https://saidrassai.github.io/Recent content on saidrassaiHugoen-usMon, 15 Jun 2026 00:00:00 +0000https://saidrassai.github.io/blog/finetune-lmf2-5/Mon, 15 Jun 2026 00:00:00 +0000https://saidrassai.github.io/blog/finetune-lmf2-5/In this notebook, we will explore the core concepts of GRPO (Group Relative Policy Optimization) by fine-tuning LFM2.5-1.2B-Instruct using Unsloth . GRPO is a reinforcement learning algorithm designed for training language models with reward signals instead of labeled examples. In contrast to supervised fine-tuning (SFT), where you tell the model the exact right answer, GRPO lets the model explore different outputs and reinforces the ones that score higher on the reward functions.https://saidrassai.github.io/blog/entreprise-cpu-rag/Fri, 05 Jun 2026 00:00:00 +0000https://saidrassai.github.io/blog/entreprise-cpu-rag/This post documents the end-to-end thinking behind reproducing a Fin-R1-style financial reasoning dataset, and what that implies for building a finance agent that can answer questions with genuine chain-of-thought reasoning instead of backward-engineered justifications. 1. HYBRID RETRIEVAL: ARCHITECTURE, FUSION & OPTIMIZATION 1.1 Key Papers & Industry Systems (2025–2026) Paper / System Venue / Org Key Finding for Production RAG Hybrid Search (BM25 + Dense) Industry Standard (Elastic, Vespa, Weaviate 2025) RRF (Reciprocal Rank Fusion) k=60 remains SOTA for zero-shot.https://saidrassai.github.io/blog/finance-agent-journey/Tue, 02 Jun 2026 00:00:00 +0000https://saidrassai.github.io/blog/finance-agent-journey/This post documents the end-to-end thinking behind reproducing a Fin-R1-style financial reasoning dataset, and what that implies for building a finance agent that can answer questions with genuine chain-of-thought reasoning instead of backward-engineered justifications. Why replicate Fin-R1-Data? The official Fin-R1-Data dataset is not publicly available. The authors at SUFE-AIFLM-Lab promised a March 2025 release but never followed through. That leaves practitioners with three options: Reconstruct it from the raw open sources using the described two-stage pipeline.