Machine Learning on /home/vigi99

Building a Chess Bot with HSTU: From Lichess Pretraining to Value Search

Thu, 16 Jul 2026 00:00:00 +0000

Build Log Snapshot#

This is my running build log for imba-chess. I update it as things change, so it reads out of order in places. That is on purpose.

Where it stands today:

I trained a small HSTU sequence model to imitate high-Elo Lichess games (next-move prediction). It reaches hr@10 around 0.92 and top-1 around 0.43 on held-out games.
On its own, that policy is weak at actual chess. Greedy play scores 0.21 against Stockfish limited to 1400 Elo. Good imitation, bad chess.
Adding a WDL value head and a value-guided search at move-selection time changes the story. The current best system scores 0.595 against Stockfish at 2200 Elo, which puts the whole thing at roughly 2250 Elo. That is about an 800 to 1000 Elo jump over the raw policy, with no reinforcement learning and no self-play yet.
The search that got me there is value_search_halving: sequential halving to decide which root move to spend compute on, plus a prior-ordered beam for the tree underneath it.
The current bet is distilling that search back into the policy head, so the policy itself gets stronger without paying for search at inference. This is grounded in the Gumbel MuZero policy-improvement result. It is designed and being built, not yet a proven win.

Two model sizes show up below: v3 (512-dim, 6 layers, about 10M params) and v4 (768-dim, 8 layers, about 27M params). The bigger v4 trunk finally hit my original 20-30M target and it mattered. Most of the dev loop runs on modest hardware: an 8GB laptop GPU (RTX 3070 class), some runs on a 4090, and a rented 5090 for a while to speed up rollout generation.

FlexAttention HSTU at 500M Events: RQ Tokens, QR Embeddings, and 1D Biases

Sun, 15 Feb 2026 00:00:00 +0000

Executive Summary#

At Yambda scale (about 500M events and 9.4M items), models rarely break because one idea is bad. They break when many small, expensive defaults pile up.

This post is the story of the choices that kept an HSTU-style recommender trainable:

Jagged, block-masked attention with PyTorch FlexAttention
Residual quantization (RQ) token prediction instead of a giant item-ID softmax
Quotient-remainder (QR) embeddings for large sparse categorical spaces
On-the-fly 1D attention bias terms (time, duration, organic) instead of dense [S, S] bias tensors
ALiBi positional bias instead of learned position embeddings

The throughline is simple: keep the inductive bias, cut the dense and quadratic costs.

Machine Learning Engineer Roadmap

Mon, 22 May 2023 00:00:00 +0000

Mastering the Art of AI Development: A Detailed Roadmap from Data to Deployment#

Introduction#

Developing an effective artificial intelligence (AI) model is akin to embarking on a long, complex journey. It demands expertise in areas ranging from dataset creation and feature engineering to model tuning, evaluation, and deployment. This blog post will walk you through the key stages involved in AI development, explain the importance of each, and provide a clear understanding of the skills required at different levels of expertise, namely Junior, Senior, and Staff Engineer.

Supervised Fine-Tuning in Large Language Models

Mon, 22 May 2023 00:00:00 +0000

The Power of Supervised Fine-Tuning in Large Language Models: An In-depth Analysis#

Introduction#

In recent years, the development of machine learning, particularly large language models (LLMs), has revolutionized the way we approach a multitude of challenges, from query-based tasks to content generation. In this post, we will dive deep into a technique gaining traction within the AI community - supervised fine-tuning using domain-specific instruction datasets - and contrast it with the more conventional prompt tuning approach, with a focus on techniques such as retrieval augmentation.

The Role of Negative Mining in Machine Learning: Bridging the Gap in Model Performance

Mon, 22 May 2023 00:00:00 +0000

Introduction#

Machine learning models are excellent tools for making predictions or classifications. However, they’re not infallible; occasionally, they may make mistakes. Some of the most enlightening mistakes are the so-called “hard negatives” — instances where the model confidently produces the incorrect output. Understanding and learning from these instances through hard negative mining can significantly improve the model’s performance.

Understanding Hard Negative Mining#

In machine learning, “hard negatives” refer to examples that are challenging for the model to classify correctly. They are the negatives that the model most often misclassifies. Hard negative mining is a strategy for improving the performance of a model by focusing on these difficult-to-classify instances.

Entity Resolution using Contrastive Learning

Sun, 01 Jan 2023 00:00:00 +0000

Introduction to Entity Resolution#

Entity resolution (also known as entity matching, record linkage, or duplicate detection) is the task of finding records that refer to the same real-world entity across different data sources (e.g., data files, books, websites, and databases).

This can be a challenging task, especially when the dataset is large and the queries mention the attributes of the entities in various ways, such as with partial information, typing errors, abbreviations, or extra information. In this blog post, we’ll be discussing how to approach the Entity Resolution Problem and the solution that was implemented to solve it.