llm-post-training

AIBuildAI LLM-Post-Train Agent: a recursive self-improving (RSI) agent for autonomous LLM post-training

Your RL second brain: 34 source-cited topics from Q-learning to GRPO and agentic RL, kept fresh monthly. Obsidian vault + AI agent layer, Brainstein SSS+.

Learning at the Right Pace: Adaptive Data Scheduling improves LLM reinforcement learning with semantic-cluster and policy-boundary sampling.

Hands-on Qwen3-1.7B code post-training lab: SFT, LoRA, DPO, PPO, GRPO, and reproducible EvalPlus benchmarks.

A novel RL algorithm that designs hierarchical token-level optimization objective for exploration-exploitation balanced policy optimization

Reproducible post-training and evaluation for long-horizon shopping agents: SFT, GRPO, budget-aware ablations, and LLM-as-Judge.

Official code for "Back to the Definition: Estimating Step-Level Advantages via Trajectory Graphs for Agentic Reinforcement Learning"

Official implementation of TRACE: Critic-Verified Credit Localization for LLM Post-Training

Preregistered research on Gemma 4 coding agents: thinking, tool use, trajectory SFT, evaluation, and train/serve consistency.

Paired-intervention toolkit for testing when to switch from on-policy distillation to GRPO.

按优化机制与反馈闭环导航 2017—2026 年关键方法、证据和失效边界｜56 项记录

Machine-verifiable evidence chain for online LLM post-training (TRL GRPO + vLLM): trajectory contracts, fault injection F1-F8, guarded updates, paired gradient replay

Reproducible demo that GRPO will game a legitimate reward model: Qwen2.5-0.5B trained against a real DistilBERT sentiment classifier drives reward up while an independent quality metric collapses, and the same run with a KL penalty keeps both healthy.

CKU Theory: Inspired by 'The Surprising Effectiveness of Deleting Weights in LLM Post-Training' (NeurIPS 2026) - Connecting task-specific Deletion Masks on one shared AI backbone via Cache-to-Cache (C2C) into One Unified AI, by AXNli Inc.

End-to-End Open-Weight LLM Post-Training & Alignment Pipeline: 4-Bit NF4 QLoRA SFT with ChatML <think> Traces, Bradley-Terry DPO Alignment, and Group-128 AWQ vLLM Serving

Critic-free GRPO policy engine for zero-downtime PostgreSQL migrations, with deterministic forge rewards, a 500-case dataset, and TRL/Unsloth training scripts.

金融工单 Agent 的数据工程与后训练实验：数据/意图/工具/政策四张注册表 → SFT / DPO / tool-use SFT 导出 → Qwen2.5-3B LoRA；held-out 政策评测 88/91，加推理时 guard 后 91/91。

CPU-only study of reward-model over-optimisation (Goodharting) in RLHF-style rejection-sampling self-improvement: optimise a verifiable reward and accuracy climbs; optimise a learned reward and the proxy climbs while true accuracy collapses.