icoa-cli 2.19.480 → 2.19.482

package/dist/commands/sprint.js

@@ -1 +1 @@
-import{existsSync as o,mkdirSync as e,readFileSync as n,writeFileSync as t}from"node:fs";import{createInterface as l}from"node:readline";import{dirname as a,join as r}from"node:path";import{fileURLToPath as s}from"node:url";import chalk from"chalk";import{getConfig as c}from"../lib/config.js";import{getMainRl as i,setMainRl as g}from"../lib/main-rl.js";import{logCommand as y}from"../lib/logger.js";import{baselineChip as d,computeChip as p,computeLegend as u,loadSprintTasks as h,mediaHelpLines as m,taskSteps as f}from"../lib/sprint-meta.js";let w=!1;export function isSprintActive(){return w}const b=a(s(import.meta.url)),$=r(process.env.HOME||".",".icoa"),k=r($,"sprint-state.json"),S={"T-torch":"裸 torch:手写 attention / causal mask / perplexity","T-umap":"降维当特征提取器,不是画图工具","T-lora":"peft LoRA 参数高效微调","T-librosa":"音频特征流水线:mel / MFCC → 定长向量"},A=["",chalk.bold.cyan("   ASTANA 2027"),chalk.gray("   — 给还没出发的你"),"",chalk.white("   草原的风从额尔齐斯河吹到阿斯塔纳的时候,会带着一点雪的味道。"),chalk.white("   你还没去过那里。但那里已经有一个座位,写着你的名字。"),"",chalk.white("   你现在坐在屏幕前,离比赛还有几天。"),chalk.white("   你可能会想:几天能改变什么?"),"",chalk.white("   足够让一个人从「看着题发呆」变成「看到题就知道该拆哪一层」。"),chalk.white("   足够让手指记住 import torch 之后该敲什么。"),chalk.white("   足够让你在比赛那天早上醒来时,不再害怕。"),"",chalk.bold.yellow("   M")+chalk.white(" —— 你总是第一个说「我先试试」的那个。别停。"),chalk.bold.yellow("   V")+chalk.white(" —— 你总是问「为什么是这样」的那个。继续问。"),chalk.bold.yellow("   A")+chalk.white(" —— 你总是在别人放弃时还在改最后一版的那个。改下去。"),"",chalk.white("   你们三个人,一个人是一条腿走路,三个人是奔跑。"),"",chalk.white("   你们要面对的不是一场考试,是一次证明:"),chalk.bold.white("   中国的高中生,可以在 AI 的赛场上,和世界站在同一排。"),"",chalk.white("   而 2027 年的阿斯塔纳,是我们下一个约定。"),"",chalk.gray("   现在,把这行字关掉,去做第 1 题。"),chalk.gray("   — 熊猫教练"),""];function q(o){try{e($,{recursive:!0}),t(k,`${JSON.stringify(o,null,1)}\n`)}catch{}}function v(){return(c().language||"en").toLowerCase().startsWith("zh")}function j(o,e){return v()?o:e}function I(o){return"★".repeat(Math.max(1,Math.min(5,o||3)))}const C=r($,"sprint-memos.json");function N(o){try{e($,{recursive:!0}),t(C,`${JSON.stringify(o,null,1)}\n`)}catch{}}function O(o,e){const n=new Set(e.done),t=new Set(e.opened);console.log(),console.log(chalk.bold.cyan(j("  ⚔ IOAI 2026 · 四工具冲刺","  ⚔ IOAI 2026 · four-tool sprint"))+chalk.gray(`   ${j("已攻克","cleared")} ${n.size} / ${o.length}`));const l=new Map;for(const e of o)l.has(e.bullseye)||l.set(e.bullseye,[]),l.get(e.bullseye)?.push(e);for(const[o,e]of l){console.log(),console.log(chalk.bold.white(`  ${o}`)+chalk.gray(`  — ${S[o]||""}`));for(const o of e){const e=n.has(o.seq)?chalk.green("✓"):t.has(o.seq)?chalk.yellow("◐"):chalk.gray("⬜");console.log(`    ${e} ${chalk.bold.white(String(o.seq).padStart(2))}  `+chalk.white(o.title.slice(0,34).padEnd(34))+chalk.gray(I(o.difficulty).padEnd(5))+chalk.gray(o.needs_gpu?" ⚡":"  ")+chalk.gray(` ${o.data_mb}M`)+("teleload"===o.status?chalk.gray(j("  · teleload","  · teleload")):"")+(o.judged?chalk.green(j("  · 可出分","  · scored")):""))}}console.log(),console.log(chalk.gray("  ")+chalk.cyan("open <n>")+chalk.gray(j(" 看题  ·  "," view  ·  "))+chalk.cyan("done <n>")+chalk.gray(j(" 标记攻克  ·  "," mark cleared  ·  "))+chalk.cyan("ai4ioai")+chalk.gray(j(" 开 AI 分屏  ·  "," AI in a split  ·  "))+chalk.cyan("memo")+chalk.gray(j(" 记问题  ·  "," report a problem  ·  "))+chalk.cyan("back")+chalk.gray(j(" 退出"," leave"))),console.log(chalk.gray(j("  ⬜ 没看过   ◐ 看过没做完   ✓ 攻克","  ⬜ unseen   ◐ opened   ✓ cleared"))),console.log(chalk.gray(u(v()))),console.log()}function T(o,e){const n=[];let t="",l=0;for(const a of o){const o=/[　-鿿＀-￯]/.test(a)?2:1;l+o>e&&(n.push(t),t="",l=0),t+=a,l+=o}return t&&n.push(t),n}export async function runSprint(){y("sprint");const e=function(){try{const e=h();try{const t=JSON.parse(n(function(e){const n=[r(b,"..","data",e),r(b,"..","..","src","data",e)];for(const e of n)if(o(e))return e;return n[0]}("sprint-cards.json"),"utf8")),l=new Map((t.cards||[]).map(o=>[o.seq,o]));for(const o of e){const e=l.get(o.seq);e&&Object.assign(o,{star:e.star,trap:e.trap,action:e.action,hint:e.hint,meta:e.meta})}}catch{}return e.sort((o,e)=>o.seq-e.seq)}catch{return[]}}();if(0===e.length)return console.log(),console.log(chalk.yellow("  冲刺题单还没装进这个版本。")),console.log(chalk.gray("  升级后再试:")+chalk.cyan("npm install -g icoa-cli@latest")),void console.log();const t=function(){try{const o=JSON.parse(n(k,"utf8"));return{who:o.who||"",opened:o.opened||[],done:o.done||[]}}catch{return{who:"",opened:[],done:[]}}}();if(0===t.opened.length&&0===t.done.length)for(const o of A)console.log(o);t.who||(t.who=""),O(e,t);const a=i(),s=null!==a,c=s?a:l({input:process.stdin,output:process.stdout,terminal:!0});s||g(c);const u=s?c.listeners("line").slice():[];c.removeAllListeners("line");const $=c.prompt.bind(c);w=!0;const S=chalk.bold.cyan("icoa ioai2026> ");c.setPrompt(S);let x=null;await new Promise(o=>{c.on("line",async l=>{const a=l.trim().replace(/^icoa\s+/i,""),[r,i]=[a.split(/\s+/)[0]?.toLowerCase()||"",a.split(/\s+/)[1]||""];if(["back","quit","exit","menu"].includes(r))return q(t),console.log(chalk.gray("  进度已保存。")),(()=>{if(w=!1,c.removeAllListeners("line"),s){for(const o of u)c.on("line",o);c.setPrompt(chalk.green("icoa> ")),c.prompt()}else g(null),c.close()})(),void o();if(""===r||["challenges","list","ls"].includes(r))O(e,t);else if(["open","o"].includes(r)||/^\d+$/.test(r)){const o=Number.parseInt(/^\d+$/.test(r)?r:i,10),n=e.find(e=>e.seq===o);n?(t.opened.includes(o)||t.opened.push(o),q(t),x=o,function(o){console.log(),console.log(chalk.cyan("  ╭─ ")+chalk.bold.white(`${o.seq} · ${o.title}`)+chalk.gray(`  ·  ${o.bullseye}`)+chalk.cyan(" ─────")),console.log(chalk.cyan("  │"));const e=f(o).length>1,n=[o.star,o.trap,e?"":o.action,o.hint?`  ${o.hint}`:""].filter(Boolean);0===n.length&&o.topic_cn&&n.push(`▸ ${o.topic_cn}`);for(const o of n){for(const e of T(o,58))console.log(chalk.cyan("  │ ")+chalk.white(e));console.log(chalk.cyan("  │"))}const t=v();console.log(chalk.cyan("  │ ")+chalk.gray(p(o,t)));const l=d(o,t);if(l)for(const o of T(l,58))console.log(chalk.cyan("  │ ")+chalk.gray(o));console.log(chalk.cyan("  │ ")+chalk.gray(o.meta||`⏱ 难度 ${I(o.difficulty)}`)+chalk.gray(o.needs_gpu?"  ⚡ 需要 GPU":"")),console.log(chalk.cyan("  ╰──────────────────────────────────────────────────────"));for(const e of m(o,t))for(const o of T(e,60))console.log(chalk.yellow("  ")+chalk.yellow(o));const a=f(o),r=function(o){const e=f(o);if(e.length>0)return e[0];const n=/(?:kaggle|teleload)\s+[a-z0-9][a-z0-9-]*/i.exec(o.action||"");return n?n[0]:null}(o);a.length>1?(console.log(chalk.bold.white(j("  ▸ 就在这里敲,两步:","  ▸ Type these here, in order:"))),a.forEach((o,e)=>{console.log(chalk.gray(`    ${e+1}. `)+chalk.bold.cyan(o)+chalk.gray(0===e?j("   ← 先把数据拉下来","   ← fetch the data first"):j("   ← 题面 + notebook","   ← statement + notebook")))})):r&&console.log(chalk.bold.white(j("  ▸ 就在这里敲: ","  ▸ Type it right here: "))+chalk.bold.cyan(r)+chalk.gray(j("   ← 数据 + notebook 落到本机","   ← pulls the data + notebook"))),console.log(chalk.gray("  ")+chalk.cyan("ai4ioai")+chalk.gray(j(" 开 AI 分屏  ·  "," AI in a split  ·  "))+chalk.cyan(`done ${o.seq}`)+chalk.gray(j(" 标记攻克  ·  "," mark cleared  ·  "))+chalk.cyan("memo")+chalk.gray(j(` 记问题(自动记在第 ${o.seq} 题)  ·  `,` report (auto-filed under ${o.seq})  ·  `))+chalk.cyan("back")+chalk.gray(j(" 回列表"," back to list"))),console.log()}(n)):console.log(chalk.yellow(`  没有第 ${i||r} 题。`)+chalk.gray(` 题号 1-${e.length}。`))}else if("done"===r){const o=Number.parseInt(i,10);e.some(e=>e.seq===o)?(t.done.includes(o)||t.done.push(o),q(t),console.log(chalk.green(`  ✓ 第 ${o} 题标记为攻克`)+chalk.gray(`  (${t.done.length}/${e.length})`))):console.log(chalk.yellow(`  题号 1-${e.length}。`))}else["kaggle","teleload","ipynb"].includes(r)?await async function(o,e){const n=e.split(/\s+/).filter(Boolean);try{if("teleload"===o){const{runTeleload:o}=await import("./teleload.js");return void await o(n[0])}if("ipynb"===o){const{runIpynb:o}=await import("./ipynb.js");return void await o(n[0])}const e=await import("./kaggle.js"),t=(n[0]||"").toLowerCase();"submit"===t?await e.runKaggleSubmit(n[1]):"check"===t?await e.runKaggleCheck(n[1]):"guide"===t?await e.runKaggleGuide(n[1]):await e.runKaggle(n[0])}catch(o){console.log(),console.log(chalk.yellow(j("  这条命令没跑成:","  That command failed: "))+chalk.gray(String(o))),console.log(chalk.gray(j("  记一条给教练:","  Tell the coach: "))+chalk.cyan(j("memo …","memo …"))),console.log()}}(r,a.split(/\s+/).slice(1).join(" ")):"memo"===r||"note"===r?await async function(o,e,t,l){const a=function(){try{const o=JSON.parse(n(C,"utf8"));return Array.isArray(o)?o:[]}catch{return[]}}(),r=o.trim();if(["list","ls","all"].includes(r.toLowerCase())){if(console.log(),0===a.length)console.log(chalk.gray(j("  还没有记录。","  Nothing noted yet.")));else{console.log(chalk.bold.white(j(`  📝 已记 ${a.length} 条`,`  📝 ${a.length} noted`)));for(const o of a.slice(-12)){const e=o.seq?chalk.cyan(`#${o.seq}`):chalk.gray("—");console.log(`   ${e} ${chalk.white(o.text.slice(0,60))}`+chalk.gray(o.synced?j("  ·已上报","  ·sent"):j("  ·仅本地","  ·local only")))}console.log(chalk.gray(`  ${C}`))}return console.log(chalk.gray(j("  再记一条:","  Note another: "))+chalk.cyan("memo")+chalk.gray(j("  (光敲这一个词就给你写字的地方)","  (just the word — it gives you a place to type)"))),void console.log()}const s=r?r.split(/\s+/):[],c=s.length>0&&/^\d+$/.test(s[0])&&e.some(o=>o.seq===Number(s[0])),i=c?Number(s[0]):t;let g=(c?s.slice(1):s).join(" ").trim();if(""===g){const o=i?chalk.gray(j(` —— 记在第 ${i} 题下`,` — filed under task ${i}`)):chalk.gray(j(" —— 通用反馈"," — general feedback"));if(console.log(),console.log(chalk.bold.white(j("  📝 撞到什么就写什么","  📝 Say what you hit"))+o),console.log(chalk.gray(j("  一行写完,回车就存下(不用等回复);直接回车 = 不记。","  One line, Enter files it (no reply comes back). Empty Enter cancels."))),g=await function(o,e){return new Promise(n=>{o.question(e,o=>n(o.trim()))})}(l,chalk.cyan("  memo > ")),""===g)return console.log(chalk.gray(j("  没记。","  Nothing noted."))),void console.log()}const d={ts:(new Date).toISOString(),seq:i,text:g,synced:!1};a.push(d),N(a),y(`sprint: memo${i?