icoa-cli 2.19.448 → 2.19.450

package/dist/lib/kaggle.js

@@ -1 +1 @@
-export function parseKaggleTaskList(e){const t=e??{},a=(t.data??t).tasks;if(!Array.isArray(a))return[];const s=[];for(const e of a){if(!e||"object"!=typeof e)continue;const t=e;"string"==typeof t.comp_id&&""!==t.comp_id&&s.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 s}export function dataFileUrls(e){const t=e.data_base;if("string"!=typeof t||""===t)return null;const a=t.replace(/\/$/,"");return{train:`${a}/train.csv`,test:`${a}/test.csv`,sample_submission:`${a}/sample_submission.csv`}}export function buildStarterCells(e,t){const a=t.id_col??"id",s=t.target_col??"target";var r;r=t.metric,/rmse|mse|mae|error/i.test(r);const n="icoa-holdout"===t.judge,i=e=>({kind:"markdown",source:e,result:null}),o=e=>({kind:"code",source:e.join("\n"),result:null}),l=[`# ${t.title}`,"",...t.description?[t.description,""]:[],`Kaggle-style practice — metric: **${t.metric}**.`,"","- `train.csv` — labelled training data",`- \`test.csv\` — rows to predict (no \`${s}\` 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",...n?["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.",...t.licence?["",`**Data licence:** ${t.licence}`]:[],...t.attribution?[`**Attribution:** ${t.attribution}`]:[]].join("\n"),c=o(["# setup — load the data (leave as is)","import os","import numpy as np","import pandas as pd","from pandas.api.types import is_numeric_dtype","from sklearn.feature_extraction.text import TfidfVectorizer","from sklearn.model_selection import train_test_split","from sklearn.metrics import accuracy_score, mean_squared_error","","# same code everywhere: in a GPU session these env vars point at the","# box's own dataset copy + workspace — locally they are unset.",`DIR = os.environ.get("ICOA_KAGGLE_DATA") or r"${e}"`,'OUT = os.environ.get("ICOA_KAGGLE_OUT") or DIR',...t.data_base?[`DATA_BASE = "${t.data_base.replace(/\/$/,"")}"`,'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 = "${a}", "${s}"   # 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'})\")"]),u=o(["# 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"','        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 the media files shipped with the task. Load them "','        f"(librosa / PIL), 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"','        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)")']),d=o(["# 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)"]),m=o(["# 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.")']),p=o(["# 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))',n?`print("score it: icoa kaggle submit ${t.comp_id}")`:'print("format check: icoa kaggle check")']),h=n?["## Submit to ICOA (scored on the server)","",`1. In the ICOA CLI run \`icoa kaggle submit ${t.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"===t.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 ${t.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."]:t.kaggle_url?["## Submit on Kaggle (manual, your own account)","",`1. Open ${t.kaggle_url}`,"2. **Submit Predictions** → upload `submission.csv`","3. Read your leaderboard score, come back, improve the MODEL cell, repeat."]:["## Submit on Kaggle (manual, your own account)","","_This dataset has no live Kaggle page linked yet._","The loop is identical: upload `submission.csv`, read the score, improve","the MODEL cell, repeat. Meanwhile the **val score** printed above is a","real signal you can work against with no account at all."];return[i(l),c,u,d,m,p,i(h.join("\n"))]}export function parseSubmitResponse(e){const t=e&&"object"==typeof e?e:{},a=t.data??t,s=a&&"object"==typeof a?a:{};return"number"==typeof s.score&&"string"==typeof s.metric?{ok:!0,score:s.score,metric:s.metric,rank:"number"==typeof s.rank_at_time?s.rank_at_time:null,fieldSize:"number"==typeof s.field_size?s.field_size:null,deduped:!0===s.deduped,boardEvent:"string"==typeof s.board_event?s.board_event:null}:"string"==typeof s.reason&&""!==s.reason?{ok:!1,reason:s.reason,error:"string"==typeof s.error?s.error:void 0,problems:Array.isArray(s.problems)?s.problems.filter(e=>"string"==typeof e):void 0}:{ok:!1,reason:"network"}}export function validateSubmissionText(e,t){const a=[],s=e.trim().split(/\r?\n/),r=t.trim().split(/\r?\n/),n=s[0]??"",i=r[0]??"";n.trim()!==i.trim()&&a.push(`header mismatch: expected "${i}", got "${n}"`);const o=s.slice(1).filter(e=>""!==e.trim()),l=r.slice(1).filter(e=>""!==e.trim());o.length!==l.length&&a.push(`row count mismatch: expected ${l.length}, got ${o.length}`);const c=e=>new Set(e.map(e=>(e.split(",")[0]??"").trim())),u=c(o),d=c(l),m=[...d].filter(e=>!u.has(e)),p=[...u].filter(e=>!d.has(e));if(m.length>0||p.length>0){const e=[];m.length>0&&e.push(`${m.length} missing (e.g. ${m[0]})`),p.length>0&&e.push(`${p.length} unknown (e.g. ${p[0]})`),a.push(`id mismatch vs sample_submission: ${e.join(", ")}`)}for(let e=0;e<o.length;e++)if(o[e].split(",").some(e=>""===e.trim())){a.push(`empty cell in data row ${e+1}`);break}return{ok:0===a.length,problems:a}}
+import{sprintTaskForComp as e}from"./sprint-meta.js";export function parseKaggleTaskList(e){const t=e??{},s=(t.data??t).tasks;if(!Array.isArray(s))return[];const a=[];for(const e of s){if(!e||"object"!=typeof e)continue;const t=e;"string"==typeof t.comp_id&&""!==t.comp_id&&a.push({comp_id:t.comp_id,title:"string"==typeof t.title?t.title:t.comp_id,metric:"string"==typeof t.metric?t.metric:"accuracy",description:"string"==typeof t.description&&""!==t.description?t.description:void 0,kaggle_url:"string"==typeof t.kaggle_url?t.kaggle_url:null,licence:"string"==typeof t.licence?t.licence:void 0,attribution:"string"==typeof t.attribution?t.attribution:void 0,data_base:"string"==typeof t.data_base&&""!==t.data_base?t.data_base:null,id_col:"string"==typeof t.id_col?t.id_col:void 0,target_col:"string"==typeof t.target_col?t.target_col:void 0,n_train:"number"==typeof t.n_train?t.n_train:void 0,n_test:"number"==typeof t.n_test?t.n_test:void 0,judge:"string"==typeof t.judge&&""!==t.judge?t.judge:void 0,board_event:"string"==typeof t.board_event&&""!==t.board_event?t.board_event:null,needs_gpu:!0===t.needs_gpu})}return a}export function dataFileUrls(e){const t=e.data_base;if("string"!=typeof t||""===t)return null;const s=t.replace(/\/$/,"");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)]}

