Starred repositories
Search engine experimentation on the TREC collections. Currently focused on zero-shot reranking implementations with typesafe.ai's JEV model
Offline search ranking evaluation from graded judgments: NDCG, MAP, MRR, per-locale (en/hi/mr) breakdowns and significance tests
非自回归System 1决策引擎，专为结构化类型决策场景设计 DeepOpen Multilingual, non-autoregressive System 1 decision engine.
🏠 Intelligent Airbnb Search & Ranking Engine that understands natural-language queries, finds relevant stays using hybrid search, applies real-world filters, and ranks results with XGBoost. 🔎 Built…
A generative recommender, built end to end and post-trained with GRPO to recommend the new items it can already represent.
Code for CIKM 2026《Rethinking Item Tokenization in Generative Recommenders: From Fixed Atoms to Semantic Subwords》DOI: 10.1145/3799682.3840979
[Pytorch] Efficient tokenization for recommendations and generative retrieval. Inspired by STATIC decoding from "Vectorizing the Trie"
Generative recommendation with residual-quantized Semantic IDs, constrained decoding, and cold-start evaluation
Explore semantic-ids with residual quantization for generative recommenders.
One protocol to benchmark and evolve recommenders: classic, generative (HSTU, TIGER, OneRec) and LLM-based, with budgeted agent evolve sessions
Generative news retrieval with MiniOneRec, exposure-aware ranking, and reproducible MIND evaluation
1
Updated Nov 10, 2026
Documents of my JAX and Scaling Book journey
Research artifacts and reproducible analysis for Scaling Laws for Generative Reward Models (arXiv:2610.30765)
reproducing the parabola of Matthias scaling law
The code repo for paper "Capability Scaling-Down Laws for LLM Compression".
All codes used in the paper "Asymmetric Scaling Laws from Sparse Features" published in 2026 NeurIPS
Проверка scaling law в прогнозировании временных рядов: как ошибка моделей зависит от объёма обучающих данных
1
Updated Nov 8, 2026
Covers pre-training data, Tokenizer, Transformer, MoE,distributed training, Scaling Laws, inference & alignment .6 progressive code assignments for full-stack LLM learning | 涵盖预训练数据、分词器、Transformer…
Hands-on LLM systems portfolio: memory anatomy, KV cache, serving SLOs, multi-GPU, distributed training, SFT/DPO/GRPO, Ray, inference optimisation, RAG/agents, with measured benchmarks
Decomposing attribute hallucination in grounded e-commerce product descriptions (CLIP/SigLIP + LLM, APG).
An e-commerce LLM post-training demo with SFT, DPO, GRPO and LoRA, plus local GPU deployment and quantization experiments.
Single-GPU agentic RL: verifiable reward, GRPO diagnostics, offline task sets, VRAM pre-flight
X/Twitter tweet scoring Skill for Claude Code
On the Recall Scaling Laws in Mamba: A Theoretical and Mechanistic Study via Hashing (NeurIPS 2026)