` ${i}`:""} — ${g}`);let p=!1;try{const{loadLearnState:o}=await import("../lib/learn-state.js"),e=o()?.token??"";if(e){const{syncCardFeedback:o}=await import("../lib/learn-curricula.js");p=await o(e,{curriculum_id:"pandahelpme31",card_number:i??0,feedback_text:g})}}catch{}p&&(d.synced=!0,N(a)),console.log(),console.log(chalk.green(j("  ✓ 记下了","  ✓ noted"))+(i?chalk.gray(j(` (第 ${i} 题)`,` (task ${i})`)):"")+chalk.gray(p?j("  ·已上报","  ·sent"):j("  ·存在本机,教练可以来取","  ·saved on this machine"))),console.log(chalk.gray(j("  继续做题就好。看全部:","  Carry on. See them all: "))+chalk.cyan("memo list")),console.log()}(a.split(/\s+/).slice(1).join(" "),e,x,c):["ai4ioai","/ai4ioai","ai"].includes(r)?(console.log(),console.log(chalk.bold.white(j("  左右分屏,左边看题右边问 AI:","  Split the screen — tasks left, AI right:"))),console.log(chalk.gray("   1. ")+chalk.bold.cyan("Ctrl-b %")+chalk.gray(j("   左右分屏(上下分屏是 ","   split left/right (up/down is "))+chalk.cyan('Ctrl-b "')+chalk.gray(j(")",")"))),console.log(chalk.gray("   2. ")+chalk.gray(j("右边那格输 ","in the right pane type "))+chalk.bold.cyan("icoa")+chalk.gray(j(" ,再输 "," , then "))+chalk.bold.cyan("ai4ioai")),console.log(chalk.gray(j("   3. 两格之间切换:","   3. move between panes: "))+chalk.cyan("Ctrl-b ←")+chalk.gray(" / ")+chalk.cyan("Ctrl-b →")),console.log(),console.log(chalk.gray(j("  它是弱模型 —— 单轮、2000 字符以内。把问题问准,就是这次要练的本事。","  A deliberately weak model — single-turn, 2000 chars. Asking precisely is the skill."))),console.log(chalk.gray(j("  在 notebook 里卡住时,那边可以直接敲 ","  Stuck inside a notebook? there you can type "))+chalk.cyan("ai")+chalk.gray(j("  —— 它会把那一格的代码和报错一起带上。","  — it carries that cell and its traceback."))),console.log()):["help","?","guide"].includes(r)?O(e,t):(console.log(chalk.gray("  ")+chalk.cyan("open <n>")+chalk.gray(j(" 看题 · "," view · "))+chalk.cyan("kaggle <题目id>")+chalk.gray(j(" 领数据 · "," get the data · "))+chalk.cyan("ipynb <路径>")+chalk.gray(j(" 开 notebook · "," open the notebook · "))+chalk.cyan("done <n>")+chalk.gray(j(" 标记 · "," mark · "))+chalk.cyan("memo")+chalk.gray(j(" 记问题 · "," report · "))+chalk.cyan("back")+chalk.gray(j(" 退出"," leave"))),console.log(chalk.gray(j("  题目 id 在卡片的 ▸ 那一行,例如 ","  The task id is on the card’s ▸ line, e.g. "))+chalk.cyan("open 17")+chalk.gray(" → ")+chalk.cyan("kaggle s17-bazarul-lui-riki")));c.prompt=$,c.setPrompt(S),c.prompt()}),s||c.once("close",()=>o()),c.prompt=$,c.prompt()})}export function registerSprintCommand(o){o.command("pandahelpme31",{hidden:!0}).action(async()=>{await runSprint()})}
+import{existsSync as o,mkdirSync as e,readFileSync as n,writeFileSync as t}from"node:fs";import{createInterface as a}from"node:readline";import{dirname as l,join as r}from"node:path";import{fileURLToPath as s}from"node:url";import chalk from"chalk";import{getConfig as c}from"../lib/config.js";import{getMainRl as i,setMainRl as g}from"../lib/main-rl.js";import{logCommand as y}from"../lib/logger.js";import{baselineChip as d,computeChip as p,computeLegend as h,loadSprintTasks as u,mediaHelpLines as m,taskSteps as f}from"../lib/sprint-meta.js";let b=!1;export function isSprintActive(){return b}const w=l(s(import.meta.url)),k=r(process.env.HOME||".",".icoa"),$=r(k,"sprint-state.json"),A={"T-torch":"裸 torch:手写 attention / causal mask / perplexity","T-umap":"降维当特征提取器,不是画图工具","T-lora":"peft LoRA 参数高效微调","T-librosa":"音频特征流水线:mel / MFCC → 定长向量"},S=["",chalk.bold.cyan("   ASTANA 2027"),chalk.gray("   — 给还没出发的你"),"",chalk.white("   草原的风从额尔齐斯河吹到阿斯塔纳的时候,会带着一点雪的味道。"),chalk.white("   你还没去过那里。但那里已经有一个座位,写着你的名字。"),"",chalk.white("   你现在坐在屏幕前,离比赛还有几天。"),chalk.white("   你可能会想:几天能改变什么?"),"",chalk.white("   足够让一个人从「看着题发呆」变成「看到题就知道该拆哪一层」。"),chalk.white("   足够让手指记住 import torch 之后该敲什么。"),chalk.white("   足够让你在比赛那天早上醒来时,不再害怕。"),"",chalk.bold.yellow("   M")+chalk.white(" —— 你总是第一个说「我先试试」的那个。别停。"),chalk.bold.yellow("   V")+chalk.white(" —— 你总是问「为什么是这样」的那个。继续问。"),chalk.bold.yellow("   A")+chalk.white(" —— 你总是在别人放弃时还在改最后一版的那个。改下去。"),"",chalk.white("   你们三个人,一个人是一条腿走路,三个人是奔跑。"),"",chalk.white("   你们要面对的不是一场考试,是一次证明:"),chalk.bold.white("   中国的高中生,可以在 AI 的赛场上,和世界站在同一排。"),"",chalk.white("   而 2027 年的阿斯塔纳,是我们下一个约定。"),"",chalk.gray("   现在,把这行字关掉,去做第 1 题。"),chalk.gray("   — 熊猫教练"),""];function v(o){try{e(k,{recursive:!0}),t($,`${JSON.stringify(o,null,1)}\n`)}catch{}}function I(){return(c().language||"en").toLowerCase().startsWith("zh")}function j(o,e){return I()?o:e}function q(o){return"★".repeat(Math.max(1,Math.min(5,o||3)))}const C=r(k,"sprint-memos.json");function x(o){try{e(k,{recursive:!0}),t(C,`${JSON.stringify(o,null,1)}\n`)}catch{}}function N(o,e){const n=new Set(e.done),t=new Set(e.opened);console.log(),console.log(chalk.bold.cyan(j("  ⚔ IOAI 2026 · 四工具冲刺","  ⚔ IOAI 2026 · four-tool sprint"))+chalk.gray(`   ${j("已攻克","cleared")} ${n.size} / ${o.length}`));const a=new Map;for(const e of o)a.has(e.bullseye)||a.set(e.bullseye,[]),a.get(e.bullseye)?.push(e);for(const[o,e]of a){console.log(),console.log(chalk.bold.white(`  ${o}`)+chalk.gray(`  — ${A[o]||""}`));for(const o of e){const e=n.has(o.seq)?chalk.green("✓"):t.has(o.seq)?chalk.yellow("◐"):chalk.gray("⬜");console.log(`    ${e} ${chalk.bold.white(String(o.seq).padStart(2))}  `+chalk.white(T(o.title,34))+chalk.gray(q(o.difficulty).padEnd(5))+chalk.gray(o.needs_gpu?" ⚡":"  ")+chalk.gray(` ${o.data_mb}M`)+("teleload"===o.status?chalk.gray(j("  · teleload","  · teleload")):"")+(o.judged?chalk.green(j("  · 可出分","  · scored")):""))}}console.log(),console.log(chalk.gray("  ")+chalk.cyan("open <n>")+chalk.gray(j(" 看题  ·  "," view  ·  "))+chalk.cyan("done <n>")+chalk.gray(j(" 标记攻克  ·  "," mark cleared  ·  "))+chalk.cyan("ai4ioai")+chalk.gray(j(" 开 AI 分屏  ·  "," AI in a split  ·  "))+chalk.cyan("memo")+chalk.gray(j(" 记问题  ·  "," report a problem  ·  "))+chalk.cyan("back")+chalk.gray(j(" 退出"," leave"))),console.log(chalk.gray(j("  ⬜ 没看过   ◐ 看过没做完   ✓ 攻克","  ⬜ unseen   ◐ opened   ✓ cleared"))),console.log(chalk.gray(h(I()))),console.log()}function O(o){if(/[　-鿿＀-￯]/.test(o))return 2;const e=o.codePointAt(0)??0;return e>=126976&&e<=129791||e>=9728&&e<=10175?2:1}function T(o,e){const n=[...o].reduce((o,e)=>o+O(e),0);if(n<=e)return o+" ".repeat(e-n);let t="",a=0;for(const n of o){const o=O(n);if(a+o>e-1)break;t+=n,a+=o}return`${t}…${" ".repeat(Math.max(0,e-a-1))}`}function L(o,e){const n=[];let t="",a=0;for(const l of o){const o=/[　-鿿＀-￯]/.test(l)?2:1;a+o>e&&(n.push(t),t="",a=0),t+=l,a+=o}return t&&n.push(t),n}export async function runSprint(){y("sprint");const e=function(){try{const e=u();try{const t=JSON.parse(n(function(e){const n=[r(w,"..","data",e),r(w,"..","..","src","data",e)];for(const e of n)if(o(e))return e;return n[0]}("sprint-cards.json"),"utf8")),a=new Map((t.cards||[]).map(o=>[o.seq,o]));for(const o of e){const e=a.get(o.seq);e&&Object.assign(o,{star:e.star,trap:e.trap,action:e.action,note:e.note,hint:e.hint,meta:e.meta})}}catch{}return e.sort((o,e)=>o.seq-e.seq)}catch{return[]}}();if(0===e.length)return console.log(),console.log(chalk.yellow("  冲刺题单还没装进这个版本。")),console.log(chalk.gray("  升级后再试:")+chalk.cyan("npm install -g icoa-cli@latest")),void console.log();const t=function(){try{const o=JSON.parse(n($,"utf8"));return{who:o.who||"",opened:o.opened||[],done:o.done||[]}}catch{return{who:"",opened:[],done:[]}}}();if(0===t.opened.length&&0===t.done.length)for(const o of S)console.log(o);t.who||(t.who=""),N(e,t);const l=i(),s=null!==l,c=s?l:a({input:process.stdin,output:process.stdout,terminal:!0});s||g(c);const h=s?c.listeners("line").slice():[];c.removeAllListeners("line");const k=c.prompt.bind(c);b=!0;const A=chalk.bold.cyan("icoa ioai2026> ");c.setPrompt(A);let O=null;await new Promise(o=>{c.on("line",async a=>{const l=a.trim().replace(/^icoa\s+/i,""),[r,i]=[l.split(/\s+/)[0]?.toLowerCase()||"",l.split(/\s+/)[1]||""];if(["back","quit","exit","menu"].includes(r))return v(t),console.log(chalk.gray("  进度已保存。")),(()=>{if(b=!1,c.removeAllListeners("line"),s){for(const o of h)c.on("line",o);c.setPrompt(chalk.green("icoa> ")),c.prompt()}else g(null),c.close()})(),void o();if(""===r||["challenges","list","ls"].includes(r))N(e,t);else if(["open","o"].includes(r)||/^\d+$/.test(r)){const o=Number.parseInt(/^\d+$/.test(r)?r:i,10),n=e.find(e=>e.seq===o);n?