Offline Algorithm Trainer

中文说明

I built this because I wanted a simple way to practise algorithms without keeping a browser, an editor, and an online judge open at the same time.

The repository contains 150 interview problems in the hot_150/ directory. Every question is an ACM-style program: its Python file reads from stdin, writes to stdout, and the judge compares the output exactly (whitespace-normalised). This mirrors a real online assessment environment, so the usual loop stays on your machine:

pick a problem -> write its question file -> run it -> inspect the failing case

It also keeps a small SQLite progress file locally. Your attempts and completed solutions are ignored by Git and are not part of this public repository.

If you are learning a pattern for the first time rather than reviewing it, start with the learning protocol. It defines the function-to-ACM transfer, hint ladder, daily session, and mastery criteria. All NeetCode problem links are collected in the central problem index.

Getting started

Python 3.12 or newer is required. There are no third-party Python dependencies.

git clone https://github.com/pllsparrow/offline-algorithm-trainer.git
cd offline-algorithm-trainer

python3 train.py list
python3 train.py show two-sum
python3 train.py run two-sum
python3 train.py run 003
python3 train.py status

Browser training camp

Install the web dependencies once, then start the local-only service:

python3 -m venv .venv
.venv/bin/python -m pip install -r requirements-web.txt
.venv/bin/python web_debug_server.py

Set ALGORITHM_TRAINER_PYTHON when the service should use a specific Python 3.12 interpreter for submitted programs.

The API uses FastAPI and exposes local documentation at /api/docs. The editor is a locally bundled CodeMirror 6 build with Python highlighting, four-space indentation, visible spaces, and breakpoint gutters. Run npm install && npm run build:web only when changing web/editor.js.

Then open http://127.0.0.1:8765. The browser app provides problem navigation, SQLite drafts, curated examples, custom runs, full submissions, breakpoints, continue, step over, locals, and a compact call stack. A submission runs the problem's complete fixed case set and writes the source back to hot_150/ after the run finishes.

Open the qNNN_problem_name.py path printed by show. The file contains only short Input and Output comments, so write the input parsing, algorithm, output, and program entry point yourself. Then run the same problem again. A failed case prints its stdin, expected stdout, and actual stdout.

For a stubborn failure, rerun only that case:

python3 train.py run two-sum --case 1

You can also filter the list:

python3 train.py list --category graph
python3 train.py list --difficulty Hard

ACM text protocol

All 150 problems use a pure text protocol. Each test case has a stdin payload and an expected stdout payload. The judge runs your file as an independent Python process, feeds it the stdin, captures stdout, and compares it (whitespace-normalised) against the expected output.

Each of 148 problems has 50 distinct cases covering boundaries, duplicates, extreme values, sizes, and structural shapes. generate-parentheses and n-queens use 9 and 10 exhaustive/boundary inputs instead; their valid input domains are too small to justify padding the corpus with duplicates.

Common input formats:

• Integer list: count n on one line, then n integers on the next.
• Single integer / float / string token: one line.
• Whole line string (may contain spaces): one line, read with readline.
• Integer matrix: r c on the first line, then r lines of c integers.
• Character board: r c on the first line, then r lines of c chars.
• Binary tree: count n, then n level-order values (null for missing).
• Graph adjacency: count n, then per node degree d and d neighbour ids.
• Linked list: count n, then n integers.
• Operations (design problems): first line q (operation count), then q lines of op args.... Output one result per operation (null for void).

For problems with multiple valid answers (e.g. group-anagrams, 3sum, subsets), the expected output is canonicalised (sorted), so you must print in the same sorted order to pass.

Inspect a problem's exact format before solving it:

python3 train.py show two-sum
python3 train.py check

scripts/build_acm.py deterministically regenerates all 150 specs from data/problems.json and data/tests.json.

Commands I use most

Command	What it does
python3 train.py list	Lists problems and local progress
python3 train.py show <slug>	Shows the problem details, ACM format, and file path
python3 train.py run <slug>	Runs the local ACM test cases
python3 train.py run <number>	Runs by problem number, for example run 001
python3 train.py run <slug> --case 1	Repeats one failing case
python3 train.py run <slug> --all	Runs the remaining cases after a failure
python3 train.py status	Shows attempted and accepted totals
python3 train.py check	Checks the repository data
python3 train.py scaffold --force	Resets all question files to their I/O comments

PyCharm users can select the shared Judge Current Solution run configuration once. After that, open any file in hot_150/ and use the green Run button to judge the current file.

All 150 Python files are stored directly under hot_150/ and named like q001_contains_duplicate.py. The judge is in judge/, while problem metadata, tests, and your local progress live under data/.

Heap and bucket follow-up exercises live in muscle_memory/heapq&buckets/. This set contains fifteen interview-focused variants with 99 deterministic cases each and semantic validators for problems that allow tied answers.

Recursion drills live in muscle_memory/recursion/. The twenty exercises cover recursion fundamentals, search, divide and conquer, and tree recursion. Ten are tree-focused, with 49 deterministic cases per exercise.

About the problem content

Problem names and links refer to exercises on LeetCode and NeetCode. This repository contains my own summaries, hints, starter templates, test harness, and local test data. It does not republish full commercial problem statements or official solutions. See NOTICE.md for details.

The original code in this repository is released under the MIT License.