弱", "description": "全部段落均在 weaken 处激活：该词用于描述不同对象的作用、强度、结构或关系变弱，触发语义稳定，但具体对象和领域多样。", "patterns": [ { "name": "削弱、减弱某种作用或关系", "trigger": "“weaken”；第2段另有“weaken or completely lose”", "context": "使原本存在的作用、能力、关系或结构变弱；对象可为黏附键、风暴、群体力量、肌肉、阻尼功能、信号关系、细胞连接等", "domain": "学术、新闻、专利...
125	256	Numeric strings and local lexical features	Most passages activate on numeric strings, especially “0059” and adjacent digits, covering mathematical numbers, years, citation anchors, project IDs, coordinates, filenames, and URLs; passage 10 instead activates on ordinary words in a discussion of syntax and does not fit the numeric pattern.	unclear	2	[ { "name": "Literal substring “0059” after a period", "trigger": "“0059” or “.0059”, such as “0.27” after “...2.27”, “...2.15.6”, “...0059 1075”", "context": "The trigger substring appears as consecutive digits within numbers, decimals, or longer numeric strings; it can be triggered in different position...	[]	0.3125	[ 0.39489999413490295, 0.6442000269889832 ]	{ "concept": "数字串与局部词语特征", "description": "多数段在数字串、尤其是“0059”及相邻数字中激活，涵盖数学数字、年份、引文锚点、项目编号、坐标、文件名和 URL；第10段则在句法论述的普通词上激活，无法归入数字模式。", "patterns": [ { "name": "句点后的“0059”字面子串", "trigger": "“0059”或“.0059”，如“...2.27”后的“0.27”、“...2.15.6”“...0059 1075”", "context": "数字、十进制数或长数字串中，触发子串以连续数字形式出现；不同位置和...
125	512	Mixed numeric and structured-text cues	Activation is distributed across multiple text types. A prominent local cue is numeric strings, especially those containing “0059” and “00”, but math problems, citations, code, medical records, and article text also have distinct trigger structures, so they cannot be grouped into one semantic pattern.	unclear	2	[ { "name": "Specific digit strings in numeric sequences", "trigger": "«0059», «00», «0060»", "context": "These character sequences appear within numbers or identifiers in math problems, bibliographic identifiers, URLs, and filenames.", "domain": "Cross-domain numeric text", "segments": [ 1,...	[]	0.625	[ 0.445499986410141, 0.5457000136375427 ]	{ "concept": "混合的数字与结构化文本线索", "description": "激活分布于多类文本；醒目的局部线索是数字串，尤其含“0059”“00”，但数学题、引文、代码、医学记录和论文正文也有各自不同的触发结构，无法归为单一语义模式。", "patterns": [ { "name": "数字序列中的特定数字串", "trigger": "«0059»、«00»、«0060»", "context": "数字或编号的中间部分含有这些字符序列；数学题、文献标识符、网址和文件名中均出现", "domain": "跨领域数字文本" }, {...
125	1,000	execut stem	Activation is concentrated on the execut stem and its continuations: it covers legal estate executors, executability/files and executors in computing, and Portuguese job titles and verb forms; meanings vary by context.	related	4	[ { "name": "Legal estate executor", "trigger": "«execut»10«ors»3", "context": "In legal proceedings or estate-tax and estate-administration contexts, refers to an executor of a will; the feminine form «execut»10«rix»1 also occurs.", "domain": "U.S. legal documents", "segments": [ 2, 3...	[]	0.8125	[ 0.46540001034736633, 0.49570000171661377 ]	{ "concept": "execut 词干", "description": "激活集中在 execut 词干及其延续部分：涵盖法律遗产执行人、计算机中的可执行性/文件与执行者，以及葡语职务和动词形式；上下文含义并不一致。", "patterns": [ { "name": "法律遗产执行人", "trigger": "«execut»10«ors»3", "context": "法律诉讼或遗产税、遗产处置语境中，指遗嘱执行人；也出现女性形式 «execut»10«rix»1", "domain": "美国法律文书" }, { "na...
