| --- |
| language: |
| - en |
| license: other |
| license_name: pleius-internal |
| tags: |
| - onnx |
| - conditional-text-generation |
| - video-feedback |
| - distillation |
| - creator-tools |
| --- |
| |
| # cortexa-create-feedback (distilled student) |
|
|
| A ~4.4M-parameter conditional decoder distilled from |
| `M725/cortexa-create-scorer` outputs. Takes CLIP-ViT-B/32 vision |
| features (mean-pooled across video frames, 768-d) + the 5 Create pillar |
| scores and emits a creator-vernacular phrase chain about the short-form |
| video: |
|
|
| ``` |
| "first frame slaps | feels intentional" |
| "thumb stopping | shareable" |
| "filler | feels rushed | first frame is nothing" |
| "feels off beat | slow open | no payoff" |
| ``` |
|
|
| ## Files |
|
|
| | file | purpose | |
| |---|---| |
| | `student_int8.onnx` | TinyTransformer decoder, 4 layers / 256-dim / 4 heads, INT8 dynamic-quantized. 6.9 MB. | |
| | `tokenizer.json` | Whole-phrase tokenizer (vocab ~138; specials `<pad>`, `<bos>`, `<eos>`, `<sep>`). | |
| | `config.json` | Encoder dim, pillar names, vocab size, special-token ids. | |
|
|
| ## Inference shape |
|
|
| ``` |
| inputs: |
| encoder_feats (1, 768) float32 # mean-pooled CLIP-ViT-B/32 vision across frames |
| scores (1, 5) float32 # [hook, hold, algorithmic_fit, brand_lift, overall] in [0,1] |
| scores_present (1,) float32 # 1.0 anchored, 0.0 fast-mode |
| input_ids (1, T) int64 |
| outputs: |
| logits (1, T, V) float32 |
| ``` |
|
|
| Same sampling recommendation as `cortexa-marketing-feedback`: temperature |
| 0.8 + top-k 20 + SEP-veto. |
|
|
| ## Training |
|
|
| 6k phrase triples from 3 real short-form videos |
| (`public/create-tutorial/*.mp4`) + 1997 synthetic "videos" built by |
| random-crop + color jitter over COCO stills (each frame goes through |
| cortexa_v10 separately, so the per-frame curve has real variation). 15 |
| epochs. Val loss 2.39 → 1.97. See |
| `research/distill_students/train_create.py` in the app repo. |
| |
| ## License |
| |
| Pleius internal — see https://pleius.com. Not for redistribution. |
| |
