AI-powered vocabulary learning. A fine-tuned 1.5B model that runs offline in your browser.
Stunning cognitive topology. Blazing retention. Built for SDG 4.
Demo account: [email protected] / Demo2026!
Features · Benchmarks · Architecture · Quick Start · Weights · 中文简介
⚡ No ads. No subscription. Works offline. The vocabulary model runs in your browser via WebGPU — your learning data never leaves your device.
|
Word library, word details and the fission graph in a three-column workspace — pick any word and its synonym network unfolds in a force-directed star map. |
Choice / spelling / recall modes driven by a memory-strength model. Every answer updates your forgetting curve in real time. |
|
Graded articles with click-to-define on any word — reading is your review session. |
An explainable AI narrative of everything you've learned: CEFR positioning, cognitive radar, growth trajectory. |
More: Ambient screensaver listening (four-season soundscapes synthesized live in Web Audio) · Cognitive Navigator (shortest activation path to your target exam) · My Libraries (built-in + custom word books) · 中英双语界面 · dark mode.
All numbers are real measurements from our 90-question benchmark (greedy decoding) — target-word injection into CEFR-constrained stories, scored on real model outputs. Full report: outputs/eval/BENCHMARK_RESULTS.md.
| Model | Target-word hit ↑ | CEFR violation ↓ | JSON compliance ↑ | FKGL |
|---|---|---|---|---|
| Qwen2.5-1.5B-Instruct (before fine-tuning) | 63.06% | 0.17% | 82.22% | 13.5 |
| + SFT (2,045 diverse stories) | 88.06% | 0.01% | 100% | 9.8 |
| + DPO (real-failure negatives) | 87.78% | 0.01% | 100% | 9.9 |
| Qwen2.5-7B-Instruct (zero-shot) | 39.72% | 0.07% | 0% | — |
| DeepSeek-V3 zero-shot (teacher) | 97.78% | 0.01% | 100% | 10.4 |
| Teacher / Oracle upper bound | 98.33% | 0.22% | 100% | — |
💡 Fine-tuning the small model: +25pp on the same 1.5B, matching the teacher's upper bound — while a larger model without fine-tuning collapses (39.72% hit, 0% valid JSON). Constraint-following is trained, not bought with parameters.
Neuro-symbolic architecture — algorithms handle scheduling, neural models handle generation:
┌───────────────────────────┐ ┌───────────────────────────────┐
│ Symbolic Engine (确定性) │ │ Neural Engine (生成) │
│ · FSRS-6 forgetting curve │ │ · Edge 1.5B (browser, WebGPU) │
│ · Levenshtein scoring │ ⇄ │ · CEFR-constrained stories │
│ · RME word matching │ │ · Real-failure DPO negatives │
└───────────────────────────┘ └───────────────────────────────┘
│ free · explainable · offline │ infinite · personal
▼ ▼
"算法管调度 —— 每毫秒、可解释" "AI 管生成 —— 把复习词写进今天的故事"
The loop: FSRS computes what you should review tomorrow → the fine-tuned model writes it into today's story → reading the story surfaces new words → your memory model updates.
Training pipeline: Data factory (generate + 5-gate filter) → SFT → DPO (real-failure negatives) → Auto-benchmark → 4-bit edge export — full report in docs/模型训练详细报告.md.
# 1. Prerequisites: Node.js ≥ 20, Docker
git clone https://github.com/orionsheep/wordlink.git
cd wordlink
npm install
# 2. Start PostgreSQL (business data)
docker compose up -d
# 3. Configure environment (see docs/部署与运行指南.md for every variable)
# DATABASE_URL is required; auth / AI keys optional
# 4. Run
npm run dev # → http://localhost:3000Demo experience: sign in with
[email protected] / Demo2026!(pre-loaded with 180 days of learning data), or browse public pages without an account. Full deployment guide:docs/部署与运行指南.md.
The complete training pipeline is in model_training/:
python model_training/08_generate_diverse_corpus.py # data factory (LLM teacher + 5-gate filter)
python model_training/02_train_sft.py # SFT
python model_training/03_train_dpo.py # DPO with real-failure negatives
python model_training/eval_real.py --backend local # benchmarkFull report: docs/模型训练详细报告.md
| Repo | Content |
|---|---|
| OrionSheep/wordlink-lexiconstrain | Final model — SFT+DPO merged weights (5.8 GB) + DPO LoRA adapter + eval artifacts |
| OrionSheep/wordlink-edge | SFT v2 merged weights (5.8 GB) + LoRA adapters (v2 / v1 legacy) + eval artifacts |
| OrionSheep/wordlink | This repository's ModelScope mirror |
Lexiverse 语宙 是一个 AI 驱动的英语词汇学习平台：自己微调的 1.5B 模型直接运行在浏览器中（WebGPU 4-bit 量化，~950MB 首次下载后永久离线），把你的到期复习词实时写进今天阅读的故事里。遗忘调度由可解释的 FSRS-6 数学引擎完成，推理零边际成本，学生数据永不出浏览器。核心实测：微调后生词注入命中 88.06%、JSON 合规 100%、CEFR 越界 0.01%。
├── src/app/ # Next.js App Router pages & API routes
├── src/components/ # UI components (welcome landing, study workspace, ambient…)
├── model_training/ # SFT / DPO / corpus factory / benchmark scripts
├── data/ # word libraries & training corpora (jsonl)
├── outputs/eval/ # benchmark metrics & per-question predictions
├── docs/ # training report, deployment guide, media specs
└── public/ # images & demo videos
Built for UNU Macau · AI for SDGs 2026 — AI for Education (SDG 4)
Stunning cognitive topology. Blazing retention.