125	2,000	character substring “amel”	All activations fall on the consecutive characters “amel”: in Dutch verzamelen, the code field Namel, the username amelius, and the same substring in Megamelus, Caramel, and Kamel. Semantics and styles vary.	single	5	[ { "name": "Dutch word ending -elen", "trigger": "«verz»2«amel»10en", "context": "In “verzamelen” (to collect), “amel” at the end of the stem is joined to the suffix “en”; in context, it is an infinitive/plural form.", "domain": "EuroParl, Dutch parliamentary text", "segments": [ 1, 2...	[]	1	[ 0.17900000512599945, 0.7468000054359436 ]	{ "concept": "字符子串“amel”", "description": "激活都落在连续字符“amel”上：包括荷兰语verzamelen、代码字段Namel、用户名amelius，以及Megamelus、Caramel、Kamel中的同一子串；语义与文体多样。", "patterns": [ { "name": "荷兰语词尾 -elen", "trigger": "«verz»2«amel»10en", "context": "“verzamelen”（收集）中词干末尾的“amel”与后缀“en”相连；在上下文中是动词不定式/复数形式。", "doma...
125	3,000	the in-word letter string “oly”	All segments mark a local span within a word containing “oly” or a similar letter sequence; these words span various domains and meanings, with no stable semantic or stylistic trigger condition evident.	single	4	[ { "name": "Words containing the substring “oly” or a variant", "trigger": "Multiple segments mark «oly» inside words, such as “am«oly»10iod”, “Klypin”, “Ver«kh»«oly»ak”, “bef«oly»ás”, “Aut«oly»cus”, “m«r_g«oly»adkin”, “«y»...«z»...”, “Kh«oly»”, “unh«oly»”, “m«oly»bdenum”; also word-internal fragments like “...	[]	null	[ 0.17820000648498535, 0.7398999929428101 ]	{ "concept": "词内“oly”字母串", "description": "所有片段均在词内部标记出含“oly”或近似字母片段的局部；这些词横跨多种领域与词义，未见稳定的语义或文体触发条件。", "patterns": [ { "name": "含子串“oly”或其变体的词", "trigger": "多段在词内标记出 «oly»，如“am«oly»10iod”“Klypin”“Ver«kh»«oly»ak”“bef«oly»ás”“Aut«oly»cus”“m«r_g«oly»adkin”“«y»...«z»...”“Kh«oly»”“unh«oly»”“m«oly»bdenu...
125	4,000	substring containing pl	The shared local trigger in all passages is the substring “pl”; it appears in English scientific words, target sequences in probability problems, and words in multilingual European Parliament texts. Word meanings and contexts vary widely, with no stable semantic or stylistic condition.	single	5	[ { "name": "pl substring in English mathematical sequence problems", "trigger": "«pl»10, followed by «lp» or p, as in “«pl»10«lp»3”, “«pl»10p”", "context": "Appears as a letter sequence whose probability is to be calculated; typically asks for the probability of obtaining that sequence when drawing sever...	[]	0.9375	[ 0.42480000853538513, 0.5034000277519226 ]	{ "concept": "包含 pl 的子串", "description": "所有段落的共同局部触发是子串“pl”；它出现在英语科学词、概率题中的目标序列，以及多语言欧洲议会词语内部。具体词义和上下文差异很大，未显示稳定的语义或文体条件。", "patterns": [ { "name": "英文数学序列题中的 pl 子串", "trigger": "«pl»10，后接«lp»或p，如“«pl»10«lp»3”“«pl»10p”", "context": "在概率题中作为待计算概率的字母序列，通常询问不放回抽取若干字母时出现该序列的概率", "domain":...
125	5,000	the in-word letter sequence “pl”	All segments mark the “pl” sequence inside words, spanning various languages, domains, and meanings; there is no evidence that activation depends on a particular topic or syntax.	single	5	[ { "name": "In-word spans beginning with pl", "trigger": "«pl»", "context": "Activation centers on the “pl” letter sequence inside words; it occurs across languages, word forms, and meanings, often at the beginning or middle of a word.", "domain": "Cross-domain; including astrophysics, mathematical p...	[]	null	[ 0.46059998869895935, 0.4390000104904175 ]	{ "concept": "词内字母序列“pl”", "description": "所有片段都标出单词内部的“pl”序列，跨越多种语言、领域和语义；没有证据表明激活依赖特定主题或句法。", "patterns": [ { "name": "pl-开头的词内片段", "trigger": "«pl»", "context": "激活集中在单词内部的“pl”字母序列；不同语言、词形和语义中均出现，常位于词首或词中。", "domain": "跨领域；包括天体物理、数学、丹麦语、瑞典语、荷兰语、斯洛文尼亚语及政治演讲" }, { "name"...
