VISHAL SINGH·Jun 22Reward hacking in Reinforcement learningA field guide to reward hacking in GRPO — why it happens, how it hides, and what actually fixes it
VISHAL SINGH·Jun 14Why Your Fine-Tuned LLM Forgets Everything — and How Self-Synthesized Replay Fixes ItFine-tuning makes a general LLM useful for your domain. But it also destroys general capabilities — coding, math, instruction following…
VISHAL SINGH·Jun 11Weight tying in language models: when and why LLMs share embeddingsMost transformer models made the same architectural choice for years without questioning it. Then they stopped. This post looks at the data…
VISHAL SINGH·Jun 3Why your fine-tuned retrieval model isn’t as good as you think it isYour model trains well, evaluates well — then quietly fails in production. The problem isn’t your architecture. It’s what you chose as your…
VISHAL SINGH·Feb 26, 2024Comparing Open Source LLMsIn recent years, large language models (LLMs) have revolutionized natural language processing and generated immense excitement in both…