(t.opened.includes(o)||t.opened.push(o),v(t),O=o,function(o){console.log(),console.log(chalk.cyan("  ╭─ ")+chalk.bold.white(`${o.seq} · ${o.title}`)+chalk.gray(`  ·  ${o.bullseye}`)+chalk.cyan(" ─────")),console.log(chalk.cyan("  │"));const e=f(o).length>1,n=[o.star,o.trap,e?"":o.action,o.note,o.hint?`  ${o.hint}`:""].filter(Boolean);0===n.length&&o.topic_cn&&n.push(`▸ ${o.topic_cn}`);for(const o of n){for(const e of L(o,58))console.log(chalk.cyan("  │ ")+chalk.white(e));console.log(chalk.cyan("  │"))}const t=I();console.log(chalk.cyan("  │ ")+chalk.gray(p(o,t)));const a=d(o,t);if(a)for(const o of L(a,58))console.log(chalk.cyan("  │ ")+chalk.gray(o));console.log(chalk.cyan("  │ ")+chalk.gray(o.meta||`⏱ 难度 ${q(o.difficulty)}`)+chalk.gray(o.needs_gpu?"  ⚡ 需要 GPU":"")),console.log(chalk.cyan("  ╰──────────────────────────────────────────────────────"));for(const e of m(o,t))for(const o of L(e,60))console.log(chalk.yellow("  ")+chalk.yellow(o));const l=f(o),r=function(o){const e=f(o);if(e.length>0)return e[0];const n=/(?:kaggle|teleload)\s+[a-z0-9][a-z0-9-]*/i.exec(o.action||"");return n?n[0]:null}(o);l.length>1?(console.log(chalk.bold.white(j("  ▸ 就在这里敲,两步:","  ▸ Type these here, in order:"))),l.forEach((o,e)=>{console.log(chalk.gray(`    ${e+1}. `)+chalk.bold.cyan(o)+chalk.gray(0===e?j("   ← 先把数据拉下来","   ← fetch the data first"):j("   ← 题面 + notebook","   ← statement + notebook")))})):r&&console.log(chalk.bold.white(j("  ▸ 就在这里敲: ","  ▸ Type it right here: "))+chalk.bold.cyan(r)+chalk.gray(j("   ← 数据 + notebook 落到本机","   ← pulls the data + notebook"))),console.log(chalk.gray("  ")+chalk.cyan("ai4ioai")+chalk.gray(j(" 开 AI 分屏  ·  "," AI in a split  ·  "))+chalk.cyan(`done ${o.seq}`)+chalk.gray(j(" 标记攻克  ·  "," mark cleared  ·  "))+chalk.cyan("memo")+chalk.gray(j(` 记问题(自动记在第 ${o.seq} 题)  ·  `,` report (auto-filed under ${o.seq})  ·  `))+chalk.cyan("back")+chalk.gray(j(" 回列表"," back to list"))),console.log()}(n)):console.log(chalk.yellow(`  没有第 ${i||r} 题。`)+chalk.gray(` 题号 1-${e.length}。`))}else if("done"===r){const o=Number.parseInt(i,10);e.some(e=>e.seq===o)?(t.done.includes(o)||t.done.push(o),v(t),console.log(chalk.green(`  ✓ 第 ${o} 题标记为攻克`)+chalk.gray(`  (${t.done.length}/${e.length})`))):console.log(chalk.yellow(`  题号 1-${e.length}。`))}else["kaggle","teleload","ipynb","board"].includes(r)?await async function(o,e){const n=e.split(/\s+/).filter(Boolean);try{if("teleload"===o){const{runTeleload:o}=await import("./teleload.js");return void await o(n[0])}if("ipynb"===o){const{runIpynb:o}=await import("./ipynb.js");return void await o(n[0])}if("board"===o){const{runBoard:o}=await import("./board.js");return void await o(n[0],n[1])}const e=await import("./kaggle.js"),t=(n[0]||"").toLowerCase();"submit"===t?await e.runKaggleSubmit(n[1]):"check"===t?await e.runKaggleCheck(n[1]):"guide"===t?await e.runKaggleGuide(n[1]):await e.runKaggle(n[0])}catch(o){console.log(),console.log(chalk.yellow(j("  这条命令没跑成:","  That command failed: "))+chalk.gray(String(o))),console.log(chalk.gray(j("  记一条给教练:","  Tell the coach: "))+chalk.cyan(j("memo …","memo …"))),console.log()}}(r,l.split(/\s+/).slice(1).join(" ")):"memo"===r||"note"===r?await async function(o,e,t,a){const l=function(){try{const o=JSON.parse(n(C,"utf8"));return Array.isArray(o)?o:[]}catch{return[]}}(),r=o.trim();if(["list","ls","all"].includes(r.toLowerCase())){if(console.log(),0===l.length)console.log(chalk.gray(j("  还没有记录。","  Nothing noted yet.")));else{console.log(chalk.bold.white(j(`  📝 已记 ${l.length} 条`,`  📝 ${l.length} noted`)));for(const o of l.slice(-12)){const e=o.seq?chalk.cyan(`#${o.seq}`):chalk.gray("—");console.log(`   ${e} ${chalk.white(o.text.slice(0,60))}`+chalk.gray(o.synced?j("  ·已上报","  ·sent"):j("  ·仅本地","  ·local only")))}console.log(chalk.gray(`  ${C}`))}return console.log(chalk.gray(j("  再记一条:","  Note another: "))+chalk.cyan("memo")+chalk.gray(j("  (光敲这一个词就给你写字的地方)","  (just the word — it gives you a place to type)"))),void console.log()}const s=r?r.split(/\s+/):[],c=s.length>0&&/^\d+$/.test(s[0])&&e.some(o=>o.seq===Number(s[0])),i=c?Number(s[0]):t;let g=(c?s.slice(1):s).join(" ").trim();if(""===g){const o=i?chalk.gray(j(` —— 记在第 ${i} 题下`,` — filed under task ${i}`)):chalk.gray(j(" —— 通用反馈"," — general feedback"));if(console.log(),console.log(chalk.bold.white(j("  📝 撞到什么就写什么","  📝 Say what you hit"))+o),console.log(chalk.gray(j("  一行写完,回车就存下(不用等回复);直接回车 = 不记。","  One line, Enter files it (no reply comes back). Empty Enter cancels."))),g=await function(o,e){return new Promise(n=>{o.question(e,o=>n(o.trim()))})}(a,chalk.cyan("  memo > ")),""===g)return console.log(chalk.gray(j("  没记。","  Nothing noted."))),void console.log()}const d={ts:(new Date).toISOString(),seq:i,text:g,synced:!1};l.push(d),x(l),y(`sprint: memo${i?` ${i}`:""} — ${g}`);let p=!1;try{const{loadLearnState:o}=await import("../lib/learn-state.js"),e=o()?.token??"";if(e){const{syncCardFeedback:o}=await import("../lib/learn-curricula.js");p=await o(e,{curriculum_id:"pandahelpme31",card_number:i??0,feedback_text:g})}}catch{}p&&(d.synced=!0,x(l)),console.log(),console.log(chalk.green(j("  ✓ 记下了","  ✓ noted"))+(i?chalk.gray(j(` (第 ${i} 题)`,` (task ${i})`)):"")+chalk.gray(p?j("  ·已上报","  ·sent"):j("  ·存在本机,教练可以来取","  ·saved on this machine"))),console.log(chalk.gray(j("  继续做题就好。看全部:","  Carry on. See them all: "))+chalk.cyan("memo list")),console.log()}(l.split(/\s+/).slice(1).join(" "),e,O,c):["ai4ioai","/ai4ioai","ai"].includes(r)?(console.log(),console.log(chalk.bold.white(j("  左右分屏,左边看题右边问 AI:","  Split the screen — tasks left, AI right:"))),console.log(chalk.gray("   1. ")+chalk.bold.cyan("Ctrl-b %")+chalk.gray(j("   左右分屏(上下分屏是 ","   split left/right (up/down is "))+chalk.cyan('Ctrl-b "')+chalk.gray(j(")",")"))),console.log(chalk.gray("   2. ")+chalk.gray(j("右边那格输 ","in the right pane type "))+chalk.bold.cyan("icoa")+chalk.gray(j(" → ↑↓ 选 "," → ↑↓ pick "))+chalk.bold.white("🧠 Learn")+chalk.gray(j(" → 在 "," → at "))+chalk.cyan("icoa learn>")+chalk.gray(j(" 输 "," type "))+chalk.bold.cyan("ai4ioai")),console.log(chalk.gray(j("   3. 两格之间切换:","   3. move between panes: "))+chalk.cyan("Ctrl-b ←")+chalk.gray(" / ")+chalk.cyan("Ctrl-b →")),console.log(),console.log(chalk.gray(j("  它是弱模型 —— 单轮、2000 字符以内。把问题问准,就是这次要练的本事。","  A deliberately weak model — single-turn, 2000 chars. Asking precisely is the skill."))),console.log(chalk.gray(j("  在 notebook 里卡住时,那边可以直接敲 ","  Stuck inside a notebook? there you can type "))+chalk.cyan("ai")+chalk.gray(j("  —— 它会把那一格的代码和报错一起带上。","  — it carries that cell and its traceback."))),console.log()):["help","?","guide"].includes(r)?function(o,e){const n=I();console.log(),console.log(chalk.bold.white(n?"  ═══ 📖 冲刺指南 ═══":"  ═══ 📖 Sprint guide ═══")),console.log(chalk.gray(n?`  这里是 IOAI 2026 四工具冲刺:${o.length} 张卡,每张 = 一个要点 + 一道真题。已攻克 ${e.done.length}/${o.length}。`:`  The IOAI 2026 four-tool sprint: ${o.length} cards, each = one idea + one real task. Cleared ${e.done.length}/${o.length}.`)),console.log(),console.log(chalk.bold.white(n?"  一张卡的循环:":"  The loop for one card:")),[["open <n>",n?"看题卡(要点/陷阱/下一步)":"read the card (idea / trap / next step)",""],["teleload / kaggle …",n?"照卡片 ▸ 行领数据":"fetch the data — copy the card’s ▸ line",""],["ipynb <路径>",n?"开 notebook,先 run all":"open the notebook, run all first",""],["kaggle check / submit",n?"验格式 → 判分出分(可出分的题)":"check the format → score it (judged tasks)",""],["board <event> <comp>",n?"看排行":"see the standings",""],["done <n>",n?"标记攻克,回列表挑下一张":"mark it cleared, pick the next card",""]].forEach(([o,e],n)=>{console.log(chalk.gray(`   ${n+1}. `)+chalk.cyan(o.padEnd(22))+chalk.gray(e))}),console.log(),console.log(chalk.gray(n?"  卡住了: ":"  Stuck: ")+chalk.cyan("ai4ioai")+chalk.gray(n?" 开 AI 分屏 · ":" AI in a split · ")+chalk.cyan("memo")+chalk.gray(n?" 把问题记给教练 · ":" report to the coach · ")+chalk.cyan("challenges")+chalk.gray(n?" 回列表 · ":" the list · ")+chalk.cyan("back")+chalk.gray(n?" 退出(进度自动存)":" leave (progress saved)")),console.log()}(e,t):(console.log(chalk.gray("  ")+chalk.cyan("open <n>")+chalk.gray(j(" 看题 · "," view · "))+chalk.cyan("kaggle <题目id>")+chalk.gray(j(" 领数据 · "," get the data · "))+chalk.cyan("ipynb <路径>")+chalk.gray(j(" 开 notebook · "," open the notebook · "))+chalk.cyan("done <n>")+chalk.gray(j(" 标记 · "," mark · "))+chalk.cyan("memo")+chalk.gray(j(" 记问题 · "," report · "))+chalk.cyan("back")+chalk.gray(j(" 退出"," leave"))),console.log(chalk.gray(j("  题目 id 在卡片的 ▸ 那一行,例如 ","  The task id is on the card’s ▸ line, e.g. "))+chalk.cyan("open 17")+chalk.gray(" → ")+chalk.cyan("kaggle s17-bazarul-lui-riki")));c.prompt=k,c.setPrompt(A),c.prompt()}),s||c.once("close",()=>o()),c.prompt=k,c.prompt()})}export function registerSprintCommand(o){o.command("pandahelpme31",{hidden:!0}).action(async()=>{await runSprint()})}

package/dist/lib/kaggle.js

@@ -1 +1 @@
