Kai9987kai/supermix-omni-collective-v5-frontier

Supermix Omni Collective V5 Frontier

Custom PyTorch checkpoint for the omni_collective_v5 frontier model.

Included files

• omni_collective_v5_frontier.pth
• omni_collective_v5_frontier_meta.json
• omni_collective_v5_frontier_summary.json
• omni_collective_v5_model.py
• omni_collective_v4_model.py
• omni_collective_model.py
• image_feature_utils.py
• image_recognition_model.py
• math_equation_model.py
• train_omni_collective_v5.py

Model summary

• Parameters: 25370219
• Stage 1 rows: 8808
• Stage 2 rows: 9675
• Stage 2 validation score: 0.5270
• Stage 2 intent accuracy: 0.8252
• Stage 2 response accuracy: 0.1508
• Stage 2 vision accuracy: 0.8846
• Stage 2 domain accuracy: 0.8072

Training sources

• 3d: 190
• bible: 180
• books: 620
• coding: 380
• coding_delta_jsonl_v5: 180
• coding_delta_v5: 14
• creative: 940
• creative_100k: 280
• delta_anchor: 169
• delta_official: 32
• dictionary: 180
• finnegans: 120
• hybrid: 859
• openscad_v5: 10
• prompt_understanding_v5: 8
• quality_anchor: 6
• reasoning: 1400
• science: 160
• science_image: 429
• science_novel: 120
• supermix_plus: 2200
• video_contact: 248
• world_events: 66
• qwen_teacher_league_total: 40

Notes

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

v5 continues the v4 omni line with:

• a slightly larger multimodal backbone
• extra coding and OpenSCAD continuation data
• prompt-understanding deltas
• multi-pass deliberative consensus at inference time
• retention of text, math, coding, and image-recognition behavior in one checkpoint

Minimal local usage

from pathlib import Path
from omni_collective_v5_model import OmniCollectiveEngineV5

engine = OmniCollectiveEngineV5(
    weights_path=Path("omni_collective_v5_frontier.pth"),
    meta_path=Path("omni_collective_v5_frontier_meta.json"),
)

print(engine.answer("Write a simple OpenSCAD example for a hollow box with 2 mm walls."))