microsoft/thinkingbox-training

microsoft/thinkingbox-training

Description: Training integration with thinkingbox framework.

Language: Python

Stars: 1

Forks: 0

Open issues: 9

Created: 2026-06-01T21:41:42Z

Pushed: 2026-06-20T02:25:44Z

Default branch: main

Fork: no

Archived: no

README:

thinkingbox-training

Fine-tuning / RL training scripts built on top of the thinkingbox framework.

Reproducing the environment

The training stack pins exact versions of every dependency. The GPU stack (PyTorch + vLLM + CUDA wheels) is installed separately because those wheels come from custom indexes and are auto-selected for your local CUDA.

Prerequisites

• Linux with an NVIDIA GPU and a recent CUDA driver
• Python 3.12
• uv (pip install uv or see uv docs)
• Sibling checkout of thinkingbox next to this repo:

parent/
thinkingbox/
thinkingbox-training/ requirements.txt

(The grep filters out the GPU-stack packages installed in step 2; those are pulled in transitively by vllm and pinned by uv's torch backend resolver.)