supermix-v39-frontier-reasoning-plus

Supermix V39 Frontier Reasoning Plus

This is the public Hugging Face package for the v39_final line from the Supermix workspace.

Important naming note

In the local app and benchmark files, this model is referred to as v39_final.

Inside the packaged training artifact, the chosen checkpoint file is:

• champion_model_chat_v39_frontier_reasoning_plus_stage2.pth
• chat_model_meta_v39_frontier_reasoning_plus_stage2.json

That is because the stage-2 checkpoint was selected as the final chosen artifact for the v39 run.

Files

• champion_model_chat_v39_frontier_reasoning_plus_stage2.pth
• chat_model_meta_v39_frontier_reasoning_plus_stage2.json
• v39_frontier_reasoning_plus_chosen_checkpoint.json
• v39_frontier_reasoning_plus_dataset_summary.json
• v39_frontier_reasoning_plus_training_summary.json
• v39_reasoning_benchmix_summary.json
• model_frontier_v39.py
• model_variants.py
• run.py

Important Note

This is a custom PyTorch checkpoint, not a standard Transformers from_pretrained model.

Local Evaluation Context

Saved local recipe holdout score:

• 29 / 528
• recipe eval accuracy: 0.0549

There was no completed common-benchmark sweep for this model because the training pod ran out of credit before that pass finished.

Usage

This checkpoint requires the included Python source files to construct the model class before loading the .pth weights.