Unit 01: First Wins (Coding Basics)
No data
0/4 lessons with claims · 0 claims
No claim data
No claims extracted in this run
Built for reliable learning, not fast content
A hands-on curriculum with measurable quality controls
You learn by writing code, getting deterministic feedback, and improving step by step. Behind the scenes, lessons are checked against source research and every release must pass quality gates before it ships.
Curriculum scope is grounded in Attention Is All You Need (Vaswani et al., 2017). Faithfulness snapshots refresh daily and engineering quality snapshots refresh with successful mainline releases.
What We Built
This product teaches core language-model mechanics through short lessons and practical coding loops, not long lectures.
Hands-On Lessons
Every lesson asks you to implement one concrete behavior, run checks, and iterate until your solution works.
Deterministic Feedback
The same code produces the same result, so learners can trust the signal and focus on understanding instead of flaky tooling.
Progress That Sticks
Your code and attempt history are saved on your device by default, so you can continue where you left off without account friction.
Built for Reliability
The learning flow is backed by release guardrails so quality, performance, and accessibility are checked before updates land.
How Learning Works
Read, implement, test, and improve in short loops.
1. Learn one concept at a time with a short explanation and a concrete task.
2. Write code directly in the lesson and run checks with case-level feedback.
3. Fix issues quickly with deterministic reruns and clear pass/fail output.
4. Build confidence through repeated, predictable progress across the catalogue.
Why You Can Trust It
Reliability comes from evidence, reproducibility, and clear gates.
Claims are checked against source evidence from the paper, not ad-hoc summaries.
Learning is active: you prove understanding by shipping working code.
Checks are reproducible, which makes progress signals stable and explainable.
Automated and human review both contribute before content is trusted.
Releases are blocked if critical quality checks fail.
Curriculum Quality Approach
Lessons are designed for clarity first, then validated for consistency and coverage.
How Lessons Are Built and Reviewed
Content starts as plain-language lessons with concrete examples and practice tasks.
Automated validation checks structure, sequence, and exercise integrity so learners get a consistent experience.
We prioritize direct explanations and practical outcomes over broad but shallow coverage.
Engineering Quality Signals
Reliability is treated as a release requirement, not a best effort.
Quality Guardrails
Automated tests verify core learning and grading behavior before release.
Static checks enforce correctness, type safety, formatting, and maintainability limits.
Complexity guardrails catch code that becomes too hard to reason about.
Build verification ensures the production app can ship safely.
Performance, accessibility, and SEO audits run on key learner-facing pages.
If any required gate fails, release is blocked until it is fixed.
Snapshot Source
snapshot
Last Updated
2/9/2026, 2:45:51 PM
Lighthouse Assertion Failures
0
Line Coverage
93.8%
Coverage Snapshot
Coverage is a guarded quality signal, not a claim that every edge case in the product is fully tested.
Lines
93.8% (91/97)
Statements
93.8% (91/97)
Functions
100.0% (6/6)
Branches
79.3% (46/58)
Lighthouse Snapshot
Scores come from repeated representative runs on key learner pages.
| Route | Performance | Accessibility | Best Practices | SEO |
|---|---|---|---|---|
| / | 100% | 100% | 100% | 100% |
| /curriculum | 100% | 100% | 100% | 100% |
| /lesson/01-01 | 93% | 100% | 100% | 100% |
Live Faithfulness Dashboard
Current catalogue-level scores from the latest pipeline run.
Lessons Scored
18/32
Total Claims
63
Faithfulness Gate
PASS
Last Updated
2/9/2026, 3:56:30 AM
How Faithfulness Is Measured
Each lesson is split into factual claims and compared against evidence from the reference paper.
Measurement Method
1. Extract candidate claims from each lesson.
2. Retrieve the most relevant evidence passages from the paper.
3. Score each claim against that evidence.
4. Label claims as entailed, neutral, or contradicted using fixed thresholds.
5. Aggregate metrics and evaluate gate checks for coverage, consistency, and risk rates.
These metrics are strong signals, not proof of absolute truth, and can be affected by retrieval quality and threshold settings.
Metric Definitions
Core formulas used in the report.
Coverage = entailed / total claims
Consistency = (entailed - contradicted) / total claims
Hallucination rate = neutral / total claims
Contradiction rate = contradicted / total claims
Live Method Snapshot
Current scoring configuration for transparency.