125	6,000	letter sequence “pl”	All snippets activate in words containing the letter sequence “pl”; meanings, languages, and domains vary, with no evident, more specific shared semantic condition.	single	5	[ { "name": "pl in English scientific contexts", "trigger": "«pl»10, embedded in “plowing”", "context": "Refers to the process of a galaxy moving through the intracluster medium; the preceding text discusses an astrophysical phenomenon.", "domain": "ArXiv astrophysics", "segments": [ 1 ]...	[]	0.875	[ 0.586899995803833, 0.2612000107765198 ]	{ "concept": "字母序列“pl”", "description": "所有片段都在含字母序列“pl”的词中激活；词义、语言和领域多样，没有明显更具体的共同语义条件。", "patterns": [ { "name": "英文科学语境中的 pl", "trigger": "«pl»10，嵌在“plowing”中", "context": "指星系穿行星系团介质的过程；前文讨论天体物理现象", "domain": "ArXiv 天体物理" }, { "name": "数学文本中的 pl 子串", "trigger": ...
125	7,000	the character sequence pl	All segments mark the same « pl» substring, followed by different word forms and appearing in different languages and domains; the examples support a response to the character sequence pl, but do not establish further contextual conditions.	single	5	[ { "name": "pl substring in a word or stem", "trigger": "« pl»10; e.g. plowing, plumbings, pllp, plads, plundrade, plenaire, plicht, plenarsm, plini", "context": "Activation centers on the consecutive characters pl within a word, often at the beginning, but also in internal spans (e.g., to«pl»5og«red»1ni...	[]	null	[ 0.6480000019073486, 0.15549999475479126 ]	{ "concept": "字符序列 pl", "description": "所有片段都标记了相同的 « pl» 子串，后续词形及语种、领域各异；现有例子支持对字符序列 pl 的响应，无法进一步确定其上下文条件。", "patterns": [ { "name": "单词或词干中的 pl 子串", "trigger": "« pl»10；如 plowing、plumbings、pllp、plads、plundrade、plenaire、plicht、plenarsm、plini", "context": "激活集中在词内的连续字符 pl，常出现在词首，但也有词内片段（如 to...
125	8,000	Letter sequence pl	All passages activate on the substring “pl” in different languages or words, usually word-initially; semantics, topics, and styles vary widely, with no consistent semantic trigger.	single	5	[ { "name": "pl substring in English words", "trigger": "« pl»10（e.g. “pl»10owing”“pl»10umb«ings»”“pl»10lp”“pl»10p”）", "context": "Activation falls on the word-initial or internal letter substring “pl” in English words; no specific semantic or syntactic context is required.", "domain": "Astrophysics, ...	[]	0.9375	[ 0.5128999948501587, 0.36800000071525574 ]	{ "concept": "字母序列 pl", "description": "所有段落都在不同语言或词汇中激活子串“pl”，通常位于词首；语义、主题和文体差异很大，未见一致的语义触发条件。", "patterns": [ { "name": "英语单词中的 pl 子串", "trigger": "« pl»10（如“pl»10owing”“pl»10umb«ings»”“pl»10lp”“pl»10p”）", "context": "激活落在英语单词的词首或内部字母子串“pl”；没有显示出对特定语义或句法环境的要求。", "domain": "天体物理、数学题" ...
125	10,000	cross-linguistic “pl” subword fragment	All passages mark the same fragment “pl”, which appears across disciplines and languages in different words: plowing, plumbings, sequence pllp, plp, plads, plundrade, plenaire, pligt, plenarsammant…, plini. Activation appears to correspond primarily to an in-word character fragment rather than a particular meaning or s...	single	5	[ { "name": "In-word pl fragment", "trigger": "«pl»10", "context": "The letter fragment pl occurs inside or at the beginning of words in different languages; the surrounding word forms and meanings vary.", "domain": "English academic texts, Danish/Swedish/Dutch/Slovenian parliamentary texts, and math ...	[]	null	[ 0.3750999867916107, 0.5968999862670898 ]	{ "concept": "跨语言的“pl”词片段", "description": "所有段落都标记了同一片段“pl”，它跨学科、跨语言出现在不同词中：plowing、plumbings、sequence pllp、plp、plads、plundrade、plenaire、pligt、plenarsammant…、plini。激活似乎主要对应词内的字符片段，而非特定语义或文体。", "patterns": [ { "name": "pl-词内片段", "trigger": "«pl»10", "context": "字母片段 pl 出现在不同语言词语内部或词首；周围词形和语义...
