Clarify Space overview and dataset framing

README.md

@@ -12,25 +12,24 @@ short_description: 'Exploring spectrograms, LWM embedding, and its evaluations.'
 
 # LWM-Spectro Lab
 
-Walk through the Sub-6 GHz → mmWave workflow end-to-end: pretrained LWM encoders digest Sub-6 GHz spectrograms, lightweight heads approximate the residual 1D-CNN beam scorer (~500k params), and MoE embeddings expose the joint SNR/mobility cues used to rank mmWave beams.
+Interactive lab for exploring **LWM spectrogram embeddings** on Sub-6 GHz I/Q baseband data. Browse the bundled DeepMIMO-derived samples (LTE / WiFi / 5G, 128×128 spectrograms, 10.5k total), compare LWM embeddings against raw spectrogram vectors, and probe precomputed MoE embeddings for joint SNR/mobility.
 
 **What you get**
-- **Scenario coverage:** Six DeepMIMO deployments (excluded from LWM pretraining) across LTE / WiFi / 5G with modulation, SNR, and mobility metadata for filtering.
-- **Pipeline checkpoints:** Inspect raw spectrograms, compare LWM embeddings vs. raw vectors via balanced t-SNE, and plug embeddings into k-NN prototypes before training the production beam-selection head.
-- **Evaluation hooks:** Joint SNR/Doppler MoE embeddings approximate the feature bank that conditions the mmWave codebook selector, enabling rapid benchmarking of routing strategies.
+- **Data slices:** LTE / WiFi / 5G Sub-6 GHz I/Q baseband spectrograms with modulation, SNR, and mobility labels; 500 samples per SNR per tech (7 SNRs).
+- **Visualization:** Peek at the raw 128×128 Sub-6 GHz I/Q baseband spectrograms, and compare LWM tech-specific embeddings vs. normalized raw vectors via balanced t-SNE.
+- **Lightweight probes:** Run k-NN prototypes for per-tech modulation recognition and joint SNR/Doppler classification using the cached MoE embeddings alongside raw baselines.
 
 ## Features
-- Visualize LWM embeddings or raw spectrograms with customizable filters.
-- Inspect joint SNR/Doppler performance using cached MoE embeddings and an adaptive k-NN classifier.
-- Upload your own datasets to compare raw channels vs. model embeddings.
+- Filtered gallery of raw Sub-6 GHz spectrograms (tech/SNR/modulation/mobility).
+- t-SNE with SNR-balanced sampling for LWM embeddings vs. raw spectrogram vectors.
+- Adaptive k-NN probes for modulation (per tech) and joint SNR/Doppler (MoE vs. raw) performance.
 
 ## Usage
 1. Select the **Spectrograms** and **t-SNE Analysis** tabs to explore embeddings.
 2. Switch to **Modulation Classification** or **Joint SNR/Doppler Evaluation** to run the k-NN prototype with adjustable train/test splits.
-3. Provide custom data (optional) to benchmark against bundled samples.
 
 ## Tab Cheat Sheet
-- **Spectrograms:** Inspect raw Sub-6 GHz spectrograms per technology/SNR/modulation/mobility before feature extraction.
-- **t-SNE Analysis:** Recreate the "colored by SNR vs. modulation" comparisons from `plot/plot_tsne.py` with balanced sampling.
-- **Modulation Classification:** Benchmark a lightweight k-NN head (standing in for the residual 1D-CNN) on LWM embeddings vs. raw inputs.
-- **Joint SNR/Doppler Evaluation:** Compare MoE embeddings and raw spectrograms on the 14-way SNR/mobility task that feeds the beam selector.
+- **Spectrograms:** Inspect raw 128×128 Sub-6 GHz I/Q baseband spectrograms per technology/SNR/modulation/mobility before feature extraction.
+- **t-SNE Analysis:** Recreate the SNR-ordered scatter plots from `plot/plot_tsne.py` with balanced sampling across SNRs.
+- **Modulation Classification:** Benchmark a lightweight k-NN probe on LWM embeddings vs. raw inputs for each technology.
+- **Joint SNR/Doppler Evaluation:** Compare cached MoE embeddings and raw spectrograms on the 14-way SNR/mobility task.