saminbbkhan/vertical-event-lm-freshretail
Document best screen reproduction recipe
Record best observed local screen
Add benchmark-aware local diagnostics
DESIGN.md: document quantile-head fix rationale (Moirai-2/Toto-2 evidence)"
eval_harness: sort quantile preds at inference (non-crossing only at eval, not training)"
Fix quantile head: replace cumsum-softplus (gradient pathology) with independent logits + pinball loss + sort at inference only (Moirai-2/Toto-2 pattern). Add noise augmentation for AR exposure bias.
Comprehensive model card: protocol, design, verified baselines, bug-fix failure analysis
Separate VEL_CODE_REPO (fetch source) from VEL_HUB_MODEL_ID (push target)
Rewrite train_job.py to fetch vel/ package from Hub at runtime (single source of truth; picks up tokenizer fix)
Fix teacher-forcing leakage: supervise regression heads at CENSOR position (before DEMAND token), not at the DEMAND token itself
Upload folder using huggingface_hub
Upload vel/trainer.py with huggingface_hub
Upload train_job.py with huggingface_hub
Upload vel/model.py with huggingface_hub
Upload train_job.py with huggingface_hub
Upload vel/eval_harness.py with huggingface_hub
Upload vel/model.py with huggingface_hub
Upload train_job.py with huggingface_hub
Upload vel/model.py with huggingface_hub
Upload train_job.py with huggingface_hub
Update ML Intern artifact metadata
Upload README.md with huggingface_hub
Upload DESIGN.md with huggingface_hub
Upload train_job.py with huggingface_hub
Upload vel/baselines.py with huggingface_hub
Upload vel/eval_harness.py with huggingface_hub
Upload vel/trainer.py with huggingface_hub
Upload vel/data.py with huggingface_hub
Upload vel/model.py with huggingface_hub
Upload vel/tokenizer.py with huggingface_hub
Update ML Intern artifact metadata
Upload vel/schema.py with huggingface_hub
initial commit