CLASSICAL ML0/10
0/ 30 SOLVED
DEEP LEARNING0/10
TRANSFORMERS0/8
INTERPRETABILITY0/2
0 wonML-ARENA BENCHMARKS
| Status | PROJECT | TOPICS (YELLOW = NEW) | BENCHMARK | INDUSTRY | MINE | EST. |
|---|---|---|---|---|---|---|
| BL-01.Bike-demand predictor MLPYTHONafter original | NEW ml.linregNEW ml.featuresNEW ml.eval | test RMSE | sklearn LinearRegression | — | 5 h | |
| BL-02.SMS spam filter MLPYTHONafter original | NEW ml.naive-bayesml.evalml.features | F1 on held-out messages | sklearn MultinomialNB | — | 5 h | |
| BL-03.Rain-tomorrow forecaster MLPYTHONafter original | NEW ml.logregml.evalml.features | log-loss; calibration | sklearn LogisticRegression | — | 5 h | |
| BL-04.Recipe recommender MLPYTHONafter originalFITS NOW | NEW ml.knnml.featuresml.eval | recall@10; queries per second | sklearn NearestNeighbors | — | 5 h | |
| BL-05.Loan decision explainer MLPYTHONafter Breiman et al. | NEW ml.treesml.evalml.features | accuracy at the same depth | sklearn DecisionTreeClassifier | — | 7 h | |
| BL-06.Colour-palette extractor MLPYTHONafter original | NEW ml.clusteringml.knnml.features | inertia | sklearn KMeans (n_init=10) | — | 5 h | |
| BL-07.Map of anything MLPYTHONafter van der Maaten & HintonFITS NOW | NEW ml.dimredml.clusteringml.eval | PCA output; neighbourhood trustworthiness | sklearn PCA · sklearn TSNE | — | 6 h | |
| BL-08.Face-or-not detector MLPYTHONafter original | NEW ml.svm-kernelsml.evalml.dimred | accuracy | sklearn SVC | — | 7 h | |
| BL-09.Tiny gradient booster MLPYTHONafter original | ml.treesml.evalml.features | test accuracy | sklearn HistGradientBoostingClassifier | — | 9 h | |
| BL-10.AutoML-lite MLPYTHONafter original | ml.evalml.featuresml.treesml.naive-bayesml.svm-kernels | CV scores; leakage tests | sklearn cross_val_score | — | 8 h | |
| BL-11.Neural net from scratch MLPYTHONafter Andrej Karpathy | NEW dl.mlpNEW dl.backpropml.logreg | gradients; accuracy | finite differences | — | 6 h | |
| BL-12.Deep-net rescue MLPYTHONafter original | NEW dl.init-normNEW dl.optimdl.backprop | training curve | PyTorch, same setup | — | 7 h | |
| BL-13.Flower classifier MLPYTHONafter original | NEW dl.regularisedl.mlpml.eval | validation accuracy | PyTorch reference model | — | 6 h | |
| BL-14.Convolution engine MLPYTHONafter original | NEW dl.cnnNEW sys.kernelsdl.backprop | conv outputs; speedup | torch.nn.functional.conv2d | — | 9 h | |
| BL-15.Style writer MLPYTHONafter Andrej Karpathy | NEW dl.rnnNEW tf.lmdl.backprop | gradients; validation loss | PyTorch LSTM · n-gram baseline | — | 8 h | |
| BL-16.Word map MLPYTHONafter Mikolov et al. | NEW dl.embeddingsNEW tf.tokenizeml.knn | analogy accuracy | gensim Word2Vec, same corpus | — | 8 h | |
| BL-17.Spoken-digit recogniser MLPYTHONafter original | NEW dl.audio-tsdl.cnndl.rnnml.eval | test accuracy | librosa spectrograms · measured | — | 7 h | |
| BL-18.Sprite generator MLPYTHONafter original | NEW dl.generativedl.cnndl.optim | reconstruction loss; sample grid | PyTorch reference VAE | — | 10 h | |
| BL-19.Game-playing agent MLPYTHONafter Mnih et al. | NEW dl.rldl.mlpdl.optim | average return on CartPole-v1 | Gymnasium solved threshold: 475 | — | 8 h | |
| BL-20.Paper recommender MLPYTHONafter Kipf & Welling | NEW dl.gnndl.embeddingsml.eval | node-classification accuracy on Cora | GCN paper: 81.5% | — | 8 h | |
| BL-21.Tiny GPT MLPYTHONafter Andrej Karpathy | NEW tf.attentiontf.lmtf.tokenizedl.init-normdl.regularise | validation loss | nanoGPT, same data and size | — | 12 h | |
| BL-22.Fast GPT MLPYTHONafter original | NEW tf.inferenceNEW sys.memorytf.attention | speed; output match; perplexity | your uncached fp32 model | — | 8 h | |
| BL-23.Faster training MLPYTHONafter original | NEW tf.trainingNEW sys.paralleldl.optim | time to target loss | your fp32 baseline · measured | — | 8 h | |
| BL-24.Teach GPT-2 a trick MLPYTHONafter Hu et al. | NEW tf.finetuneNEW tf.evaltf.lmtf.training | eval loss; trainable params | full fine-tune · measured | — | 8 h | |
| BL-25.Ask your notes MLPYTHONafter original | NEW tf.retrievaltf.evaldl.embeddingsml.knn | recall@5 on 50 questions you write | BM25 (rank_bm25) | — | 8 h | |
| BL-26.Search photos by text MLPYTHONafter original | NEW tf.multimodaldl.embeddingstf.retrieval | recall@5 | pretrained CLIP zero-shot · measured | — | 9 h | |
| BL-27.Serve your GPT MLPYTHONafter original | NEW sys.servingtf.inferencesys.parallelsys.memory | p50/p95 latency; throughput | one-request-at-a-time server | — | 7 h | |
| BL-28.Mini compiler MLPYTHONafter original | NEW sys.compilerssys.kernelsdl.backprop | speed; output match | your unfused version | — | 10 h | |
| BL-29.X-ray your GPT MLPYTHONafter nostalgebraist | NEW interp.lensNEW interp.probesNEW interp.circuitstf.attentionml.dimred | probe accuracy; induction scores | TransformerLens on GPT-2 small | — | 10 h | |
| BL-30.Feature finder and steering wheel MLPYTHONafter Anthropic | NEW interp.saeNEW interp.steeringtf.lmml.clustering | reconstruction error and sparsity; steering success | SAELens defaults, same activations | — | 12 h |
INDUSTRY is the library, paper or record you measure against. MINE is your latest recorded result. Base Form projects add one or two new topics each and reuse the rest.