Paper Source
aiayn.pdf
Target Files
32
Paper Chunks
27
Retrieval Top-K
5
Embedding Model
Xenova/all-MiniLM-L6-v2
NLI Model
Xenova/nli-deberta-v3-xsmall
Claim Label Thresholds
entailedMin: 0.015
contradictedMin: 0.999
contradictionMargin: 0.95
minSimilarityForSupport: 0.2
minSimilarityForContradiction: 0.25
Gate Thresholds
minCoverage: 0.2
minConsistency: 0.1
minEvidenceQuality: 0.35
maxHallucinationRate: 0.7
maxContradictionRate: 0.15
Coverage
85.7%
Consistency
0.857
Evidence Quality
0.356
Hallucination Rate
14.3%
Contradiction Rate
0.0%
Gate Checks
Threshold checks from the latest faithfulness run.
coverage: PASS
consistency: PASS
evidenceQuality: PASS
hallucinationRate: PASS
contradictionRate: PASS
highConfidenceUnsupportedRate: PASS
Score Legend
Status combines coverage, hallucination risk, contradiction risk, and data availability.
Good
Watch
Poor
No data
Unit Faithfulness
Unit-level quality includes claim depth, label mix, and status from the latest run.
Unit 02: First Working LM (Bigram)
No data
0/4 lessons with claims · 0 claims
No claim data
No claims extracted in this run
Unit 03: Trainable Neural LM
Good
1/4 lessons with claims · 2 claims
2 entailed
0 neutral
0 contradicted
100.0% coverage · 0.0% neutral
Unit 04: Embeddings and Better Context
Good
1/4 lessons with claims · 4 claims
4 entailed
0 neutral
0 contradicted
100.0% coverage · 0.0% neutral
Unit 05: Attention Fundamentals
Good
4/4 lessons with claims · 17 claims
14 entailed
3 neutral
0 contradicted
82.4% coverage · 17.6% neutral
Unit 06: Transformer Block
Good
4/4 lessons with claims · 5 claims
5 entailed
0 neutral
0 contradicted
100.0% coverage · 0.0% neutral
Unit 07: Tiny GPT Training Loop
Good
4/4 lessons with claims · 22 claims
20 entailed
2 neutral
0 contradicted
90.9% coverage · 9.1% neutral
Unit 08: Capstone Quality and Interpretability
Good
4/4 lessons with claims · 13 claims
9 entailed
4 neutral
0 contradicted
69.2% coverage · 30.8% neutral
| Unit | Evidence Depth | Claim Mix | Signal | Status |
|---|---|---|---|---|
| Unit 01: First Wins (Coding Basics) | 0/4 lessons with claims 0 claims extracted | No claim data | No claims extracted in this run | No data |
| Unit 02: First Working LM (Bigram) | 0/4 lessons with claims 0 claims extracted | No claim data | No claims extracted in this run | No data |
| Unit 03: Trainable Neural LM | 1/4 lessons with claims 2 claims extracted | 2 entailed 0 neutral 0 contradicted | 100.0% coverage · 0.0% neutral | Good |
| Unit 04: Embeddings and Better Context | 1/4 lessons with claims 4 claims extracted | 4 entailed 0 neutral 0 contradicted | 100.0% coverage · 0.0% neutral | Good |
| Unit 05: Attention Fundamentals | 4/4 lessons with claims 17 claims extracted | 14 entailed 3 neutral 0 contradicted | 82.4% coverage · 17.6% neutral | Good |
| Unit 06: Transformer Block | 4/4 lessons with claims 5 claims extracted | 5 entailed 0 neutral 0 contradicted | 100.0% coverage · 0.0% neutral | Good |
| Unit 07: Tiny GPT Training Loop | 4/4 lessons with claims 22 claims extracted | 20 entailed 2 neutral 0 contradicted | 90.9% coverage · 9.1% neutral | Good |
| Unit 08: Capstone Quality and Interpretability | 4/4 lessons with claims 13 claims extracted | 9 entailed 4 neutral 0 contradicted | 69.2% coverage · 30.8% neutral | Good |
Lesson Faithfulness
Lesson status shows confidence signals when claims are present, and explicit no-data states when no claims were extracted.
