cortexa-marketing-feedback (distilled student)

A ~4.4M-parameter conditional decoder distilled from M725/cortexa-marketing-scorer outputs. Takes CLIP-ViT-B/32 vision features (768-d) + the 4 Marketing pillar scores (or a "no-scores" sentinel for fast mode) and emits a creator-vernacular phrase chain:

"scroll stopping | clear cta | thumb stopping"
"forgettable | looks clean | low contrast text"
"lazy design | model looks fake | low contrast"

The student is meant to be the feedback callout shown on the result screen for paid users — plain-language pros and cons that go alongside the scorer's numeric output.

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 ~115; specials <pad>, <bos>, <eos>, <sep>).
config.json	Encoder dim, pillar names, vocab size, special-token ids — read by the TS/JS runtime to shape inputs.

Inference shape

inputs:
  encoder_feats   (1, 768)  float32   # mean-pooled CLIP-ViT-B/32 vision output
  scores          (1, 4)    float32   # [universal_appeal, demographic_appeal, audience_drive, engagement] in [0,1]
  scores_present  (1,)      float32   # 1.0 anchored, 0.0 fast-mode
  input_ids       (1, T)    int64     # decoder context
outputs:
  logits          (1, T, V) float32

Greedy decode works; temperature 0.8 + top-k 20 + SEP-veto is the recommended sampling config when running on more than one input (prevents the greedy "forgettable | forgettable | forgettable" collapse the v0 model exhibited).

Training

15k phrase triples from 5k COCO photos. Each photo scored locally against the cortexa_v10 head; phrase chains generated by research.distill_adjectives.phrase_rules.scores_to_phrase. 12 epochs, AdamW, cosine schedule. Val loss 2.31 → 1.87. See research/distill_students/train_marketing.py in the app repo.

License

Pleius internal — see https://pleius.com. Not for redistribution.