EpiDebug Multimodal v1 (CPU reasoning seed)

Sibling of punyajain/epidebug-mm-v0.

Philosophy

1. Cannot encode every failure scenario.
2. Encode physics/maths/reasoning capability.
3. Train on failure data.
4. Clean messy multimodal input before the model.
5. Encode cleaned extracts multimodally.

Pipeline

Messy dump → Cleanse/Extract → Multimodal encode → Reason (compete hypotheses / follow-ups)

Files

• model_trainable_only.pt — heads + tabular/image/fusion (load frozen MiniLM from Hub at runtime)
• config.json, metrics.json

Honest limits

CPU seed only — not a frontier reasoner. Scale-up: GPU LoRA on a larger reasoning LM; keep the cleanse gate.