01-01
First Green Test
No data
No claims
No claim data
No claims extracted in this run
Open lesson01-02
Counting What the Model Sees
No data
No claims
No claim data
No claims extracted in this run
Open lesson01-03
Random Choice for Generation
No data
No claims
No claim data
No claims extracted in this run
Open lesson01-04
Unigram Baseline Generator
No data
No claims
No claim data
No claims extracted in this run
Open lesson02-01
Tokenize and Vocabulary
No data
No claims
No claim data
No claims extracted in this run
Open lesson02-02
Build Bigram Counts
No data
No claims
No claim data
No claims extracted in this run
Open lesson02-03
Bigram Probabilities and Predict
No data
No claims
No claim data
No claims extracted in this run
Open lesson02-04
Generate with Bigram
No data
No claims
No claim data
No claims extracted in this run
Open lesson03-01
Make Context Target Examples
No data
No claims
No claim data
No claims extracted in this run
Open lesson03-02
One-Hot and Linear Logits
No data
No claims
No claim data
No claims extracted in this run
Open lesson03-03
Softmax and Cross-Entropy
Good
2 claims
2 entailed
0 neutral
0 contradicted
100.0% coverage · 0.0% neutral
Open lesson03-04
Gradient Step That Reduces Loss
No data
No claims
No claim data
No claims extracted in this run
Open lesson04-01
Embedding Lookup
No data
No claims
No claim data
No claims extracted in this run
Open lesson04-02
Combine Context Vectors
No data
No claims
No claim data
No claims extracted in this run
Open lesson04-03
MLP Language Head
No data
No claims
No claim data
No claims extracted in this run
Open lesson04-04
Train Neural Language Model
Good
4 claims
4 entailed
0 neutral
0 contradicted
100.0% coverage · 0.0% neutral
Open lesson05-01
Query, Key, Value Intuition
Good
3 claims
3 entailed
0 neutral
0 contradicted
100.0% coverage · 0.0% neutral
Open lesson05-02
Attention Scores and Weights
Good
5 claims
4 entailed
1 neutral
0 contradicted
80.0% coverage · 20.0% neutral
Open lesson05-03
Weighted Sum of Values
Good
4 claims
4 entailed
0 neutral
0 contradicted
100.0% coverage · 0.0% neutral
Open lesson05-04
Causal Mask (No Peeking)
Watch
5 claims
3 entailed
2 neutral
0 contradicted
60.0% coverage · 40.0% neutral
Open lesson06-01
Multi-Head Attention
Good
1 claims
1 entailed
0 neutral
0 contradicted
100.0% coverage · 0.0% neutral
Open lesson06-02
Feed-Forward Layer
Good
1 claims
1 entailed
0 neutral
0 contradicted
100.0% coverage · 0.0% neutral
Open lesson06-03
Residual and LayerNorm
Good
1 claims
1 entailed
0 neutral
0 contradicted
100.0% coverage · 0.0% neutral
Open lesson06-04
Full Transformer Block Forward Pass
Good
2 claims
2 entailed
0 neutral
0 contradicted
100.0% coverage · 0.0% neutral
Open lesson07-01
Positional Information
Good
6 claims
5 entailed
1 neutral
0 contradicted
83.3% coverage · 16.7% neutral
Open lesson07-02
Stack Blocks Into MiniGPT
Good
5 claims
4 entailed
1 neutral
0 contradicted
80.0% coverage · 20.0% neutral
Open lesson07-03
Batching and Train Step
Good
7 claims
7 entailed
0 neutral
0 contradicted
100.0% coverage · 0.0% neutral
Open lesson07-04
Validation and Checkpoints
Good
4 claims
4 entailed
0 neutral
0 contradicted
100.0% coverage · 0.0% neutral
Open lesson08-01
Temperature Sampling
Good
4 claims
3 entailed
1 neutral
0 contradicted
75.0% coverage · 25.0% neutral
Open lesson08-02
Top-k Sampling
Good
2 claims
2 entailed
0 neutral
0 contradicted
100.0% coverage · 0.0% neutral
Open lesson08-03
Visualize Attention Maps
Poor
2 claims
0 entailed
2 neutral
0 contradicted
0.0% coverage · 100.0% neutral
Open lesson08-04
Final Model Comparison
Good
5 claims
4 entailed
1 neutral
0 contradicted
80.0% coverage · 20.0% neutral
Open lesson| Lesson | Claims | Claim Mix | Signal | Status | Action |
|---|---|---|---|---|---|
01-01 First Green Test | No claims | No claim data | No claims extracted in this run | No data | Open lesson |
01-02 Counting What the Model Sees | No claims | No claim data | No claims extracted in this run | No data | Open lesson |
01-03 Random Choice for Generation | No claims | No claim data | No claims extracted in this run | No data | Open lesson |
01-04 Unigram Baseline Generator | No claims | No claim data | No claims extracted in this run | No data | Open lesson |
