EDCO, Huawei’s entropy‑based dynamic curriculum method, continuously selects the most uncertain samples for domain‑specific LLM fine‑tuning, achieving higher accuracy and more stable gradients across communication, medical, and legal tasks while cutting entropy‑estimation cost by over 80 %.
This article outlines a systematic approach for creating agentic factual SFT and Mid‑train data, covering the definition of training goals, query filtering, two‑layer classification and labeling, trajectory format, differences between Mid‑train and SFT, a practical synthesis pipeline, and common pitfalls to avoid.
A systematic experiment by the Oxford Internet Institute shows that adding a friendly, empathetic personality to large language models via supervised fine‑tuning dramatically raises factual error rates—especially under emotional prompts—while cold, concise tuning leaves accuracy intact.
The 360 AI Research Institute tackled the zero‑tolerance translation demands of airline maintenance by building a specialized parallel corpus and applying RAG‑enhanced, SFT‑fine‑tuned, and RL‑reinforced models, raising Chinese‑to‑English translation accuracy from 70% to 96% and enabling a one‑month rollout.
The preview model Qwopus3.6-27B‑v1, distilled from Claude Opus onto Qwen3.6‑27B using SFT with the Unsloth stack and a curated 12 K high‑quality inference sample set, is evaluated on agentic reasoning, front‑end design, and Canvas/WebGL tasks with an RTX 5090, and can be deployed locally via llama.cpp GGUF quantizations with detailed memory guidelines.
The article explains why pretrained LLMs cannot call tools, then breaks down the three‑stage training pipeline—Supervised Fine‑Tuning, Reinforcement Learning from Human Feedback, and knowledge distillation—showing how each step teaches models to read tool schemas, decide when to invoke a tool, generate JSON calls, and finally transfer the capability to smaller models with LoRA.
SePT (Self‑evolving Post‑Training) shows that a large language model can improve its mathematical reasoning ability by about ten percentage points using a reward‑free online self‑training loop that decouples generation temperature from standard SFT, matching or surpassing RL‑based methods without harming general performance.
This article systematically reviews the core post‑training techniques for large language models—including supervised fine‑tuning, RLHF, PPO, GRPO, DPO, RLVR and Agentic RL—explains their evolution, compares their trade‑offs, and highlights the most promising research directions for 2025‑2026.
The authors show that high SFT accuracy does not guarantee strong RL performance because over‑fitting reduces output diversity, and they propose Adaptive Early‑Stop Loss (AESL), a diversity‑aware early‑stopping objective that dynamically weights token and subsequence losses, yielding consistently better RL results on multiple LLMs and math benchmarks.
This guide maps the end-to-end process for building a small AI model—from leveraging open-source base models and applying SFT with LoRA/QLoRA, through alignment techniques like DPO or ORPO, to low-cost distillation and final quantization for local deployment, while recommending free GPU resources and essential tooling.
The article breaks down how AI product managers can transform a raw base model into a market‑ready, high‑impact product by applying supervised fine‑tuning, tool‑use routing, RLHF alignment, and chain‑of‑thought reasoning, while highlighting trade‑offs, cost shifts, and evaluation metrics.
This article critically examines recent RL‑for‑LLM studies, revealing that reinforcement learning improves search efficiency but does not extend the intrinsic reasoning capabilities of base models, and explores the underlying model‑conditioned optimization bias, comparisons with SFT distillation, and the trade‑off with catastrophic forgetting.
Gaode transforms its map app into a dynamic, AI‑driven “living map” by fine‑tuning the large Spacetime‑GR model through embedding‑based and generative ranking SFT, DPO alignment, and multimodal augmentation, achieving significant offline CTR‑AUC improvements and online CTR gains in POI recommendation.
The article examines the fundamental similarity between SFT and RL loss functions for large language models, explains why RL training is prone to instability, discusses infrastructure and data quality challenges, and reviews practical tricks and reward‑model considerations for more reliable RL fine‑tuning.
This article details how Bilibili’s multimedia lab leveraged a multimodal training pipeline combining data‑compressed SFT and the GRPO reinforcement‑learning algorithm to achieve a 13.5% metric boost and secure second place in the ICCV MIPI Detailed Image Quality Assessment competition.
This article examines the necessity of integrating Supervised Fine‑Tuning (SFT) with Reinforcement Learning (RL) for large language models, surveys alternating, sample‑reuse, simultaneous, and hint‑guided fusion methods, presents the underlying loss functions, and discusses practical trade‑offs such as entropy collapse and importance‑sampling corrections.
This guide explains how to use the EasyDistill framework and Alibaba Cloud PAI to distill large language models for high‑quality text generation, covering model deployment, SFT and DPO training data construction, code examples, configuration files, and best practices for achieving resource‑efficient, high‑performance student models.
This article explains post‑training concepts, compares SFT, DPO, and GRPO fine‑tuning methods, and provides step‑by‑step guidance for using Alibaba Cloud's PAI platform—including Model Gallery and DSW—to fine‑tune large language models with code examples and practical tips.
In an interview at the DA Digital Intelligence Conference, ByteDance AI specialist Cai Conghuai explains how large language models, combined with techniques like SFT, DPO, and RAG, are reshaping Douyin's user‑experience signal detection, root‑cause analysis, and evaluation, while outlining future AI‑agent breakthroughs.
In a detailed interview, Vivo AI engineer Liang Tianan explains how the BlueHeart Small V assistant leverages large language models, multi‑stage recall, ranking, and reward‑model fine‑tuning (SFT/DPO) to generate high‑quality, diverse post‑dialogue recommendation items while balancing latency, cost, and evaluation challenges.