-import{sprintTaskForComp as e}from"./sprint-meta.js";export function parseKaggleTaskList(e){const t=e??{},s=(t.data??t).tasks;if(!Array.isArray(s))return[];const a=[];for(const e of s){if(!e||"object"!=typeof e)continue;const t=e;"string"==typeof t.comp_id&&""!==t.comp_id&&a.push({comp_id:t.comp_id,title:"string"==typeof t.title?t.title:t.comp_id,metric:"string"==typeof t.metric?t.metric:"accuracy",description:"string"==typeof t.description&&""!==t.description?t.description:void 0,kaggle_url:"string"==typeof t.kaggle_url?t.kaggle_url:null,licence:"string"==typeof t.licence?t.licence:void 0,attribution:"string"==typeof t.attribution?t.attribution:void 0,data_base:"string"==typeof t.data_base&&""!==t.data_base?t.data_base:null,id_col:"string"==typeof t.id_col?t.id_col:void 0,target_col:"string"==typeof t.target_col?t.target_col:void 0,n_train:"number"==typeof t.n_train?t.n_train:void 0,n_test:"number"==typeof t.n_test?t.n_test:void 0,judge:"string"==typeof t.judge&&""!==t.judge?t.judge:void 0,board_event:"string"==typeof t.board_event&&""!==t.board_event?t.board_event:null,needs_gpu:!0===t.needs_gpu})}return a}export function dataFileUrls(e){const t=e.data_base;if("string"!=typeof t||""===t)return null;const s=t.replace(/\/$/,"");return{train:`${s}/train.csv`,test:`${s}/test.csv`,sample_submission:`${s}/sample_submission.csv`}}export function buildStarterCells(t,s){const a=s.id_col??"id",i=s.target_col??"target";var n;n=s.metric,/rmse|mse|mae|error/i.test(n);const r="icoa-holdout"===s.judge,o=e=>({kind:"markdown",source:e,result:null}),l=e=>({kind:"code",source:e.join("\n"),result:null}),c=[`# ${s.title}`,"",...s.description?[s.description,""]:[],`Kaggle-style practice — metric: **${s.metric}**.`,"","- `train.csv` — labelled training data",`- \`test.csv\` — rows to predict (no \`${i}\` column)`,"- `sample_submission.csv` — the exact format Kaggle expects","","Run all cells top to bottom: a baseline model trains on your CPU and","`submission.csv` appears next to the data. Then improve the model cell",...r?["and re-run — `icoa kaggle submit` scores it and ranks you on the board."]:["and re-run — your score is whatever Kaggle says it is."],"","**AI help:** split your terminal (`Ctrl-b %`), run `icoa` in the right pane","and type `ai4ioai` for a live Gemma 3n E4B. Ask it for feature ideas or bug fixes —","then verify every suggestion with the val score. A tip that lowers it, drop it.",...s.licence?["",`**Data licence:** ${s.licence}`]:[],...s.attribution?[`**Attribution:** ${s.attribution}`]:[]].join("\n"),p=l(["# setup — load the data (leave as is)","import os","import numpy as np","import pandas as pd","from pandas.api.types import is_numeric_dtype","from sklearn.feature_extraction.text import TfidfVectorizer","from sklearn.model_selection import train_test_split","from sklearn.metrics import accuracy_score, mean_squared_error","","# same code everywhere: in a GPU session these env vars point at the","# box's own dataset copy + workspace — locally they are unset.",`DIR = os.environ.get("ICOA_KAGGLE_DATA") or r"${t}"`,'OUT = os.environ.get("ICOA_KAGGLE_OUT") or DIR',...s.data_base?[`DATA_BASE = "${s.data_base.replace(/\/$/,"")}"`,'if not os.path.exists(DIR + "/train.csv"):',"    import urllib.request","    os.makedirs(DIR, exist_ok=True)",'    for f in ("train.csv", "test.csv", "sample_submission.csv"):','        urllib.request.urlretrieve(DATA_BASE + "/" + f, DIR + "/" + f)']:[],'train = pd.read_csv(DIR + "/train.csv")','test = pd.read_csv(DIR + "/test.csv")','sub_fmt = pd.read_csv(DIR + "/sample_submission.csv")',"","# 🔴 The submission column names are NOT always the training column names.","# On tasks imported from other platforms the two genuinely differ: train.csv",'# labels the answer "class" or "result" while sample_submission.csv calls it','# "answer". Hard-coding the submission name here used to raise KeyError on the',"# very first cell. So trust the DATA: whatever column train has and test lacks","# is the thing to predict.",`SUB_ID, SUB_TARGET = "${a}", "${i}"   # names the submission file wants`,"only_train = [c for c in train.columns if c not in test.columns]","TARGET = only_train[-1] if only_train else SUB_TARGET","shared = [c for c in train.columns if c in test.columns]","# Only treat a column as the row id when the submission actually names it AND",'# both files carry it. Falling back to "first shared column" silently spent a',"# real feature as an id on tasks whose id lives only in test.csv.","ID = SUB_ID if SUB_ID in shared else None",'assert TARGET in train.columns, f"cannot find the label column in {list(train.columns)}"',"","# Regression or classification? Decided from the labels, not from metadata —","# several tasks ship continuous labels, and handing those to a classifier",'# raises "Unknown label type: continuous".',"REG = is_numeric_dtype(train[TARGET]) and train[TARGET].nunique() > 20",'print(f"train {train.shape} · test {test.shape} · predicting {TARGET!r} "',"      f\"({'regression' if REG else 'classification'})\")","# 前几行 —— 看一眼真实数据长什么样,比看列名有用得多","print(train.head(3).to_string(max_colwidth=24)[:700])"]),d=e(s.comp_id),u=d?.media,m="audio"===u?.kind?"scipy.io.wavfile.read(path) -> numpy, then scipy.signal for a spectrogram":"image"===u?.kind?"PIL.Image.open(path) -> numpy":"weights"===u?.kind?"torch.load(path, weights_only=True) — the model itself is the input here":"unpack the archive inside the pack, then read what is in it",h=l(["# FEATURES — turn raw columns into numbers a model can use (improve THIS cell).","# Numbers: fill gaps with the median. Category (<= 20 distinct): one-hot.","# Free text: TF-IDF — on a text task the sentences ARE the signal, and dropping",'# them as "high-cardinality" leaves nothing to train on (that used to crash the','# whole notebook with "No objects to concatenate").',"# only columns present in BOTH files — a train-only column cannot be a feature","# at prediction time, and indexing test with it raises KeyError.","feat = [c for c in shared if c not in (ID, TARGET)]","Xtr_raw, Xte_raw = train[feat].copy(), test[feat].copy()","num = [c for c in feat if is_numeric_dtype(Xtr_raw[c])]","cat = [c for c in feat if c not in num and Xtr_raw[c].nunique() <= 20]","# Everything else is vectorised as text. Requiring a long average length here","# used to throw away high-cardinality SHORT strings — ingredient names, single","# words, category codes — which on several tasks are the entire signal, leaving","# the notebook with zero features.","txt = [c for c in feat if c not in num and c not in cat]","for c in num:","    med = Xtr_raw[c].median()","    Xtr_raw[c] = Xtr_raw[c].fillna(med)","    Xte_raw[c] = Xte_raw[c].fillna(med)","blocks_tr, blocks_te = [], []","if num or cat:","    d_tr = pd.get_dummies(Xtr_raw[num + cat], columns=cat, dummy_na=True)","    d_te = pd.get_dummies(Xte_raw[num + cat], columns=cat, dummy_na=True)",'    d_tr, d_te = d_tr.align(d_te, join="left", axis=1, fill_value=0)',"    blocks_tr.append(d_tr.to_numpy(dtype=float))","    blocks_te.append(d_te.to_numpy(dtype=float))","for c in txt:","    # fillna BEFORE astype(str): a real NaN reaching the vectoriser raises",'    # "np.nan is an invalid document", and astype alone does not catch it.','    tr_txt = Xtr_raw[c].fillna("").astype(str)','    te_txt = Xte_raw[c].fillna("").astype(str)',"    # min_df=1: on a short-token column (one ingredient per row) min_df=2 can","    # filter the vocabulary down to nothing.","    long_col = tr_txt.str.len().mean() >= 12","    # Try the good settings, then a permissive fallback, then give up on THIS","    # column only. A column whose vocabulary prunes to nothing used to abort the",'    # whole run with "After pruning, no terms remain".','    for kw in ({"min_df": 2 if long_col else 1,','                "analyzer": "word" if long_col else "char_wb"},','               {"min_df": 1, "analyzer": "char_wb"}):',"        try:","            vec = TfidfVectorizer(max_features=3000, ngram_range=(1, 2), **kw)","            b_tr = vec.fit_transform(tr_txt).toarray()","            blocks_tr.append(b_tr)","            blocks_te.append(vec.transform(te_txt).toarray())","            break","        except ValueError as e:","            last = e","    else:",'        print(f"  skipped text column {c!r}: {last}")',"if not blocks_tr:","    raise SystemExit(",'        f"No usable feature columns. The shared columns are {feat} — none of them "','        f"is numeric, categorical or text.\\n"',...u?[`        "The signal for this task is in ${u.kind} files, not in the CSVs.\\n"`,`        "  1. fetch them (${u.mb}MB):  teleload ${u.pack}\\n"`,...u.glob?[`        "     files land at <pack dir>/${u.glob}\\n"`]:[],`        "  2. replace this cell: ${m}\\n"`,'        "     build one feature row per file, then keep the rest of the notebook.\\n"']:['        f"Two kinds of task land here, and both are real work rather than a bug:\\n"','        f"  1. audio/image — train.csv only lists sample ids and labels and the "','        f"signal is in media files. Read them with scipy.io.wavfile (audio) or "','        f"PIL (images) and turn each file into a feature row.\\n"','        f"  2. train.csv and test.csv describe different things (a retrieval or "','        f"matching task). Read both files, work out what links a test row to a "','        f"train row, and build that mapping yourself.\\n"'],...u?