Jinsong-Zhou/MiniMax-H3-Postrain

MiniMax-H3 causal post-training (T1–T6)

Code and results for the causal teacher-forcing LoRA experiments in code/tutorial.md. That covers 17 table rows run as 13 unique configs, 2 epochs / 2560 steps each, on 8× H100 per experiment.

Path	Contents
code/	Working tree at commit 42ad3ac: RAVEN training code, cluster/ Slurm scripts, and cluster/RUNBOOK.md
git/MiniMax-H3-Postrain.bundle	Full git history of all branches. Restore with git clone git/MiniMax-H3-Postrain.bundle MiniMax-H3-Postrain
runs/minimax_h3_results_20261109_lora/	Final results: 2-minute inference with the trained LoRA (13 experiments × 10 cases), loss curves, configs, and the HTML analysis in analysis/index.html
runs/minimax_h3_results_20261109_noLoRA/	Base-model baseline. These videos were made by an inference run whose LoRA did not load (bug since fixed), so they show the base model

Known issues:

• The T2 (chunk 8/10/15) prompt-switch times use a hard-coded 17-frame chunk (infer/long_video.py), so T2 videos switch prompts early.
• Audio from the LoRA runs is about 10 dB quieter than the base model and the training data; see the analysis notes.

Checkpoints are not included. They remain on the cluster at $W/runs/causal_h3_tf/<exp>/checkpoints.