DEDUPED Hot AI repos, turned into use cases Get 5 a day

How to distill a teacher model into a smaller student with Mini Verl

Train a student on its own rollouts using feedback from a teacher model with on-policy distillation.

  • Research
  • Advanced
  • 2 hours
  • From the README
  • Any agent

Who it is forA researcher with one NVIDIA GPU who wants a smaller model to match a larger teacher on its own outputs.

The problemTeacher-student distillation on a single GPU is hard to set up and to inspect, and most tooling assumes large clusters.

The outcomeA student adapter trained on its own rollouts with teacher targets, with visible losses and teacher targets in the logs.

What you need first

  • A single NVIDIA GPU with a CUDA-enabled PyTorch build installed from the PyTorch installer
  • Python with pip available
  • Enough GPU memory for the chosen model (see the single-GPU guide and hardware planning docs)

Steps

  1. Read the OPD quickstart at docs/opd-quickstart.md to choose the teacher and student models.
  2. Check memory needs with the hardware planning doc at docs/hardware-planning.md before you start.
  3. Follow the on-policy distillation steps in docs/opd-quickstart.md and run a dry run first.
  4. Review the logged rollouts, losses and teacher targets to confirm the student is learning.
  5. Export the adapter and configuration using the hand-off steps in docs/verl-opd-scaleout.md.

Prompt for your agent

Using the miniverl repo (DaoyuanLi2816/mini-verl), follow docs/opd-quickstart.md to set up on-policy distillation with teacher <teacher-model> and student <student-model> on one GPU. Report the path to the exported adapter and the final teacher loss. Verify that the peak reserved VRAM stays within the GPU memory you have.

Caveat. The README reports measured VRAM and results for specific Qwen3 and SmolLM2 workloads on an RTX 4080, so other models will need their own memory check.

Share: X LinkedIn Hacker News Email Copy link

Disclaimer

Deduped lists open-source repositories that belong to other people. We do not own, maintain, audit, or vouch for any of them, and a repo appearing here is not an endorsement or a safety check. The use cases, steps, and prompts are written by AI and may be incomplete, outdated, or just wrong. Where a use case is marked "Seen in the wild", we link to public posts where someone described something similar, but we have not verified those posts and the use case itself is still our AI's own write-up. Use all of this at your own risk. Only run code, commands, or agents on systems and accounts you are authorized to use, and read the code and every command before you run it, especially anything that touches credentials, files, or the network. Nothing here is professional, security, legal, or financial advice. All trademarks and project names belong to their respective owners, and their use here does not imply affiliation. Content is provided as is, without warranties of any kind. Found a mistake, a safety problem, or something you want removed? Email [email protected] and we will look at it promptly. Full disclaimer page