TrainRL · Post-training research

From learning signal
to useful behavior.

TrainRL is Vikram Kharvi's independent research log on post-training language models. The work is released one technical blog at a time so each finding has room to be read, tested, and discussed.

Read the current release Follow future releases

Current release / Blog 01

Post-training is a stack.

Start with the end-to-end map: assistant contracts, supervised fine-tuning, preference data, reward models, verifiable RL, PPO, DPO, tools, memory, security boundaries, and independent evaluation.

Read the foundational guide

Public research / 01 blog

Available now

The remaining field notes are preserved off the public site and will be released gradually.

  1. 01
    Blog 01 · 13 Aug 202615 min read

    Post-training Is a Stack

    The full map of supervised tuning, synthetic data, preferences, RL, safety, agent trajectories, distillation, and continual improvement—with guidance on when each method fits.

    Read ↗

Across Art of Cyber AI

Follow the model lifecycle

The post-training research is active. Pre-training and inference are coming soon.