hyp1231 / Latte

What is Latte?

Latte implements a T5-based generative recommendation model that boosts expressiveness in autoregressive semantic ID generation, tackling limitations where standard GR models fail to capture simple collaborative filtering patterns. You train it on Amazon Reviews 2023 categories like Industrial_and_Scientific or Video_Games, generating conflict-free item IDs via vector quantization (Faiss OPQ, RQ K-means, or RQ-VAE) from sentence embeddings. Python setup with uv sync and one-liner training via train_latte.sh delivers 3.45% NDCG@10 gains over baselines—ideal for dl paper implementations github like this arXiv work.

Why is it gaining traction?

Unlike pure semantic ID baselines (PSID here), Latte prepends learnable latent tokens to sidestep decoding bottlenecks, with agg_max/sum for multi-path aggregation during beam search (num_beams=50-500). Quickstart scripts handle caching, inference, and WandB/TensorBoard logging, letting you swap VQ methods or resume checkpoints seamlessly. Developers grab it for reproducible recsys experiments without rebuilding tokenizers from scratch.

Who should use this?

Recsys researchers replicating generative retrieval papers, ML engineers benchmarking semantic compression on text-rich catalogs like e-commerce reviews, or GenAI teams exploring agentic recommendation limits. Perfect if you're tuning T5 decoders on leave-one-out splits for categories like Musical_Instruments.

Verdict

Solid for academic repros—run main.py --model=Latte --vq_method=rqkmeans and evaluate instantly—but 19 stars and 1.0% credibility signal early-stage maturity; docs are README-focused with no tests. Worth forking for latte review github if you're in dl paper implementations github, else wait for more adoption.

(198 words)