02-01 Tokenize and Vocabulary | No claims | No claim data | No claims extracted in this run | No data | Open lesson |
02-02 Build Bigram Counts | No claims | No claim data | No claims extracted in this run | No data | Open lesson |
02-03 Bigram Probabilities and Predict | No claims | No claim data | No claims extracted in this run | No data | Open lesson |
02-04 Generate with Bigram | No claims | No claim data | No claims extracted in this run | No data | Open lesson |
03-01 Make Context Target Examples | No claims | No claim data | No claims extracted in this run | No data | Open lesson |
03-02 One-Hot and Linear Logits | No claims | No claim data | No claims extracted in this run | No data | Open lesson |
03-03 Softmax and Cross-Entropy | 2 claims | 2 entailed 0 neutral 0 contradicted | 100.0% coverage · 0.0% neutral | Good | Open lesson |
03-04 Gradient Step That Reduces Loss | No claims | No claim data | No claims extracted in this run | No data | Open lesson |
04-01 Embedding Lookup | No claims | No claim data | No claims extracted in this run | No data | Open lesson |
04-02 Combine Context Vectors | No claims | No claim data | No claims extracted in this run | No data | Open lesson |
04-03 MLP Language Head | No claims | No claim data | No claims extracted in this run | No data | Open lesson |
04-04 Train Neural Language Model | 4 claims | 4 entailed 0 neutral 0 contradicted | 100.0% coverage · 0.0% neutral | Good | Open lesson |
05-01 Query, Key, Value Intuition | 3 claims | 3 entailed 0 neutral 0 contradicted | 100.0% coverage · 0.0% neutral | Good | Open lesson |
05-02 Attention Scores and Weights | 5 claims | 4 entailed 1 neutral 0 contradicted | 80.0% coverage · 20.0% neutral | Good | Open lesson |
05-03 Weighted Sum of Values | 4 claims | 4 entailed 0 neutral 0 contradicted | 100.0% coverage · 0.0% neutral | Good | Open lesson |
05-04 Causal Mask (No Peeking) | 5 claims | 3 entailed 2 neutral 0 contradicted | 60.0% coverage · 40.0% neutral | Watch | Open lesson |
06-01 Multi-Head Attention | 1 claims | 1 entailed 0 neutral 0 contradicted | 100.0% coverage · 0.0% neutral | Good | Open lesson |
06-02 Feed-Forward Layer | 1 claims | 1 entailed 0 neutral 0 contradicted | 100.0% coverage · 0.0% neutral | Good | Open lesson |
06-03 Residual and LayerNorm | 1 claims | 1 entailed 0 neutral 0 contradicted | 100.0% coverage · 0.0% neutral | Good | Open lesson |
06-04 Full Transformer Block Forward Pass | 2 claims | 2 entailed 0 neutral 0 contradicted | 100.0% coverage · 0.0% neutral | Good | Open lesson |
07-01 Positional Information | 6 claims | 5 entailed 1 neutral 0 contradicted | 83.3% coverage · 16.7% neutral | Good | Open lesson |
07-02 Stack Blocks Into MiniGPT | 5 claims | 4 entailed 1 neutral 0 contradicted | 80.0% coverage · 20.0% neutral | Good | Open lesson |
07-03 Batching and Train Step | 7 claims | 7 entailed 0 neutral 0 contradicted | 100.0% coverage · 0.0% neutral | Good | Open lesson |
07-04 Validation and Checkpoints | 4 claims | 4 entailed 0 neutral 0 contradicted | 100.0% coverage · 0.0% neutral | Good | Open lesson |
08-01 Temperature Sampling | 4 claims | 3 entailed 1 neutral 0 contradicted | 75.0% coverage · 25.0% neutral | Good | Open lesson |
08-02 Top-k Sampling | 2 claims | 2 entailed 0 neutral 0 contradicted | 100.0% coverage · 0.0% neutral | Good | Open lesson |
08-03 Visualize Attention Maps | 2 claims | 0 entailed 2 neutral 0 contradicted | 0.0% coverage · 100.0% neutral | Poor | Open lesson |
08-04 Final Model Comparison | 5 claims | 4 entailed 1 neutral 0 contradicted | 80.0% coverage · 20.0% neutral | Good | Open lesson |
Known Gaps and What's Next
We publish limitations directly so learners can judge progress honestly.
Faithfulness scoring continues to improve, and some valid teaching claims can still be marked unsupported when retrieval misses context.
Coverage metrics represent a guarded test scope, not exhaustive verification of every edge case in the product.
Some advanced topics are intentionally simplified for beginners; precision notes are added where those simplifications matter.
Faithfulness snapshots refresh daily and quality snapshots refresh with successful mainline releases.