package/dist/lib/sprint-meta.d.ts

@@ -0,0 +1,65 @@
+/** Media a task needs that its CSV three-pack does not carry. */
+export interface SprintMedia {
+    /** what the files are — decides which library the guidance names */
+    kind: 'audio' | 'image' | 'weights' | 'archive';
+    /** teleload pack id (also the folder it unpacks into) */
+    pack: string;
+    /** unpacked size, MB — from the live teleload manifest at authoring time */
+    mb: number;
+    /** where the files sit inside the pack, for the "replace this cell" hint */
+    glob?: string;
+}
+export interface SprintTaskMeta {
+    seq: number;
+    bullseye: string;
+    title: string;
+    platform: string;
+    difficulty: number;
+    data_mb: number;
+    needs_gpu: boolean;
+    cli_mode: string;
+    dir?: string;
+    topic_cn?: string;
+    comp?: string | null;
+    judged?: boolean;
+    status?: string;
+    /**
+     * What `run all` prints on the shipped starter, measured by actually running
+     * every notebook (not read off the code). `'none'` = it stops before scoring.
+     * `'pack'` = the teleload pack ships its own baseline notebook.
+     */
+    baseline?: string;
+    media?: SprintMedia;
+}
+/** Data files live next to the built JS (copied by the build), fall back to repo layout. */
+export declare function sprintDataPath(name: string): string;
+export declare function loadSprintTasks(): SprintTaskMeta[];
+/** Look a task up by the competition id the student types (`s01-sound-of-nature`). */
+export declare function sprintTaskForComp(comp: string | null | undefined): SprintTaskMeta | null;
+/**
+ * The commands to type, in order. A media task is TWO steps and the media comes
+ * first — pulling the CSVs first only to hit a wall in cell 3 is the dead end
+ * this function exists to remove.
+ */
+export declare function taskSteps(t: SprintTaskMeta): string[];
+/** One-line answer to "will `run all` give me a number?" */
+export declare function baselineChip(t: SprintTaskMeta, zh: boolean): string;
+/**
+ * What a student can do RIGHT NOW.
+ *
+ * 🔴 The GPU half used to read "⚡ 需 GPU · 排晚场", which sent students to
+ * `gpu on` — and that path is currently broken for sprint tasks: the box's data
+ * tree is `/data/aiproblems/<platform>/<source-name>` while the notebook asks
+ * for `/data/kaggle/<comp_id>`, so the kernel lands on a directory that does not
+ * exist (and these comps carry no `data_base` fallback either). Meanwhile we
+ * MEASURED every one of them producing a baseline on a laptop CPU. So the chip
+ * says the true thing: heavy, GPU optional, starts here. Put "evening window"
+ * back once the box-side hydration lands.
+ */
+export declare function computeChip(t: SprintTaskMeta, zh: boolean): string;
+/**
+ * The media block: what is missing, the command that fetches it, and the first
+ * line of code that reads it. Used by the download guide AND baked into the
+ * starter notebook's stop message, so the two can never disagree.
+ */
+export declare function mediaHelpLines(t: SprintTaskMeta, zh: boolean): string[];