The article analyzes the DAPO technical report, showing how dynamic‑sampling pipelines and token‑level loss handling in SFT and RL training outperform ad‑hoc algorithm tricks, and compares the training dynamics of reinforce_baseline and GRPO with concrete code examples.
We built an LLM‑powered system for Bilibili that automatically creates ad titles from user keywords, employing fluency, style, and quality classifiers, mixed domain data cleaning, and alignment methods such as SFT, DPO and KTO, resulting in a product that now generates about ten percent of daily titles and drives significant ad spend.
This guide explains when to rely on prompt engineering versus SFT fine‑tuning for Vision‑Language Models, emphasizing data quality, appropriate dataset sizes, training epochs, hyper‑parameter tuning, and practical steps to build robust VLM pipelines.
Cleaning SFT data for LLMs is surprisingly complex, as subtle JSON formatting variations, inconsistent markdown wrappers, intent settings, and unit handling can cause model inconsistencies, requiring unified standards, careful prompt design, and extensive manual review to ensure reliable training outputs.
This article provides a comprehensive, experience‑driven guide to supervised fine‑tuning (SFT) of large language models, covering special tokens, latency considerations, data diversity and production, training frameworks and hyper‑parameters, over‑/under‑fitting diagnostics, and evaluation metrics such as helpfulness, honesty, and harmlessness.
The article analyzes why supervised fine‑tuning (SFT) cannot replace reinforcement learning from human feedback (RLHF), highlighting SFT's lack of negative feedback and backward‑looking capability, and explains how RLHF’s reward model addresses these fundamental shortcomings.
The answer explains that joining a pre‑training team accelerates a newcomer’s engineering abilities through hands‑on work with large‑scale data pipelines, distributed training code, and debugging, while SFT teams focus mainly on data labeling, making pre‑training the more effective path for rapid skill growth.
The article analyzes why Reinforcement Learning from Human Feedback (RLHF) cannot be replaced by Supervised Fine‑Tuning (SFT), highlighting SFT's lack of negative feedback, its one‑directional attention limitation, and how RLHF's reward models provide crucial safety and performance improvements for large language models.
The answer argues that fresh graduates should join pre‑training teams because the required engineering tasks—large‑scale data crawling, Hadoop/Spark pipelines, torch and CUDA setup, megatron code debugging, and scaling‑law experiments—rapidly sharpen coding skills, while SFT work focuses mainly on data labeling and offers slower technical growth.
The article breaks down multiple practical approaches for data preparation, training code handling, and experiment analysis in large‑language‑model fine‑tuning, showing how deeper engagement in each step can boost personal expertise even when final model performance appears similar.
Over the past year our team explored applying large language models to specialized domains, detailing their professional benefits, unique challenges such as accuracy and knowledge‑base maintenance, and presenting solutions like alignment enhancement via BPO, Text2API, RAG, and advanced SFT/DPO techniques.
Alibaba’s logistics AI team shares their year‑long journey building a vertical‑domain large language model for logistics, detailing model alignment, Text2API, RAG, SFT techniques, challenges like accuracy and knowledge‑base maintenance, and showcasing real‑world applications such as chatbots, DingTalk assistants, and custom AI assistants.
Vertical domain large language models, fine‑tuned on specialized data, deliver higher expertise and task performance, but require continual knowledge updates and careful alignment; techniques such as BPO‑guided instruction tuning (+1.8% accuracy), Reflexion‑based Text2API (+4% API correctness), advanced RAG preprocessing, and SFT combined with ORPO (+5.2% gain) demonstrate notable improvements while underscoring remaining challenges and collaborative opportunities.
This article details the design, implementation, and evaluation of an AI‑powered tool that automatically creates compliant Git commit messages by leveraging large language models, custom plugins, and performance‑focused optimizations to improve developer productivity and commit quality.
This article presents a comprehensive overview of the development, training, data preparation, algorithmic strategies, evaluation methods, and RAG integration for a domain‑specific patent large language model, highlighting challenges, practical results, and future research directions.
This article examines JD Retail's end‑to‑end large language model framework that combines supervised fine‑tuning, retrieval‑augmented generation, and ReAct‑based AI agents to overcome retail‑specific challenges, improve model accuracy, reduce hallucinations, and enable autonomous multi‑step business workflows.
This article shares hands‑on guidance on using continual pre‑training (CPT), supervised fine‑tuning (SFT), and LoRA adapters for large language models, covering dataset size requirements, learning‑rate scheduling, warm‑up ratios, epoch strategies, and practical routing choices based on real‑world experiments.
This article investigates how mixing data from mathematical reasoning, code generation, and general instruction-following tasks influences supervised fine‑tuning of LLaMA models, revealing distinct scaling curves, resource‑dependent performance conflicts, and a two‑stage DMT strategy that mitigates catastrophic forgetting while boosting overall capability.
This article shares practical techniques for domain‑specific large model continue pre‑training, including data selection, mixing ratios with general data, multi‑task instruction pre‑training, resource‑aware fine‑tuning strategies, evaluation set design, vocabulary considerations, and deployment constraints for 7‑13B models.
Llama 2 advances open‑source large‑model research by expanding context length to 4096, adopting GQA attention, scaling training data to 2 trillion tokens, and introducing refined SFT and RLHF techniques such as Ghost Attention, margin‑based reward modeling, and iterative rejection sampling, all detailed in Meta’s 76‑page report.
