The Vocabulary System
Your Mind Deserves
Stunning cognitive topology. Blazing retention. Built by FSRS-6, refined by authentic contexts. This is vocabulary learning, wildly reimagined.
Pro features. Zero complexity.
Designed for cognition.
Built to retain.
Every sentence is intentional. In our four-season ambient focus space, sentences emerge with liquid-glass staggered reveal, paired with synthesized Web Audio nature soundscapes and authentic IndexTTS voices.
Explore Reading SpaceIt gets smarter.
Automatically.
Your vocabulary stream evolves with your learning state. The Playlist Vocabulary Hub allows real-time syllabus selection, shuffle curation, and cadence pacing (8s/12s/16s). Zero UI discipline ensures seamless passive learning.
Launch Word StreamArchitecture
A neuro-symbolic hybrid engine,where algorithms and models each play their part
We refuse AI for AI's sake. Memory scheduling is handled by an explainable math engine (FSRS forgetting curve + edit-distance grading); language generation is handled by a domain-fine-tuned neural model — determinism guarantees pedagogical rigor, generative power guarantees endlessly fresh content.
Symbolic · Symbolic Engine
Mathematical Deterministic Core
- · FSRS forgetting curve: millisecond-level solving of stability S / difficulty D / retrievability R
- · Levenshtein edit distance: zero-error spelling grading
- · RME multi-objective matching: precisely recalls due words from the lexicon
- · Traits: free, explainable, works offline
Neural · Neural Engine
Domain-Fine-Tuned Generative Model
- · CEFR-level constraints: vocabulary difficulty locked within the learner's level
- · Target-word slot injection: due words woven seamlessly into real contexts
- · Structured output: bilingual JSON + vocabulary explanations
- · Traits: unlimited content, personalized, runs on-device
Purpose-Fine-Tuned Models
WordLink-Edge
On-Device in Browser · LiveQwen2.5-1.5B + domain-fine-tuned LoRA (full SFT + DPO pipeline)
Model
WordLink-Edge, a 1.5B model born for teaching,every number reproducible
Built on Qwen2.5-1.5B and domain-fine-tuned through a full pipeline of diverse-corpus SFT and DPO on real failure negatives — all genuinely trained on local Apple Silicon. A small model that matches its commercial-flagship teacher, and lives in the browser offline forever.
Target-Word Injection Hit Rate · 90-Question Benchmark
(82.22% before fine-tuning)
(0.17% before fine-tuning)
(13.5 → matched to A2/B1)
Same Questions Compared · Why Domain Fine-Tuning Is a Must
temperature=0 · greedy decoding
DPO Behavior Alignment, Real Convergence
Negatives all come from the 7B model's real failure outputs; four acceptance checks all passed with zero reward hacking — validating alignment safety, not gaming metrics.
In the Browser, Offline Forever
WebGPU 4-bit on-device inference — zero network and zero marginal cost after the first download; student data never leaves the device (GDPR / COPPA friendly) — equally usable in low-connectivity regions.
Pedagogical-Grade Data Engineering
DeepSeek-V3 teacher generation × 44 topics × A2/B1/B2 tiers, five hard filters (90.5% pass rate); de-duplicated against the test set by prompt for auditable evaluation hygiene.
Pipeline
Five-Stage Training Pipeline
Data Factory
Teacher model generates diverse micro-stories; CEFR word list + difficulty band + structure — five hard filters; 2,045 entries selected (90.5% pass rate)
SFT Supervised Fine-Tuning
LoRA r=64 injection; learns target-word slot injection + CEFR boundaries + JSON structured output
DPO Preference Alignment
All negatives from the model's real failure outputs; trust-region CLIP prevents degradation; margins genuinely climb 0→11
Automated Evaluation
90-question benchmark · greedy decoding · four dimensions: hit rate, out-of-bounds rate, JSON compliance, FKGL difficulty
On-Device Quantized Export
4-bit quantization → WebGPU in-browser inference; ~950MB, offline forever after the first download
Real Evaluation Comparison · 90-question benchmark · greedy decoding · temperature=0
| Model | Word Injection Hit ↑ | JSON Compliance ↑ | FKGL Difficulty | Notes |
|---|---|---|---|---|
| Qwen2.5-1.5B base (pre-tuning) | 63.06% | 82.22% | 13.5 | Product baseline |
| + Domain fine-tuning SFT | 88.06% | 100% | 9.8 | Same model, +25pp |
| + DPO preference alignment | 87.78% | 100% | 9.9 | On par with SFT, more stable behavior |
| DeepSeek-V3 zero-shot | 97.78% | 100% | — | Commercial flagship reference |
| Teacher upper bound (reference answers) | 98.33% | 100% | — | Fine-tuning catches up with it |
* All metrics are computed by an automated evaluator on real model outputs (temperature=0, greedy decoding); evaluation details and per-question outputs are stored in the repository at outputs/eval/. Word hit = share of due review words naturally embedded in the text; JSON compliance = share of structured outputs directly consumable by programs.
Product Tour
Ten feature modules,one screenshot each
All of the following are real screenshots of the running product (not design mockups). Every module is one click away from a live experience.
Workspace
Workspace
Your post-login learning hub: lexicon, study modes, star map and quiz entries one screen away, with day/night themes that follow your habits
Try it live →
Study
Star-Map Workbench
Lexicon outline × word details × fission star map, three panes in sync: select any word and its synonym network unfolds instantly in a force-directed star map
Try it live →
Quiz
Smart Quiz
Three question types — multiple choice, spelling, recall — generated intelligently with a memory-strength model; every answer updates your forgetting curve
Try it live →
Immersive
Immersive Star Map
A full-screen force-directed star map with draggable floating panels — your library and word details hover above the galaxy, and every node click fissions further
Try it live →
Ambient
Four-Season Ambient
Real scenes of spring, summer, autumn and winter plus natural soundscapes synthesized in real time with Web Audio — words drift by with the passing landscape
Try it live →
Read
Context Reading
A graded article stream: click any word while reading for instant definitions, save it to your wordbook with one tap — reading is reviewing
Try it live →
Passport
Learning Passport
AI translates all your learning data into an explainable English credential narrative + cognitive radar, making learning outcomes visible
Try it live →
Report
Learning Report
Streaks, accuracy trends, vocabulary mastery distribution and recent activity — your progress at a glance
Try it live →
Libraries
My Libraries
Built-in syllabus wordbooks plus custom libraries in any combination — grouped study, import and export on demand
Try it live →
Settings
Settings
Interface language, sound, display and learning preferences all adjustable, with seamless Chinese-English bilingual switching
Try it live →
You dream it. We anchor it.
Connect every word into a second brain. FSRS-6 handles memory curves, RME-V5 serves the contexts, and AI Dictation closes the loop.