package/dist/lib/sprint-meta.js

@@ -0,0 +1 @@
+import{existsSync as e,readFileSync as t}from"node:fs";import{dirname as n,join as i}from"node:path";import{fileURLToPath as a}from"node:url";const o=n(a(import.meta.url));export function sprintDataPath(t){const n=[i(o,"..","data",t),i(o,"..","..","src","data",t)];for(const t of n)if(e(t))return t;return n[0]}let r=null;export function loadSprintTasks(){if(r)return r;try{const e=JSON.parse(t(sprintDataPath("sprint-tasks.json"),"utf8"));r=Array.isArray(e?.tasks)?e.tasks.slice().sort((e,t)=>e.seq-t.seq):[]}catch{r=[]}return r}export function sprintTaskForComp(e){if(!e)return null;const t=e.trim().toLowerCase();return loadSprintTasks().find(e=>(e.comp||"").toLowerCase()===t)??null}export function taskSteps(e){const t=[];return e.media&&t.push(`teleload ${e.media.pack}`),"teleload"===e.status?t.length>0?t:[`teleload ${e.comp??""}`.trim()]:(e.comp&&t.push(`kaggle ${e.comp}`),t)}export function baselineChip(e,t){const n=(e.baseline||"").trim();if(!n)return"";if("none"===n)return t?"📊 无 baseline — 特征要你自己建(见下)":"📊 no baseline — you build the features (below)";if("pack"===n)return t?"📊 官方 baseline.ipynb 在数据包里":"📊 the pack ships the official baseline.ipynb";const i=t?`📊 开箱 baseline ${n}`:`📊 out-of-the-box baseline ${n}`;return/\b0\.(0\d|1\d|2\d)/.test(n)||/\b0\.000/.test(n)?i+(t?" — 线性模型对这题本就无效,不是坏了":" — a linear model cannot express this task; not a bug"):i}export function computeChip(e,t){return e.needs_gpu?t?"⚡ 数据大/训练重 · GPU 可选 —— baseline 本机就能跑":"⚡ heavy (GPU optional) — the baseline still runs on your laptop":t?"💻 CPU 可做 · 现在就能开":"💻 CPU is enough · start now"}function s(e,t){switch(e){case"audio":return t?"本机 aienv:scipy.io.wavfile.read → numpy → scipy.signal 求谱(没有 librosa)· GPU 箱上 librosa/torchaudio 都有":"on your laptop: scipy.io.wavfile.read → numpy → scipy.signal (no librosa here) · on the GPU box librosa/torchaudio both exist";case"image":return t?"PIL.Image.open 读图 → numpy":"PIL.Image.open → numpy";case"weights":return t?"torch.load(..., weights_only=True) 读权重(输入就是模型)":"torch.load(..., weights_only=True) — the model IS the input here";default:return t?"先解开包里的压缩档再读":"unpack the archive inside the pack first"}}export function mediaHelpLines(e,t){const n=e.media;if(!n)return[];const i="audio"===n.kind?t?"音频":"audio":"image"===n.kind?t?"图片":"images":"weights"===n.kind?t?"模型权重":"model weights":t?"附件档":"an archive",a="none"===(e.baseline||"").trim()?t?`⚠ 这题的信号在${i}里,CSV 只有 id 和标签 —— 通用 starter 会在第 3 格停下。`:`⚠ The signal is in the ${i}; the CSVs carry only ids and labels — the generic starter stops at cell 3.`:t?`⚠ 真信号在${i}里。只喂 CSV 也能跑完,但分数接近瞎猜(${e.baseline})。`:`⚠ The real signal is in the ${i}. CSV-only does finish, but scores near chance (${e.baseline}).`;return t?[a,`▸ 先拉数据(${n.mb}MB):teleload ${n.pack}`,`▸ 然后把 FEATURES 那一格换成:${s(n.kind,!0)}`,...n.glob?[`   文件在 <数据目录>/${n.glob}`]:[]]:[a,`▸ Pull it first (${n.mb}MB):  teleload ${n.pack}`,`▸ Then replace the FEATURES cell:  ${s(n.kind,!1)}`,...n.glob?[`   files live at <data dir>/${n.glob}`]:[]]}

package/package.json

@@ -1,6 +1,6 @@
 {
   "name": "icoa-cli",
-  "version": "2.19.448",
+  "version": "2.19.450",
   "description": "ICOA CLI — The world's first CLI-native cyber & AI security olympiad terminal: AI4CTF (Day 1), CTF4AI (Day 2), VLA4CTF (Pioneer Round — embodied AI)",
   "type": "module",
   "bin": {