#supervised fine-tuning SFT
181 related articles
·2 min
Alibaba's Qwen Surges to #2 on Text Arena — A Deep Dive into the Model's Capabilities
Alibaba's Qwen LLM surges to #2 on Text Arena via blind human evaluation, showcasing top-tier alignment quality. Analysis of Qwen's technical strengths, open-source strategy, and industry impact.
Read more →
·3 min
Unsloth Now Supports AMD GPUs: 2x Training Speedup and 70% VRAM Savings Across 500+ Models
Unsloth officially supports AMD GPUs across RDNA 3-4, Strix Halo, and MI300 series, delivering 2x training speedup and 70% VRAM savings on 500+ models with RL and vLLM weight sharing support.
Read more →
·3 min
The Debate Over Delayed AI Model Releases: Can a Two-Month Delay Close the Gap with Opus?
Frequent AI model delays have become industry norm. Do delays mean better performance? This article analyzes the tension between delays and expectations, why Claude Opus became the benchmark, and how delays erode user trust.
Read more →
·4 min
The Evolution of Policy Gradient Algorithms: The Complete Lineage from REINFORCE to PPO to GRPO
Trace the evolution of policy gradient algorithms: from REINFORCE's high variance, through Actor-Critic baselines, TRPO's trust regions, PPO's clipping, to GRPO's group baselines for reasoning models.
Read more →
·3 min
Gemini's Hilarious Reply Goes Viral on Reddit: A Deep Dive into the AI Personification Trend
Google Gemini went viral on Reddit for a humorous reply, dubbed an "undercover wasp." This article explores the technical origins of AI humor, RLHF-driven personality shaping, and the future of AI personification.
Read more →
·2 min
Google Releases Gemini 3.6 Flash and Two Other New Models: Comprehensive Expansion of the Flash Family
Google launches Gemini 3.6 Flash, 3.5 Flash-Lite, and 3.5 Flash Cyber, further expanding its lightweight AI product line. Analysis of positioning, differentiation strategy, and developer impact.
Read more →
·3 min
Evaluating Open-Source AI Mathematical Reasoning: Latest Progress and Core Challenges
In-depth analysis of open-source AI models' latest progress in mathematical reasoning, exploring evaluation challenges like data contamination and benchmark saturation, and how formal verification and chain-of-thought methods drive more objective assessment.
Read more →
·3 min
LLM Post-Training on an 8GB GPU: A Practical Guide to SFT, DPO, and GRPO
Learn how to complete LLM post-training on a consumer GPU with just 8GB VRAM, covering SFT, DPO, and GRPO methods using LoRA quantization and other techniques.
Read more →
·3 min
Explorative Modeling: How Best-of-K Training Strategies Improve Model Learning Efficiency
Explorative modeling lets models generate K candidate predictions and learn from the best one, introducing exploration into training. This article analyzes Best-of-K training strategy principles, applications, and challenges.
Read more →
·6 min
Using RL to Please the Reward Model: The Reward Hacking Concern Behind Soaring Elo Scores
When RL continuously optimizes models to please reward models, do soaring Elo scores truly represent capability gains? A deep dive into Reward Hacking in RLHF, Goodhart's Law in AI, and industry countermeasures.
Read more →
·4 min
Heretic Uncensoring Technology Explained: Practical Applications on Jamba2 and Qwen3.5 Open-Source Models
Deep dive into Heretic uncensoring technology applied to Jamba2-Mini, Qwen3.5-9B, and 27B open-source models, exploring how refusal rates dropped from 97% to 4% and the safety debates involved.
Read more →
·5 min
The Complete Learning Path to Understanding the Kimi K3 Technical Report
A systematic learning path for understanding the Kimi K3 technical report, covering MoE, MLA, distributed training, and modern post-training techniques.
Read more →
·2 min
Distilling DeepSeek to GPT-OSS: Why Censorship Doesn't Transfer
When distilling DeepSeek to GPT-OSS, censorship mechanisms don't transfer. This article analyzes the separability of capabilities vs. behavioral constraints in knowledge distillation and its implications for AI safety.
Read more →
·2 min
Distilling DeepSeek to GPT-OSS: Why Censorship Doesn't Transfer
When distilling DeepSeek to GPT-OSS, model censorship doesn't transfer. This article analyzes the separability of capabilities and behavioral constraints in knowledge distillation and its implications for AI safety.
Read more →
·4 min
Freesolo Flash: Deep Dive into the Full-Stack Platform for Enterprise Small Model Training
Freesolo Flash is a full-stack platform for enterprise small language model (SLM) training that commoditizes reinforcement learning, enabling teams to train specialized AI models at low cost.
Read more →
·2 min
Why Do AI Assistants Always Say "Absolutely"? The Technical Truth Behind Excessive Flattery
Why do AI chatbots always start with "Absolutely" and agree with everything? A deep dive into LLM sycophancy, RLHF training side effects, and how to get honest feedback from AI.
Read more →
·2 min
OpenReviewer: An Open-Source LLM Built Specifically for Scientific Paper Review
OpenReviewer is an open-source LLM for generating critical scientific paper reviews. This article analyzes its technical approach, use cases, and limitations.
Read more →
·2 min
OpenReviewer: An Open-Source LLM Built Specifically for Scientific Paper Review
OpenReviewer is an open-source LLM for generating critical scientific paper reviews. This article analyzes its technical approach, use cases, and limitations.
Read more →
·3 min
Lessons from Claude Code's 80% Prompt Reduction: Less Is More in Prompt Engineering
Anthropic cut Claude Code's system prompt by 80% and got better performance. Learn why verbose prompts hurt, how to streamline them, and key takeaways for AI developers.
Read more →
·3 min
Lessons from Claude Code's 80% Prompt Reduction: Less Is More in Prompt Engineering
Anthropic cut Claude Code's system prompt by 80% and got better results. Learn why verbose prompts hurt performance, how to streamline them, and key lessons for AI developers.
Read more →