['        f"columns — train={list(train.columns)} test={list(test.columns)}"']:['        f"Either way: replace this cell. train={list(train.columns)} "','        f"test={list(test.columns)}"'],"    )","X = np.hstack(blocks_tr)","X_test = np.hstack(blocks_te)","y = train[TARGET]","X_tr, X_val, y_tr, y_val = train_test_split(X, y, test_size=0.2, random_state=0)",'print(f"{X.shape[1]} model-ready features "','      f"({len(num)} numeric, {len(cat)} one-hot, {len(txt)} TF-IDF text)")']),f=l(["# EDA — look at the labels before you model. Produces ONE figure:","#   left  = what you are predicting (class balance / value spread)","#   right = what the FEATURES cell will have to work with","# icoa ipynb saves it as a .png and opens your image viewer (type `open` to reopen).","import matplotlib.pyplot as plt","try:","    fig, ax = plt.subplots(1, 2, figsize=(11, 4))","    if REG:",'        vals = pd.to_numeric(train[TARGET], errors="coerce").dropna()',"        ax[0].hist(vals, bins=30)",'        ax[0].set_title(f"{TARGET} — {len(vals)} values, mean {vals.mean():.3g}")',"    else:","        vc = train[TARGET].astype(str).value_counts().head(12)","        ax[0].bar(range(len(vc)), vc.values)","        ax[0].set_xticks(range(len(vc)))",'        ax[0].set_xticklabels([s[:14] for s in vc.index], rotation=45, ha="right")','        # An imbalanced target is the single most common reason a "good"',"        # accuracy means nothing — show the majority share right in the title.","        top = vc.iloc[0] / len(train)",'        ax[0].set_title(f"{TARGET} — {train[TARGET].nunique()} classes · biggest {top:.0%}")','    kinds = {"numeric": 0, "category": 0, "text": 0}',"    for c in [c for c in shared if c not in (ID, TARGET)]:","        if is_numeric_dtype(train[c]):",'            kinds["numeric"] += 1',"        elif train[c].nunique() <= 20:",'            kinds["category"] += 1',"        else:",'            kinds["text"] += 1',"    miss = train.isna().mean()","    miss = miss[miss > 0].sort_values(ascending=False).head(10)","    if len(miss) > 0:","        ax[1].barh(range(len(miss)), miss.values)","        ax[1].set_yticks(range(len(miss)))","        ax[1].set_yticklabels([s[:16] for s in miss.index])",'        ax[1].set_title("missing rate (top columns)")',"    else:","        ax[1].bar(list(kinds), list(kinds.values()))",'        ax[1].set_title(f"{sum(kinds.values())} usable columns · no missing values")',"    plt.tight_layout()","    plt.show()",'    print("columns by kind:", kinds)',"except Exception as e:","    # EDA is never allowed to stop the run — `run all` halts on the first","    # error, and a plotting hiccup must not cost the student their baseline.",'    print("EDA skipped:", type(e).__name__, e)']),g=l(["# MODEL — a working baseline. Improve THIS cell to climb the board.","from sklearn.linear_model import LogisticRegression, Ridge","# REG was decided from the labels back in setup.","model = Ridge() if REG else LogisticRegression(max_iter=2000)"]),_=l(["# validate — quick local score on a held-out fifth of train","model.fit(X_tr, y_tr)","p_val = model.predict(X_val)","if REG:",'    print("val RMSE:", round(mean_squared_error(y_val, p_val) ** 0.5, 4))',"else:","    acc = accuracy_score(y_val, p_val)",'    print("val accuracy:", round(acc, 4))',"    if acc < 0.3:","        # A near-zero baseline is the expected result on several of these tasks,","        # not a broken notebook — say so, or the student reads it as a bug.",'        print("    A score this low usually means the linear baseline cannot "','              "represent the pattern at all (combination rules, sequences, "','              "spelling). That IS the task: build features that expose it, "','              "or switch to a model that can. The chain itself works — "','              "submission.csv is written either way.")']),b=l(["# submission — retrain on ALL data, write submission.csv (Kaggle format)","model.fit(X, y)","pred = model.predict(X_test)",'out = OUT + "/submission.csv"',"# Copy the layout of sample_submission.csv and swap in the predictions, rather","# than rebuilding the file from column names. Some tasks want more than two","# columns (e.g. subtaskID,datapointID,answer) and a hand-built 2-column frame","# is rejected by the grader even when every prediction is right.","if len(sub_fmt) == len(pred):","    sub = sub_fmt.copy()","    sub[sub.columns[-1]] = pred","else:","    keys = {}","    if ID is not None and ID in test.columns:","        keys[SUB_ID] = test[ID]","    keys[SUB_TARGET] = pred","    sub = pd.DataFrame(keys)",'    print(f"NOTE: sample_submission has {len(sub_fmt)} rows but test has "','          f"{len(pred)} — built the file from column names instead; "','          f"check it against the task page before submitting.")',"sub.to_csv(out, index=False)",'print("wrote", out, "·", list(sub.columns))',r?`print("score it: icoa kaggle submit ${s.comp_id}")`:'print("format check: icoa kaggle check")']),y=r?["## Submit to ICOA (scored on the server)","",`1. In the ICOA CLI run \`icoa kaggle submit ${s.comp_id}\``,"2. The server scores `submission.csv` against the secret test labels","   and your rank lands on the live board (`icoa board`).","3. Improve the MODEL cell, re-run, submit again — best score counts."]:"nitro"===s.judge?["## Where this one gets scored","","This task comes from **judge.nitro-ai.org** and the answer key stays on","their server, so ICOA cannot score it for you. That does not make it a","practice stub — it is a real task, just judged elsewhere.","",`1. \`icoa kaggle check ${s.comp_id}\` — checks the FORMAT of`,"   `submission.csv` (headers, row count, ids). Format errors are the most","   common way a correct answer still scores zero.","2. Watch the **val score** printed above — that is your real feedback loop","   during the sprint, and it needs no account anywhere.","3. Want the official number? Register at judge.nitro-ai.org and submit","   there. Otherwise compare against the official solution, or ask","   `ai4ioai` to review your approach."]:s.kaggle_url?["## Submit on Kaggle (manual, your own account)","",`1. Open ${s.kaggle_url}`,"2. **Submit Predictions** → upload `submission.csv`","3. Read your leaderboard score, come back, improve the MODEL cell, repeat."]:["## Submit on Kaggle (manual, your own account)","","_This dataset has no live Kaggle page linked yet._","The loop is identical: upload `submission.csv`, read the score, improve","the MODEL cell, repeat. Meanwhile the **val score** printed above is a","real signal you can work against with no account at all."];return[o(c),p,f,h,g,_,b,o(y.join("\n"))]}export function parseSubmitResponse(e){const t=e&&"object"==typeof e?e:{},s=t.data??t,a=s&&"object"==typeof s?s:{};return"number"==typeof a.score&&"string"==typeof a.metric?{ok:!0,score:a.score,metric:a.metric,rank:"number"==typeof a.rank_at_time?a.rank_at_time:null,fieldSize:"number"==typeof a.field_size?a.field_size:null,deduped:!0===a.deduped,boardEvent:"string"==typeof a.board_event?a.board_event:null}:"string"==typeof a.reason&&""!==a.reason?{ok:!1,reason:a.reason,error:"string"==typeof a.error?a.error:void 0,problems:Array.isArray(a.problems)?a.problems.filter(e=>"string"==typeof e):void 0}:{ok:!1,reason:"network"}}export function validateSubmissionText(e,t){const s=[],a=e.trim().split(/\r?\n/),i=t.trim().split(/\r?\n/),n=a[0]??"",r=i[0]??"";n.trim()!==r.trim()&&s.push(`header mismatch: expected "${r}", got "${n}"`);const o=a.slice(1).filter(e=>""!==e.trim()),l=i.slice(1).filter(e=>""!==e.trim());o.length!==l.length&&s.push(`row count mismatch: expected ${l.length}, got ${o.length}`);const c=e=>new Set(e.map(e=>(e.split(",")[0]??"").trim())),p=c(o),d=c(l),u=[...d].filter(e=>!p.has(e)),m=[...p].filter(e=>!d.has(e));if(u.length>0||m.length>0){const e=[];u.length>0&&e.push(`${u.length} missing (e.g. ${u[0]})`),m.length>0&&e.push(`${m.length} unknown (e.g. ${m[0]})`),s.push(`id mismatch vs sample_submission: ${e.join(", ")}`)}for(let e=0;e<o.length;e++)if(o[e].split(",").some(e=>""===e.trim())){s.push(`empty cell in data row ${e+1}`);break}return{ok:0===s.length,problems:s}}export function buildMediaStarterCells(e,t){const s=e=>({kind:"markdown",source:e,result:null}),a=e=>({kind:"code",source:e.join("\n"),result:null}),i=t.media?.kind??"archive",n=t.title??"ICOA sprint task",r=[`# ${t.seq?`${t.seq} · `:""}${n}`,"",...t.topic_cn?[t.topic_cn,""]:[],"This task ships **data, not a baseline** — the CSV starter cannot help here",`because the signal lives in the ${i} files you just pulled.