125	12,000	The string “ pl”	All segments activate on the string “ pl”; the segment occurs across languages and word forms, with no stable semantic or contextual conditions.	single	5	[ { "name": "Word-initial pl segment", "trigger": "« pl»10", "context": "Appears in various complete words, such as “plowing,” “plumbings,” “plp,” “plp,” “plenaire,” “plicht,” “plenum,” “plundrade,” and “plini”; there are no stable lexical or syntactic conditions.", "domain": "Astrophysics, math probl...	[]	1	[ 0.6205999851226807, 0.1979999989271164 ]	{ "concept": "字母串“ pl”", "description": "所有片段都激活在字符串“ pl”上；该片段跨越多种语言和词形，未显示出稳定的语义或语境条件。", "patterns": [ { "name": "pl 词首片段", "trigger": "« pl»10", "context": "出现在多个不同完整词中：如“plowing”“plumbings”“plp”“plp”“plenaire”“plicht”“plenum”“plundrade”“plini”；无稳定的词义或句法条件。", "domain": "天体物理、数学题、欧洲议会...
125	14,000	substring ful	Activation consistently falls on the substring ful in various spellings and word forms, including the fullerene family, fully/fulfil, and fullness, Fulgoroid, fulcrum, and fulminant. This is mostly cross-lexical substring matching, with substantial differences in meaning and context.	related	4	[ { "name": "fullerene stem", "trigger": "« ful» in « ful»10lerenes and « ful»10leride", "context": "Matches ful/ful-l spans, especially in fullerene, fullerenes, and fulleride; often occurs in paper introductions, reference lists, or materials science discussions.", "domain": "ArXiv scientific papers...	[]	null	[ 0.15279999375343323, 0.7861999869346619 ]	{ "concept": "含 ful 子串", "description": "激活稳定落在多种拼写和词形中的子串 ful：包括 fullerene 词族、fully/fulfil，以及 fullness、Fulgoroid、fulcrum、fulminant 等。多为跨词汇的子串匹配，语义和语境差异很大。", "patterns": [ { "name": "fullerenes 词干", "trigger": "« ful»10lerenes、« ful»10leride 中的 «ful»", "context": "命中 ful/ful-l 的片段，尤其是 fuller...
125	16,000	In-word pl string	All segments mark the contiguous in-word characters pl; different letters follow it, and it occurs across languages, word forms, and domains. The available context shows no stable semantic or topical conditions.	single	5	[ { "name": "pl segment in English words", "trigger": "« pl»10", "context": "Activation centers on the in-word pl substring; complete words vary by context and include plowing, plumbings, sequence pllp, sequence plp, plads, plundrade, plenaire, plicht, plenum, and plini.", "domain": "Cross-domain; sci...	[]	0.9375	[ 0.6319000124931335, 0.1761000007390976 ]	{ "concept": "词内 pl 字符串", "description": "所有片段都标记了词内连续字符 pl；其后接不同字母，在多种语言、词形和领域中出现。现有上下文未显示稳定的语义或主题条件。", "patterns": [ { "name": "英文单词片段 pl", "trigger": "« pl»10", "context": "激活集中在词内的 pl 子串；完整词因上下文而异，包括 plowing、plumbings、sequence pllp、sequence plp、plads、plundrade、plenaire、plicht、plenum 和 pl...
125	20,000	word-initial “pl” substring	All segments mark the same substring, “pl”, at the beginning of a word; the following word forms and languages vary across astrophysics, mathematics, parliamentary, and policy texts, with no stable semantic or stylistic condition evident.	single	5	[ { "name": "Word-initial “pl” fragment in English technical texts", "trigger": "« pl»10, continued as “plowing”“plumbings”“pllp”“plp”", "context": "Activation centers on the word-initial “pl” fragment, followed by different word forms; some occurrences are in sentences and others in math problems.", ...	[]	null	[ 0.5533000230789185, 0.2750000059604645 ]	{ "concept": "词首“pl”子串", "description": "所有段都标记了同一子串“pl”，且均位于词首；后续词形和语言各异，覆盖天体物理、数学、议会及政策文本，未见稳定的语义或文体条件。", "patterns": [ { "name": "英语科技文本中的“pl”词片", "trigger": "« pl»10，分别续接为“plowing”“plumbings”“pllp”“plp”", "context": "激活集中在词首“pl”片段，后续词形不同；部分出现在句中，部分位于数学题中。", "domain": "英语学术文本、数学题" ...

