Technical scenario notes, parameter names and reference answers are currently in English.
Attention & masking
Read the method, change one input, and compare the exported runs.
Causal baseline
Future positions must receive zero probability.
Local calculation / simulation · Open lab ↗Bidirectional context
Removing the mask exposes future keys.
Local calculation / simulation · Open lab ↗Sharp attention
Larger logits concentrate probability on the largest visible score.
Local calculation / simulation · Open lab ↗Flat attention
Equal visible logits yield uniform weights.
Local calculation / simulation · Open lab ↗Soft attention
A smaller scale moves weights toward uniform without changing visibility.
Local calculation / simulation · Open lab ↗Embedding geometry
Read the method, change one input, and compare the exported runs.
Cosine nearest neighbors
A large document norm should not dominate cosine ranking.
Local calculation / simulation · Open lab ↗Dot-product norm bias
The long vector may outrank a better-aligned short vector.
Local calculation / simulation · Open lab ↗Query rotation
Rotating the query changes which document aligns best.
Local calculation / simulation · Open lab ↗Opposite direction
Opposite vectors receive negative cosine scores.
Local calculation / simulation · Open lab ↗Zero-vector edge case
The lab returns zero scores for an undefined cosine denominator.
Local calculation / simulation · Open lab ↗KV cache & paging
Read the method, change one input, and compare the exported runs.
GQA baseline
Eight KV heads use one quarter the cache of 32 heads.
Local calculation / simulation · Open lab ↗Multi-head cache
More KV heads scale cache bytes linearly.
Local calculation / simulation · Open lab ↗Multi-query cache
One KV head sharply reduces logical cache size.
Local calculation / simulation · Open lab ↗Long-context batch
Sequence and batch size multiply memory demand.
Local calculation / simulation · Open lab ↗Page fragmentation
One extra token allocates an additional page per sequence.
Local calculation / simulation · Open lab ↗Train a neural network
Train a tiny neural network and watch its decision boundary change.
Learn XOR quadrants
A nonlinear hidden layer can fit this non-linearly separable task.
Actual browser training · Open lab ↗Underpowered hidden layer
One hidden unit limits the available decision boundaries.
Actual browser training · Open lab ↗Slow learning
A lower learning rate may need more epochs.
Actual browser training · Open lab ↗Circle classification
The hidden layer can approximate a curved boundary.
Actual browser training · Open lab ↗Seed sensitivity
Initialization and sampled data affect the learning trajectory.
Actual browser training · Open lab ↗Bounded agent workflows
Read the method, change one input, and compare the exported runs.
Human approval gate
A write pauses without the mock approval toggle.
Local calculation / simulation · Open lab ↗Authorized mock write
Approval permits only the local mock write.
Local calculation / simulation · Open lab ↗Tool allowlist
A disallowed calculator call stops the workflow.
Local calculation / simulation · Open lab ↗Budget exhaustion
The second paid action is blocked before exceeding budget.
Local calculation / simulation · Open lab ↗Memory truncation
Only the latest result survives FIFO truncation.
Local calculation / simulation · Open lab ↗Hybrid retrieval & reranking
Read the method, change one input, and compare the exported runs.
Hybrid fusion baseline
Combining independent ranks can promote shared candidates.
Local calculation / simulation · Open lab ↗Lexical only
Exact query words alone set the ranking.
Local calculation / simulation · Open lab ↗Dense only
Only vector direction determines the ranking.
Local calculation / simulation · Open lab ↗Oracle reranking ceiling
Perfect labels can improve ordering only inside the candidate pool.
Local calculation / simulation · Open lab ↗Candidate starvation
Even an oracle cannot recover documents excluded before reranking.
Local calculation / simulation · Open lab ↗LLM judge audit
Read the method, change one input, and compare the exported runs.
Mixed judge agreement
Agreement and order consistency measure different properties.
Local calculation / simulation · Open lab ↗Perfect agreement
Perfect labels yield agreement and consistency of one.
Local calculation / simulation · Open lab ↗Always choose first
Selecting the first position fails swap consistency.
Local calculation / simulation · Open lab ↗Tie-heavy dataset
A large tie class can make aggregate agreement misleading.
Local calculation / simulation · Open lab ↗Consistently wrong
A judge can be order-consistent and disagree with every human label.
Local calculation / simulation · Open lab ↗Model routing & fallbacks
Read the method, change one input, and compare the exported runs.
Two-tier routing
Most traffic uses the cheap path; failures add strong-model cost.
Local calculation / simulation · Open lab ↗All strong model
Strong-only routing provides a simple cost baseline.
Local calculation / simulation · Open lab ↗Frequent fallback
High fallback frequency can erase routing savings.
Local calculation / simulation · Open lab ↗Cheap model outage
Every request pays both cost and latency before successful fallback.
Local calculation / simulation · Open lab ↗No primary failures
Expected cost is a weighted average without retries.
Local calculation / simulation · Open lab ↗Quantization & outliers
Read the method, change one input, and compare the exported runs.
Four-bit baseline
Fewer representable values introduce rounding error.
Local calculation / simulation · Open lab ↗Eight-bit reconstruction
More quantization levels usually reduce reconstruction error.
Local calculation / simulation · Open lab ↗Two-bit stress test
Only three signed levels preserve little detail.
Local calculation / simulation · Open lab ↗Outlier dominates scale
An outlier enlarges the global step and collapses small values.
Local calculation / simulation · Open lab ↗All-zero tensor
The zero-range safeguard reconstructs all zeros exactly.
Local calculation / simulation · Open lab ↗Request caching & TTL
Read the method, change one input, and compare the exported runs.
Locality baseline
Repeated keys can avoid repeated model work.
Local calculation / simulation · Open lab ↗Capacity thrashing
A tiny cache evicts useful entries before reuse.
Local calculation / simulation · Open lab ↗TTL expiry
Entries expire at the next request with this interval.
Local calculation / simulation · Open lab ↗High reuse workload
Temporal locality improves the hit rate.
Local calculation / simulation · Open lab ↗All unique requests
Unique keys provide no exact-match cache savings.
Local calculation / simulation · Open lab ↗Tokenizers · RAG chunking · Sampling · Evals
Prompt diff & token impact
Compare two prompt versions line by line and measure their token and UTF-8 byte changes.
Open tool →JSONL dataset inspector
Find malformed records, repeated samples, long inputs, and exact train/validation overlap.
Open tool →RAG chunking lab
Explore Unicode-safe chunk boundaries, overlap, and exact text-token counts.
Open tool →Paired evaluation comparison
Compare two systems on the same examples with a reproducible bootstrap interval.
Open tool →GPU memory estimator
Estimate weights, KV cache, and full-training states before scheduling a run.
Open tool →Sampling lab: logits → tokens
Test temperature, top-k and nucleus sampling against a reproducible local distribution.
Open tool →Speculative decoding lab
Explore when a draft model helps—or hurts—under explicit acceptance and latency assumptions.
Open tool →Local educational experiment. No external API calls. Synthetic fixtures are not production measurements.