WordLink

See every English word together with its neighbours.

An open-source vocabulary workspace built on a graph of 326k typed word links: synonyms, near-synonyms, word families and look-alike spellings, all mined from open dictionaries. Quizzes, graded reading and an AI tutor sit on top of the graph. The repo also contains the complete SFT + DPO pipeline and benchmark for WordLink-Edge, a 1.5B model fine-tuned to write your review words into graded stories.

Live demo · Features · Quick start · How it works · Model · Roadmap · 中文

Why

Most vocabulary apps drill words one at a time. But you remember abandon better when you also see desert, forsake and give up next to it, and abandonment, abandoned around it. WordLink puts that neighbourhood next to every word you study, so each review also revisits the words around it.

Features

Word graph

The right-hand panel of the workspace is a map of the selected word.

• Colour = relation type. Toggle synonyms, near-synonyms, antonyms and look-alikes from the toolbar.
• Word family as a bubble. Derivatives and inflections (reach → reached, reaching, reachable) sit inside a dashed bubble around the centre word instead of adding more lines.
• One sector per relation. Each group fans out in its own direction with a label such as Synonyms 10/12 +; click it to show the rest.
• Your progress on the map. A ring around each word shows how well you know it: a quarter once you have seen it, two thirds while learning, closed when mastered, red when it is weak.
• Second-degree links on demand (press H), dashed outlines for words without an entry yet, and short Chinese glosses when you zoom in.

Around the graph

• Quizzes: multiple choice, spelling and recall. Every answer updates a per-word memory state that decides when the word comes back.
• Graded reading: articles with click-to-define on every word and the graph in a side drawer.
• Ambient mode: a full-screen word stream with four seasonal soundscapes, synthesized live with Web Audio.
• Also: learning passport (a report of what you have studied), custom word books from CSV, AI tutor chat, English/Chinese interface, light and dark themes.

Quick start

Requirements: Node.js 20+, Docker.

git clone https://github.com/orionsheep/wordlink.git
cd wordlink
npm install

docker compose up -d                 # PostgreSQL on localhost:5432
cp .env.example .env.local           # then edit, see below
npx prisma migrate deploy            # create the tables
npm run migrate                      # load the dictionaries in data/ (takes a while)
npm run import-edges                 # load the 326k typed word links used by the graph

npm run dev                          # http://localhost:3000

Minimal .env.local:

DATABASE_URL="postgresql://postgres:postgres@localhost:5432/english_word_fission_dev?schema=LPT_english"
DEMO_MODE=true          # no auth server needed: every visitor uses the built-in demo account

• Accounts: real sign-up and login use Supabase Auth (supabase start locally) — set the three SUPABASE variables and turn DEMO_MODE off.
• AI features (tutor chat, story generation, passport report) call any OpenAI-compatible endpoint: set LLM_BASE_URL, LLM_MODEL and an API key (SILICONFLOW_APIKEY). Everything else works without one.
• Demo data: npm run demo-data gives [email protected] six months of study history.

Full guide (in Chinese): docs/部署与运行指南.md.

How it works

Layer	What it is
App	Next.js 16 App Router, React 19, Tailwind CSS 4, next-intl
Data	PostgreSQL via Prisma; word data loaded from data/
Graph	GET /api/fission?word=… returns the word's neighbours and their neighbours from the word_relation table; rendered with react-force-graph-2d in src/components/PanelFissionGraph.tsx
Memory	each quiz answer moves a word through UNFAMILIAR → FAMILIAR → LEARNED → MASTERED and sets its next review date (src/lib/memory.ts)
AI	one OpenAI-compatible gateway (src/lib/ai/gateway.ts), prompts in data/ai_prompts/, responses cached on disk
Auth	Supabase Auth, or DEMO_MODE for a single built-in account

The word links come from open data plus a few mining scripts in scripts/:

Relation	Links	Source
Inflections	125,612	ECDICT word-form codes
Look-alike spellings	120,946	longest-common-subsequence score ≥ 0.75, same-family pairs excluded
Synonyms	87,971	synonym table in data/word_fission_data.csv
Near-synonyms (with usage notes)	74,008	"commonly confused" notes in data/word_chinese/
Derivatives	6,221	derivative links in the dictionary entries
Antonyms	0	not collected yet — see the roadmap

The model: WordLink-Edge

WordLink-Edge is Qwen2.5-1.5B-Instruct fine-tuned with SFT + DPO for one job: write a short story at a given CEFR level that uses a given list of target words, returned as JSON. The pipeline in model_training/ covers data generation with a teacher model and five filters, SFT, DPO on real failure cases, and the benchmark.

Measured on a 90-question benchmark with greedy decoding (full report):

Model	Target-word hit ↑	CEFR violations ↓	Valid JSON ↑
Qwen2.5-1.5B-Instruct, before fine-tuning	63.06%	0.17%	82.22%
+ SFT (2,045 stories)	88.06%	0.01%	100%
+ DPO (real failure negatives)	87.78%	0.01%	100%
Qwen2.5-7B-Instruct, zero-shot	39.72%	0.07%	0%
DeepSeek-V3, zero-shot (teacher)	97.78%	0.01%	100%

Fine-tuning adds 25 points on the same 1.5B model, while a 7B model without fine-tuning fails the format entirely.

Weights: OrionSheep/wordlink-lexiconstrain (SFT + DPO merged) and OrionSheep/wordlink-edge (SFT).

Status. The model is trained and benchmarked, but the web app does not run it yet: its AI features call an OpenAI-compatible API. Running WordLink-Edge in the browser with WebGPU is on the roadmap.

Roadmap

• Antonym links (Open English WordNet)
• Run WordLink-Edge in the browser (WebGPU, 4-bit)
• Schedule reviews with the FSRS-6 implementation already in src/lib/reader-engine/fsrs.ts
• Sense-level links: which meaning of a word a synonym belongs to
• Tabs and split view in the workspace
• Clear the remaining ESLint errors

Ideas and pull requests are welcome — see CONTRIBUTING.md.

Repository map

src/app/            pages and API routes
src/components/     UI (workspace, graph, word modules, ambient mode, …)
src/lib/            data access, memory model, AI gateway, reader engine
prisma/             schema and migrations
data/               dictionaries, word links, word lists, AI prompts, training corpora
model_training/     SFT / DPO / corpus / benchmark scripts
outputs/eval/       benchmark metrics and per-question predictions
scripts/            data import and maintenance tools
docs/               deployment guide, training report

License and data

Code: MIT. data/ bundles lexical resources from other projects, including ECDICT (MIT); check each source's terms before reusing the data on its own. Model weights follow the Qwen2.5 license.

Acknowledgements

• ECDICT for phonetics, translations and word forms
• Qwen2.5 as the base model
• The graph's relation chips, word-family bubbles and familiarity marks were inspired by 词汇星图 in HytidelLegend/ai-aug-edu

Built for UNU Macau · AI for SDGs 2026, AI for Education (SDG 4).