maximg1/ced-demo-ensemble

Put the demo's model weights here

make demo-live loads the ensemble from this folder. It ships empty: the weights are ~1.7 GB, far past what belongs in git (GitHub rejects any file over 100 MB, and each model.safetensors here is 480-740 MB), so everything except this README is gitignored.

What goes here

demo/models/phase2_config.json
demo/models/v2_threshold.json
demo/models/checkpoints/roberta_inverse_seed42/
demo/models/checkpoints/deberta_inverse_seed42/
demo/models/checkpoints/modernbert_sqrt_inverse_seed42/

Each checkpoint is a directory (weights + tokenizer + training_meta.json); copy it whole.

phase2_config.json is not optional. It records the encoders, seeds, input_format and class-weight strategies those checkpoints were trained with, and the server reads the ensemble's composition from it rather than assuming today's defaults.

v2_threshold.json carries the tuned operating point (0.52, the one the paper reports). It is a copy of outputs/phase1_broad/threshold.json, the Phase 1 run the operating point is tuned on. Without it the server falls back to the same 0.52 and logs a warning, so the demo still runs, but copy it and the number is traceable.

Where it comes from

The Phase 3 (mixed-era) run on the cluster, under outputs/phase3/. Three of its thirty checkpoints: one seed of the macro-F1 arm, which is what a laptop can hold. The strategy names come from outputs/phase1_broad/strategy_tables.json; a different Phase 1 selection run would name different directories.

rsync -avP -e 'ssh -J <user>@slurm-client.cs.tau.ac.il' \
  --include='phase2_config.json' \
  --include='checkpoints/' \
  --include='checkpoints/roberta_inverse_seed42/***' \
  --include='checkpoints/deberta_inverse_seed42/***' \
  --include='checkpoints/modernbert_sqrt_inverse_seed42/***' \
  --exclude='*' \
  <user>@c-003:/home/yandex/DLWorkShop2025b/<user>/congressional-empirical-discourse/outputs/phase3/ \
  demo/models/

Then the operating point, from the Phase 1 run (it is in the local outputs/ once that is synced):

cp outputs/phase1_broad/threshold.json demo/models/v2_threshold.json

Replace the whole folder rather than adding to it: an older phase2_config.json names older checkpoints, and the server reads the ensemble's composition from it.

Already have the run locally? cp the same four paths out of outputs/phase3/ instead.

Using a different ensemble

make demo-live CHECKPOINT_ROOT=outputs/phase2   # the legacy-only instrument

Any root works as long as it has the checkpoints/ + phase2_config.json layout above.