End of preview. Expand in Data Studio

Pythia-160M Neuron Evolution

How 240 MLP neurons of Pythia-160M change over training: LLM-written explanations at 40 checkpoints, concept lineages, and quantitative metrics for all 36,864 neurons.

Configs

Config	Rows	One row is
`neurons`	240	a sampled neuron: final category/label, lineage-birth shares, trajectory judgments, Claude audit
`explanations`	9,600	a neuron at one checkpoint: concept, description, patterns (each tagged with its lineage), clarity, detection score, semantic-map position
`lineages`	17,575	a concept lineage of a neuron: parent, birth mechanism, activation share at each of the 40 steps
`contexts`	249,600	a probe context: top-10 (with per-token activations), held-out (rank 11–18) and random negatives
`detections`	4,800	a detection-scoring trial: labels, predictions, balanced accuracy

from datasets import load_dataset
ds = load_dataset("LangGao/pythia-160m-neuron-evolution", "explanations", split="train")

Other files

steps.json: the 40 analysis steps, the 20-step subset used by judges and detection (steps_judged), the 240 [layer, neuron] pairs, and probe loss per step.
metrics/all.npz: arrays [154 steps, 12 layers, 3072 neurons] for every neuron at every checkpoint, computed without an LLM: top1, snr, frac_pos, hist_entropy, selectivity, coarse_corr_final, tok_overlap_final, doc_jacc_final, doc_jacc_prev, plus steps.
metrics/ref.npz: coarse, tok, doc similarity of each step to two references (refs = 64000, 143000), shape [2, 154, 12, 3072].
probe/: the 10M-token probe set (stratified sample of the Pile validation split). tokens.npy (uint16), doc_offsets.npy, doc_subset.npy, doc_source.npy (row in monology/pile-uncopyrighted val), subsets.json. contexts.doc / contexts.token_index index into these.

Notes

Model: EleutherAI/pythia-160m (non-deduped). Activation = output of mlp.dense_h_to_4h (pre-GELU). Neurons: 20 per layer, uniform random (seed 0).
Texts were generated by gpt-6-luna (reasoning effort low) with Chinese prompts, then machine-translated to English by the same model. Chinese originals are kept in the zh columns. Quoted source text (tokens, triggers) can itself be non-English.
birth in lineages: branch (derived from an earlier lineage), replace (takes over from a lineage that just died), coexist (unrelated newcomer), sporadic (one step only). Shares are activation-weighted over each step's top-10 contexts.
Trajectory judgments (neurons.recursive, neurons.diffusion) and detection_score cover the 20-step subset only; explanations and lineages cover all 40 steps. claude_audit is set for 24 neurons (2 per layer).
LLM labels are noisy; treat them as annotations, not ground truth.

Downloads last month: -