| --- |
| license: other |
| library_name: pytorch |
| tags: |
| - supermix |
| - multimodal |
| - pytorch |
| - custom-model |
| - reasoning |
| - vision |
| - math |
| - protein-folding |
| - grounded-generation |
| - experimental |
| --- |
| |
| # Supermix Omni Collective V7 Frontier |
|
|
| Custom PyTorch checkpoint for the `omni_collective_v7` frontier model. |
|
|
| ## Included files |
| - `omni_collective_v7_frontier.pth` |
| - `omni_collective_v7_frontier_meta.json` |
| - `omni_collective_v7_frontier_summary.json` |
| - `omni_collective_v7_model.py` |
| - `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` |
| - `protein_folding_model.py` |
| - `train_omni_collective_v7.py` |
|
|
| ## Model summary |
| - Parameters: 77560031 |
| - Stage 1 rows: 28076 |
| - Stage 2 rows: 28254 |
| - Stage 2 validation score: 0.4115 |
| - Stage 2 intent accuracy: 0.7280 |
| - Stage 2 response accuracy: 0.1072 |
| - Stage 2 vision accuracy: 0.5385 |
| - Stage 2 domain accuracy: 0.6844 |
|
|
| ## Training sources |
| - all-model distillation rows: 133 |
| - v33 benchmax rows: 1788 |
| - v39 benchmax rows: 2047 |
| - conversation supermix plus v7: 2600 |
| - conversation creative v7: 1400 |
| - conversation reasoning v7: 1320 |
| - conversation books v7: 920 |
| - conversation science v7: 240 |
| - conversation science novel v7: 200 |
| - conversation coding v7: 380 |
| - math exact v7 added: 548 |
| - protein folding v7 added: 120 |
| - protein pack v7: 120 |
| - science image: 430 |
| - video contact: 248 |
|
|
| ## Notes |
| This is a custom checkpoint, not a standard Transformers `from_pretrained` model. |
|
|
| `v7` extends the omni line with: |
| - all-model distillation across the local Supermix model families |
| - longer multi-pass deliberation with grounded-response guards |
| - broader conversation-focused continuation data |
| - extra math and protein-folding supervision |
| - preservation of text, vision, reasoning, math, and specialist-profile behavior in one checkpoint |
|
|
| The teacher league for this run included: |
| - `v40_benchmax` |
| - `qwen_v28` |
| - `qwen_v30` |
| - `omni_collective_v1` through `omni_collective_v6` |
| - specialist models including `math_equation_micro_v1`, `protein_folding_micro_v1`, and `science_vision_micro_v1` |
|
|
| ## Minimal local usage |
| ```python |
| from pathlib import Path |
| from omni_collective_v7_model import OmniCollectiveEngineV7 |
| |
| engine = OmniCollectiveEngineV7( |
| weights_path=Path("omni_collective_v7_frontier.pth"), |
| meta_path=Path("omni_collective_v7_frontier_meta.json"), |
| ) |
| |
| print(engine.answer("Give a grounded summary of what hydrophobic collapse does during protein folding.")) |
| ``` |
|
|
