hf_train_job: install huggingface_hub before downloading source/dataset

hf_train_job: use snapshot_download (v1/v3 working pattern)

hf_train_job: lazy-import huggingface_hub to match v3 pattern

hf_train_job: add upload diagnostics and skip guard

hf_train_job: revert to mount-based writable registry pattern from v3

hf_train_job: snapshot_download source+data, writable registry

Lazy-load embedding engine; force HF_REGISTRY_PATH to /data/registry.db

hf_train_job: persistent diagnostic log + registry path tracing

hf_train_job: flush logs, trap import errors for HF debugging

v4: three-tier descriptor targets (v3), collate padding sentinel, HF job v3 dataset

Update HF job config to mount source and dataset volumes

v4: time-aware objective head, physics-derived v2 targets, Pyrfume vocab

fix(loss): prevent objective-starvation with fixed task weights; target pimt_v3.pt

feat: replace sigmoid gating with stabilized FiLM affine modulation; train pimt_v2.pt

fix(hf_job): prepare writable registry before pino imports

fix(hf_job): writable registry + output dirs for read-only source mounts

feat(hf_job): allow mounted data/registry paths via env vars

feat(ingest): TGSC demo formula scraper + odor cache

feat(naturals,shalimar): hierarchical natural oil tokenization

fix(embeddings): handle registry records with empty openpom/shape embeddings

feat(embeddings): expand input token from 145-D to 151-D with physical-functional payload

feat(artifacts): add paper artifact generator with validation metrics and UMAP POM figure

chore(hf): download warm registry.db cache from HF dataset repo before training

chore(hf): download warm registry.db cache from HF dataset repo before training

perf(embeddings): implement write-through SQLite cache for 3D shape features

perf(train): avoid Hugging Face Dataset round-trip during formula-disjoint split

docs(readme): recommend t4-medium (30 GB RAM) to avoid OOM on full dataset

build: lower Python requirement to >=3.10 for official PyTorch image compatibility

feat(semantics): add literature-based odor descriptor vocabulary builder

feat(scripts): add Pyrfume descriptor and ODT ingestion pipeline

feat(scripts): add dataset split validation utilities

feat(scripts): add shape-feature backfill script for AromaRegistry

feat(scripts): add UNIFAC subgroup audit tool for registry compounds

chore(gitignore): ignore generated artifacts, datasets, checkpoints, and logs

fix(train): keep AdaptiveLossBalancer forward signature intact, expose last_weights attribute

feat(train): log AdaptiveLossBalancer weights per batch for telemetry verification

docs(readme): add HF Hub card metadata and MIT LICENSE

docs(readme): overhaul public README for HF source repo

chore(hf): point hf_train_job.py to empirical_dataset_v1.jsonl

chore(train): point default data path to empirical dataset and confirm balancer/loss setup

feat(data): zero-extend early-depleted trajectories to fixed 49-step grid

feat(data,verifier): add skip_ifra to verifier for synthetic dataset generation

feat(data): add commercial solvent dilution layer and clean natural profiles

feat(data): deploy integrated empirical bootstrap generator

fix(data,verifier): key trajectories by CAS and read precomputed targets

feat(train): pin 597 single-molecule controls exclusively to training

refactor(loss): rename DeepSeekLossBalancer to AdaptiveLossBalancer

feat(train,heads,loss): DeepSeekLossBalancer + continuous cosine alignment

feat(structure,thermo): 3D shape + natural oil reconstruction, 145-D input

fix(pyproject): add datasets dependency