`,"","Run all cells: they inventory the pack and draw one picture of what is inside,","so you start from evidence instead of from the folder name. The last cell says","what you have to build — that part is the task.","","**Reading the files:** on your laptop the notebook env has numpy / scipy /","pandas / scikit-learn / torch / transformers / Pillow — but NOT librosa,","torchaudio or soundfile. The GPU contest box has librosa and torchaudio too."].join("\n"),o=a(["# inventory — what is actually in this pack (counts, sizes, a few real names)","import os","import numpy as np",`DIR = os.environ.get("ICOA_PACK_DIR") or r"${e}"`,"by_ext, total, sample = {}, 0, {}","for root, _dirs, files in os.walk(DIR):","    for f in files:",'        ext = os.path.splitext(f)[1].lower() or "(no ext)"',"        p = os.path.join(root, f)","        by_ext[ext] = by_ext.get(ext, 0) + 1","        total += os.path.getsize(p)","        sample.setdefault(ext, p)",'print(f"{DIR}\\n{sum(by_ext.values())} files · {total/1048576:.0f} MB")',"for ext, n in sorted(by_ext.items(), key=lambda kv: -kv[1])[:8]:",'    print(f"  {ext:>10}  {n:>6}   e.g. {os.path.relpath(sample[ext], DIR)}")',"# 🔴 a leftover .zip means the pack still needs unpacking before you can read it",'zips = [p for e, p in sample.items() if e == ".zip"]',"if zips:",'    print("\\nNOTE: there is still a .zip in here — unpack it first:")','    print(f"  import zipfile; zipfile.ZipFile(r\\"{zips[0]}\\").extractall(DIR)")']),l=["# EDA — one picture of what is inside. icoa ipynb saves it as a .png and","# opens your image viewer (`open` reopens it).","import matplotlib.pyplot as plt","try:"];"image"===i?l.push("    from PIL import Image","    imgs = []","    for root, _d, files in os.walk(DIR):","        for f in sorted(files):",'            if os.path.splitext(f)[1].lower() in (".jpg", ".jpeg", ".png"):',"                imgs.append(os.path.join(root, f))","        if len(imgs) >= 9:","            break","    fig, axes = plt.subplots(3, 3, figsize=(7, 7))","    for ax, p in zip(axes.ravel(), imgs[:9]):","        ax.imshow(Image.open(p))","        ax.set_title(os.path.basename(p)[:18], fontsize=8)","    for ax in axes.ravel():",'        ax.axis("off")','    plt.suptitle(f"{len(imgs)}+ images · first 9")'):"weights"===i?l.push("    import torch","    wfiles = []","    for root, _d, files in os.walk(DIR):","        for f in sorted(files):",'            if os.path.splitext(f)[1].lower() in (".pt", ".pth"):',"                wfiles.append(os.path.join(root, f))","    fig, ax = plt.subplots(1, max(1, min(2, len(wfiles))), figsize=(10, 4), squeeze=False)","    for i, p in enumerate(wfiles[:2]):","        # weights_only=True: never unpickle arbitrary objects from a data file",'        obj = torch.load(p, map_location="cpu", weights_only=True)','        t = obj if hasattr(obj, "flatten") else list(obj.values())[0]',"        v = np.asarray(t, dtype=float).ravel()","        ax[0][i].hist(v, bins=60)",'        ax[0][i].set_title(f"{os.path.basename(p)} · shape {tuple(np.shape(t))}", fontsize=9)','        print(os.path.basename(p), "shape", tuple(np.shape(t)))'):l.push("    import pandas as pd","    csvs = []","    for root, _d, files in os.walk(DIR):","        for f in files:",'            if f.lower().endswith(".csv"):',"                csvs.append(os.path.join(root, f))","    csvs.sort(key=os.path.getsize, reverse=True)","    df = pd.read_csv(csvs[0])","    print(os.path.basename(csvs[0]), df.shape)","    print(df.head(3).to_string(max_colwidth=24)[:600])","    # the last low-cardinality column is the usual label; if there is none,","    # plot the row-length spread so the picture still says something true","    cand = [c for c in df.columns if df[c].nunique() <= 30]","    fig, ax = plt.subplots(figsize=(8, 4))","    if cand:","        vc = df[cand[-1]].astype(str).value_counts().head(12)","        ax.bar(range(len(vc)), vc.values)","        ax.set_xticks(range(len(vc)))",'        ax.set_xticklabels([s[:14] for s in vc.index], rotation=45, ha="right")','        ax.set_title(f"{cand[-1]} — {df[cand[-1]].nunique()} values")',"    else:","        ax.hist([len(str(x)) for x in df[df.columns[-1]]], bins=30)",'        ax.set_title(f"length of {df.columns[-1]}")'),l.push("    plt.tight_layout()","    plt.show()","except Exception as e:","    # never stop the run on a plotting problem — `run all` halts on first error",'    print("EDA skipped:", type(e).__name__, e)');const c=["## Your job","","There is no baseline to beat here — you build the first one. The shape of it:","","1. **turn each file into a row of numbers** — that is the whole trick","image"===i?"   (`PIL.Image.open` → resize → `np.asarray` → flatten, or a pretrained encoder)":"weights"===i?"   (`torch.load(..., weights_only=True)` → the tensors themselves ARE the features)":"   (read the CSVs / unpack the archive, then vectorise the text columns)","2. keep a held-out fifth of the data and print a val score every time","3. write your predictions in exactly the format the task page asks for","","`add` makes a new code cell · `add md` a note cell · `run all` runs everything ·","a cell that raises offers `ai N`, which sends that cell and its traceback to the AI."].join("\n");return[s(r),o,a(l),s(c)]}
+import{sprintTaskForComp as e}from"./sprint-meta.js";export function parseKaggleTaskList(e){const t=e??{},s=(t.data??t).tasks;if(!Array.isArray(s))return[];const a=[];for(const e of s){if(!e||"object"!=typeof e)continue;const t=e;"string"==typeof t.comp_id&&""!==t.comp_id&&a.push({comp_id:t.comp_id,title:"string"==typeof t.title?t.title:t.comp_id,metric:"string"==typeof t.metric?t.metric:"accuracy",description:"string"==typeof t.description&&""!==t.description?t.description:void 0,kaggle_url:"string"==typeof t.kaggle_url?t.kaggle_url:null,licence:"string"==typeof t.licence?t.licence:void 0,attribution:"string"==typeof t.attribution?t.attribution:void 0,data_base:"string"==typeof t.data_base&&""!==t.data_base?t.data_base:null,id_col:"string"==typeof t.id_col?t.id_col:void 0,target_col:"string"==typeof t.target_col?t.target_col:void 0,n_train:"number"==typeof t.n_train?t.n_train:void 0,n_test:"number"==typeof t.n_test?t.n_test:void 0,judge:"string"==typeof t.judge&&""!==t.judge?t.judge:void 0,board_event:"string"==typeof t.board_event&&""!==t.board_event?t.board_event:null,needs_gpu:!0===t.needs_gpu})}return a}export function dataFileUrls(e){const t=e.data_base;if("string"!=typeof t||""===t)return null;const s=t.replace(/\/$/,"");return{train:`${s}/train.csv`,test:`${s}/test.csv`,sample_submission:`${s}/sample_submission.csv`}}export function buildStarterCells(t,s){const a=s.id_col??"id",i=s.target_col??"target";var n;n=s.metric,/rmse|mse|mae|error/i.test(n);const o="icoa-holdout"===s.judge,r=e=>({kind:"markdown",source:e,result:null}),l=e=>({kind:"code",source:e.join("\n"),result:null}),c=[`# ${s.title}`,"",...s.description?[s.description,""]:[],`Kaggle-style practice — metric: **${s.metric}**.`,"","- `train.csv` — labelled training data",`- \`test.csv\` — rows to predict (no \`${i}\` column)`,"- `sample_submission.csv` — the exact format Kaggle expects","","Run all cells top to bottom: a baseline model trains on your CPU and","`submission.csv` appears next to the data. Then improve the model cell",...o?["and re-run — `icoa kaggle submit` scores it and ranks you on the board."]:["and re-run — your score is whatever Kaggle says it is."],"","**AI help:** split your terminal (`Ctrl-b %`), run `icoa` in the right pane","and type `ai4ioai` for a live Gemma 3n E4B. Ask it for feature ideas or bug fixes —","then verify every suggestion with the val score. A tip that lowers it, drop it.",...s.licence?["",`**Data licence:** ${s.licence}`]:[],...s.attribution?[`**Attribution:** ${s.attribution}`]:[]].join("\n"),p=l(["# setup — load the data (leave as is)","import os","import numpy as np","import pandas as pd","from pandas.api.types import is_numeric_dtype","from sklearn.feature_extraction.text import TfidfVectorizer","from sklearn.model_selection import train_test_split","from sklearn.metrics import accuracy_score, mean_squared_error","","# same code everywhere: in a GPU session these env vars point at the","# box's own dataset copy + workspace — locally they are unset.",`DIR = os.environ.get("ICOA_KAGGLE_DATA") or r"${t}"`,'OUT = os.environ.get("ICOA_KAGGLE_OUT") or DIR',...s.data_base?[`DATA_BASE = "${s.data_base.replace(/\/$/,"")}"`,'if not os.path.exists(DIR + "/train.csv"):',"    import urllib.request","    os.makedirs(DIR, exist_ok=True)",'    for f in ("train.csv", "test.csv", "sample_submission.csv"):','        urllib.request.urlretrieve(DATA_BASE + "/" + f, DIR + "/" + f)']:[],'train = pd.read_csv(DIR + "/train.csv")','test = pd.read_csv(DIR + "/test.csv")','sub_fmt = pd.read_csv(DIR + "/sample_submission.csv")',"","# 🔴 The submission column names are NOT always the training column names.","# On tasks imported from other platforms the two genuinely differ: train.csv",'# labels the answer "class" or "result" while sample_submission.csv calls it','# "answer". Hard-coding the submission name here used to raise KeyError on the',"# very first cell. So trust the DATA: whatever column train has and test lacks","# is the thing to predict.",`SUB_ID, SUB_TARGET = "${a}", "${i}"   # names the submission file wants`,"only_train = [c for c in train.columns if c not in test.columns]","TARGET = only_train[-1] if only_train else SUB_TARGET","shared = [c for c in train.columns if c in test.columns]","# Only treat a column as the row id when the submission actually names it AND",'# both files carry it. Falling back to "first shared column" silently spent a',"# real feature as an id on tasks whose id lives only in test.csv.","ID = SUB_ID if SUB_ID in shared else None",'assert TARGET in train.columns, f"cannot find the label column in {list(train.columns)}"',"","# Regression or classification? Decided from the labels, not from metadata —","# several tasks ship continuous labels, and handing those to a classifier",'# raises "Unknown label type: continuous".',"REG = is_numeric_dtype(train[TARGET]) and train[TARGET].nunique() > 20",'print(f"train {train.shape} · test {test.shape} · predicting {TARGET!r} "',"      f\"({'regression' if REG else 'classification'})\")","# 前几行 —— 看一眼真实数据长什么样,比看列名有用得多","print(train.head(3).to_string(max_colwidth=24)[:700])"]),u=e(s.comp_id),d=u?.media,m="audio"===d?.kind?"scipy.io.wavfile.read(path) -> numpy, then scipy.signal for a spectrogram":"image"===d?.kind?"PIL.Image.open(path) -> numpy":"weights"===d?.kind?"torch.load(path, weights_only=True) — the model itself is the input here":"unpack the archive inside the pack, then read what is in it",h=l(["# FEATURES — turn raw columns into numbers a model can use (improve THIS cell).","# Numbers: fill gaps with the median. Category (<= 20 distinct): one-hot.","# Free text: TF-IDF — on a text task the sentences ARE the signal, and dropping",'# them as "high-cardinality" leaves nothing to train on (that used to crash the','# whole notebook with "No objects to concatenate").',"# only columns present in BOTH files — a train-only column cannot be a feature","# at prediction time, and indexing test with it raises KeyError.","feat = [c for c in shared if c not in (ID, TARGET)]","Xtr_raw, Xte_raw = train[feat].copy(), test[feat].copy()","num = [c for c in feat if is_numeric_dtype(Xtr_raw[c])]","cat = [c for c in feat if c not in num and Xtr_raw[c].nunique() <= 20]","# Everything else is vectorised as text. Requiring a long average length here","# used to throw away high-cardinality SHORT strings — ingredient names, single","# words, category codes — which on several tasks are the entire signal, leaving","# the notebook with zero features.","txt = [c for c in feat if c not in num and c not in cat]","for c in num:","    med = Xtr_raw[c].median()","    Xtr_raw[c] = Xtr_raw[c].fillna(med)","    Xte_raw[c] = Xte_raw[c].fillna(med)","blocks_tr, blocks_te = [], []","if num or cat:","    d_tr = pd.get_dummies(Xtr_raw[num + cat], columns=cat, dummy_na=True)","    d_te = pd.get_dummies(Xte_raw[num + cat], columns=cat, dummy_na=True)",'    d_tr, d_te = d_tr.align(d_te, join="left", axis=1, fill_value=0)',"    blocks_tr.append(d_tr.to_numpy(dtype=float))","    blocks_te.append(d_te.to_numpy(dtype=float))","for c in txt:","    # fillna BEFORE astype(str): a real NaN reaching the vectoriser raises",'    # "np.nan is an invalid document", and astype alone does not catch it.','    tr_txt = Xtr_raw[c].fillna("").astype(str)','    te_txt = Xte_raw[c].fillna("").astype(str)',"    # min_df=1: on a short-token column (one ingredient per row) min_df=2 can","    # filter the vocabulary down to nothing.","    long_col = tr_txt.str.len().mean() >= 12","    # Try the good settings, then a permissive fallback, then give up on THIS","    # column only. A column whose vocabulary prunes to nothing used to abort the",'    # whole run with "After pruning, no terms remain".','    for kw in ({"min_df": 2 if long_col else 1,','                "analyzer": "word" if long_col else "char_wb"},','               {"min_df": 1, "analyzer": "char_wb"}):',"        try:","            vec = TfidfVectorizer(max_features=3000, ngram_range=(1, 2), **kw)","            b_tr = vec.fit_transform(tr_txt).toarray()","            blocks_tr.append(b_tr)","            blocks_te.append(vec.transform(te_txt).toarray())","            break","        except ValueError as e:","            last = e","    else:",'        print(f"  skipped text column {c!r}: {last}")',"if not blocks_tr:","    raise SystemExit(",'        f"No usable feature columns. The shared columns are {feat} — none of them "','        f"is numeric, categorical or text.\\n"',...d?[`        "The signal for this task is in ${d.kind} files, not in the CSVs.\\n"`,`        "  1. fetch them (${d.mb}MB):  teleload ${d.pack}\\n"`,...d.glob?[`        "     files land at <pack dir>/${d.glob}\\n"`]:[],`        "  2. replace this cell: ${m}\\n"`,'        "     build one feature row per file, then keep the rest of the notebook.\\n"']:['        f"Two kinds of task land here, and both are real work rather than a bug:\\n"','        f"  1. audio/image — train.csv only lists sample ids and labels and the "','        f"signal is in media files. Read them with scipy.io.wavfile (audio) or "','        f"PIL (images) and turn each file into a feature row.\\n"','        f"  2. train.csv and test.csv describe different things (a retrieval or "','        f"matching task). Read both files, work out what links a test row to a "','        f"train row, and build that mapping yourself.\\n"'],...d?['        f"columns — train={list(train.columns)} test={list(test.columns)}"']:['        f"Either way: replace this cell. train={list(train.columns)} "','        f"test={list(test.columns)}"'],"    )","X = np.hstack(blocks_tr)","X_test = np.hstack(blocks_te)","y = train[TARGET]","X_tr, X_val, y_tr, y_val = train_test_split(X, y, test_size=0.2, random_state=0)",'print(f"{X.shape[1]} model-ready features "','      f"({len(num)} numeric, {len(cat)} one-hot, {len(txt)} TF-IDF text)")']),f=l(["# EDA — look at the labels before you model. Produces ONE figure:","#   left  = what you are predicting (class balance / value spread)","#   right = what the FEATURES cell will have to work with","# icoa ipynb saves it as a .png and opens your image viewer (type `open` to reopen).","import matplotlib.pyplot as plt","try:","    fig, ax = plt.subplots(1, 2, figsize=(11, 4))","    if REG:",'        vals = pd.to_numeric(train[TARGET], errors="coerce").dropna()',"        ax[0].hist(vals, bins=30)",'        ax[0].set_title(f"{TARGET} — {len(vals)} values, mean {vals.mean():.3g}")',"    else:","        vc = train[TARGET].astype(str).value_counts().head(12)","        ax[0].bar(range(len(vc)), vc.values)","        ax[0].set_xticks(range(len(vc)))",'        ax[0].set_xticklabels([s[:14] for s in vc.index], rotation=45, ha="right")','        # An imbalanced target is the single most common reason a "good"',"        # accuracy means nothing — show the majority share right in the title.","        top = vc.iloc[0] / len(train)",'        ax[0].set_title(f"{TARGET} — {train[TARGET].nunique()} classes · biggest {top:.0%}")','    kinds = {"numeric": 0, "category": 0, "text": 0}',"    for c in [c for c in shared if c not in (ID, TARGET)]:","        if is_numeric_dtype(train[c]):",'            kinds["numeric"] += 1',"        elif train[c].nunique() <= 20:",'            kinds["category"] += 1',"        else:",'            kinds["text"] += 1',"    miss = train.isna().mean()","    miss = miss[miss > 0].sort_values(ascending=False).head(10)","    if len(miss) > 0:","        ax[1].barh(range(len(miss)), miss.values)","        ax[1].set_yticks(range(len(miss)))","        ax[1].set_yticklabels([s[:16] for s in miss.index])",'        ax[1].set_title("missing rate (top columns)")',"    else:","        ax[1].bar(list(kinds), list(kinds.values()))",'        ax[1].set_title(f"{sum(kinds.values())} usable columns · no missing values")',"    plt.tight_layout()","    plt.show()",'    print("columns by kind:", kinds)',"except Exception as e:","    # EDA is never allowed to stop the run — `run all` halts on the first","    # error, and a plotting hiccup must not cost the student their baseline.",'    print("EDA skipped:", type(e).__name__, e)']),g=l(["# MODEL — a working baseline. Improve THIS cell to climb the board.","from sklearn.linear_model import LogisticRegression, Ridge","# REG was decided from the labels back in setup.","model = Ridge() if REG else LogisticRegression(max_iter=2000)"]),_=l(["# validate — quick local score on a held-out fifth of train","model.fit(X_tr, y_tr)","p_val = model.predict(X_val)","if REG:",'    print("val RMSE:", round(mean_squared_error(y_val, p_val) ** 0.5, 4))',"else:","    acc = accuracy_score(y_val, p_val)",'    print("val accuracy:", round(acc, 4))',"    if acc < 0.3:","        # A near-zero baseline is the expected result on several of these tasks,","        # not a broken notebook — say so, or the student reads it as a bug.",'        print("    A score this low usually means the linear baseline cannot "','              "represent the pattern at all (combination rules, sequences, "','              "spelling). That IS the task: build features that expose it, "','              "or switch to a model that can. The chain itself works — "','              "submission.csv is written either way.")']),b=l(["# submission — retrain on ALL data, write submission.csv (Kaggle format)","model.fit(X, y)","pred = model.predict(X_test)",'out = OUT + "/submission.csv"',"# Copy the layout of sample_submission.csv and swap in the predictions, rather","# than rebuilding the file from column names. Some tasks want more than two","# columns (e.g. subtaskID,datapointID,answer) and a hand-built 2-column frame","# is rejected by the grader even when every prediction is right.","if len(sub_fmt) == len(pred):","    sub = sub_fmt.copy()","    sub[sub.columns[-1]] = pred","else:","    keys = {}","    if ID is not None and ID in test.columns:","        keys[SUB_ID] = test[ID]","    keys[SUB_TARGET] = pred","    sub = pd.DataFrame(keys)",'    print(f"NOTE: sample_submission has {len(sub_fmt)} rows but test has "','          f"{len(pred)} — built the file from column names instead; "','          f"check it against the task page before submitting.")',"sub.to_csv(out, index=False)",'print("wrote", out, "·", list(sub.columns))',o?`print("score it: icoa kaggle submit ${s.comp_id}")`:'print("format check: icoa kaggle check")']),y=o?["## Submit to ICOA (scored on the server)","",`1. In the ICOA CLI run \`icoa kaggle submit ${s.comp_id}\``,"2. The server scores `submission.csv` against the secret test labels","   and your rank lands on the live board (`icoa board`).","3. Improve the MODEL cell, re-run, submit again — best score counts."]:"nitro"===s.judge?["## Where this one gets scored","","This task comes from **judge.nitro-ai.org** and the answer key stays on","their server, so ICOA cannot score it for you. That does not make it a","practice stub — it is a real task, just judged elsewhere.","",`1. \`icoa kaggle check ${s.comp_id}\` — checks the FORMAT of`,"   `submission.csv` (headers, row count, ids). Format errors are the most","   common way a correct answer still scores zero.","2. Watch the **val score** printed above — that is your real feedback loop","   during the sprint, and it needs no account anywhere.","3. Want the official number? Register at judge.nitro-ai.org and submit","   there. Otherwise compare against the official solution, or ask","   `ai4ioai` to review your approach."]:s.kaggle_url?["## Submit on Kaggle (manual, your own account)","",`1. Open ${s.kaggle_url}`,"2. **Submit Predictions** → upload `submission.csv`","3. Read your leaderboard score, come back, improve the MODEL cell, repeat."]:["## Submit on Kaggle (manual, your own account)","","_This dataset has no live Kaggle page linked yet._","The loop is identical: upload `submission.csv`, read the score, improve","the MODEL cell, repeat. Meanwhile the **val score** printed above is a","real signal you can work against with no account at all."];return[r(c),p,f,h,g,_,b,r(y.join("\n"))]}export function parseSubmitResponse(e){const t=e&&"object"==typeof e?e:{},s=t.data??t,a=s&&"object"==typeof s?s:{};return"number"==typeof a.score&&"string"==typeof a.metric?{ok:!0,score:a.score,metric:a.metric,rank:"number"==typeof a.rank_at_time?a.rank_at_time:null,fieldSize:"number"==typeof a.field_size?a.field_size:null,deduped:!0===a.deduped,boardEvent:"string"==typeof a.board_event?a.board_event:null,used:"number"==typeof a.submissions_used?a.submissions_used:null,cap:"number"==typeof a.submissions_cap?a.submissions_cap:null}:"string"==typeof a.reason&&""!==a.reason?{ok:!1,reason:a.reason,error:"string"==typeof a.error?a.error:void 0,problems:Array.isArray(a.problems)?a.problems.filter(e=>"string"==typeof e):void 0}:{ok:!1,reason:"network"}}export function validateSubmissionText(e,t){const s=[],a=e.trim().split(/\r?\n/),i=t.trim().split(/\r?\n/),n=a[0]??"",o=i[0]??"";n.trim()!==o.trim()&&s.push(`header mismatch: expected "${o}", got "${n}"`);const r=a.slice(1).filter(e=>""!==e.trim()),l=i.slice(1).filter(e=>""!==e.trim());r.length!==l.length&&s.push(`row count mismatch: expected ${l.length}, got ${r.length}`);const c=e=>new Set(e.map(e=>(e.split(",")[0]??"").trim())),p=c(r),u=c(l),d=[...u].filter(e=>!p.has(e)),m=[...p].filter(e=>!u.has(e));if(d.length>0||m.length>0){const e=[];d.length>0&&e.push(`${d.length} missing (e.g. ${d[0]})`),m.length>0&&e.push(`${m.length} unknown (e.g. ${m[0]})`),s.push(`id mismatch vs sample_submission: ${e.join(", ")}`)}for(let e=0;e<r.length;e++)if(r[e].split(",").some(e=>""===e.trim())){s.push(`empty cell in data row ${e+1}`);break}return{ok:0===s.length,problems:s}}export function buildMediaStarterCells(e,t){const s=e=>({kind:"markdown",source:e,result:null}),a=e=>({kind:"code",source:e.join("\n"),result:null}),i=t.media?.kind??"archive",n=t.title??"ICOA sprint task",o=[`# ${t.seq?`${t.seq} · `:""}${n}`,"",...t.topic_cn?[t.topic_cn,""]:[],"This task ships **data, not a baseline** — the CSV starter cannot help here",`because the signal lives in the ${i} files you just pulled.`,"","Run all cells: they inventory the pack and draw one picture of what is inside,","so you start from evidence instead of from the folder name. The last cell says","what you have to build — that part is the task.","","**Reading the files:** on your laptop the notebook env has numpy / scipy /","pandas / scikit-learn / torch / transformers / Pillow — but NOT librosa,","torchaudio or soundfile. The GPU contest box has librosa and torchaudio too."].join("\n"),r=a(["# inventory — what is actually in this pack (counts, sizes, a few real names)","import os","import numpy as np",`DIR = os.environ.get("ICOA_PACK_DIR") or r"${e}"`,"by_ext, total, sample = {}, 0, {}","for root, _dirs, files in os.walk(DIR):","    for f in files:",'        ext = os.path.splitext(f)[1].lower() or "(no ext)"',"        p = os.path.join(root, f)","        by_ext[ext] = by_ext.get(ext, 0) + 1","        total += os.path.getsize(p)","        sample.setdefault(ext, p)",'print(f"{DIR}\\n{sum(by_ext.values())} files · {total/1048576:.0f} MB")',"for ext, n in sorted(by_ext.items(), key=lambda kv: -kv[1])[:8]:",'    print(f"  {ext:>10}  {n:>6}   e.g. {os.path.relpath(sample[ext], DIR)}")',"# 🔴 a leftover .zip means the pack still needs unpacking before you can read it",'zips = [p for e, p in sample.items() if e == ".zip"]',"if zips:",'    print("\\nNOTE: there is still a .zip in here — unpack it first:")','    print(f"  import zipfile; zipfile.ZipFile(r\\"{zips[0]}\\").extractall(DIR)")']),l=["# EDA — one picture of what is inside. icoa ipynb saves it as a .png and","# opens your image viewer (`open` reopens it).","import matplotlib.pyplot as plt","try:"];"image"===i?l.push("    from PIL import Image","    imgs = []","    for root, _d, files in os.walk(DIR):","        for f in sorted(files):",'            if os.path.splitext(f)[1].lower() in (".jpg", ".jpeg", ".png"):',"                imgs.append(os.path.join(root, f))","        if len(imgs) >= 9:","            break","    fig, axes = plt.subplots(3, 3, figsize=(7, 7))","    for ax, p in zip(axes.ravel(), imgs[:9]):","        ax.imshow(Image.open(p))","        ax.set_title(os.path.basename(p)[:18], fontsize=8)","    for ax in axes.ravel():",'        ax.axis("off")','    plt.suptitle(f"{len(imgs)}+ images · first 9")'):"weights"===i?l.push("    import torch","    wfiles = []","    for root, _d, files in os.walk(DIR):","        for f in sorted(files):",'            if os.path.splitext(f)[1].lower() in (".pt", ".pth"):',"                wfiles.append(os.path.join(root, f))","    fig, ax = plt.subplots(1, max(1, min(2, len(wfiles))), figsize=(10, 4), squeeze=False)","    for i, p in enumerate(wfiles[:2]):","        # weights_only=True: never unpickle arbitrary objects from a data file",'        obj = torch.load(p, map_location="cpu", weights_only=True)','        t = obj if hasattr(obj, "flatten") else list(obj.values())[0]',"        v = np.asarray(t, dtype=float).ravel()","        ax[0][i].hist(v, bins=60)",'        ax[0][i].set_title(f"{os.path.basename(p)} · shape {tuple(np.shape(t))}", fontsize=9)','        print(os.path.basename(p), "shape", tuple(np.shape(t)))'):l.push("    import pandas as pd","    csvs = []","    for root, _d, files in os.walk(DIR):","        for f in files:",'            if f.lower().endswith(".csv"):',"                csvs.append(os.path.join(root, f))","    csvs.sort(key=os.path.getsize, reverse=True)","    df = pd.read_csv(csvs[0])","    print(os.path.basename(csvs[0]), df.shape)","    print(df.head(3).to_string(max_colwidth=24)[:600])","    # the last low-cardinality column is the usual label; if there is none,","    # plot the row-length spread so the picture still says something true","    cand = [c for c in df.columns if df[c].nunique() <= 30]","    fig, ax = plt.subplots(figsize=(8, 4))","    if cand:","        vc = df[cand[-1]].astype(str).value_counts().head(12)","        ax.bar(range(len(vc)), vc.values)","        ax.set_xticks(range(len(vc)))",'        ax.set_xticklabels([s[:14] for s in vc.index], rotation=45, ha="right")','        ax.set_title(f"{cand[-1]} — {df[cand[-1]].nunique()} values")',"    else:","        ax.hist([len(str(x)) for x in df[df.columns[-1]]], bins=30)",'        ax.set_title(f"length of {df.columns[-1]}")'),l.push("    plt.tight_layout()","    plt.show()","except Exception as e:","    # never stop the run on a plotting problem — `run all` halts on first error",'    print("EDA skipped:", type(e).__name__, e)');const c=["## Your job","","There is no baseline to beat here — you build the first one. The shape of it:","","1. **turn each file into a row of numbers** — that is the whole trick","image"===i?"   (`PIL.Image.open` → resize → `np.asarray` → flatten, or a pretrained encoder)":"weights"===i?"   (`torch.load(..., weights_only=True)` → the tensors themselves ARE the features)":"   (read the CSVs / unpack the archive, then vectorise the text columns)","2. keep a held-out fifth of the data and print a val score every time","3. write your predictions in exactly the format the task page asks for","","`add` makes a new code cell · `add md` a note cell · `run all` runs everything ·","a cell that raises offers `ai N`, which sends that cell and its traceback to the AI."].join("\n");return[s(o),r,a(l),s(c)]}