Skip to content

NOTES.md

Latest commit

 

History

History
59 lines (47 loc) · 2.39 KB

NOTES.md

File metadata and controls

59 lines (47 loc) · 2.39 KB

Notes

Scratchpad for ideas and follow-ups that are not yet decisions. Excluded from the built package via .Rbuildignore.

Publishing the merged SPECTER2 model to HuggingFace Hub

Currently, users must run inst/scripts/prepare_specter2_merged.py themselves to produce the merged specter2_base + proximity-adapter model directory that TEI can serve. This requires a Python environment with transformers + adapters, ~500 MB on disk, and a few minutes of compute.

If SPECTER2 becomes a first-class supported backend, the cleaner path is to publish the merged model to HuggingFace Hub once and let users pull it directly.

Why not bundle the weights in the package

  • CRAN size limit is 5 MB; merged model is ~500 MB. Hard no.
  • Off-CRAN bundling still bloats git history, slows install_local(), and couples model release cadence to code release cadence. R packages are source code, not weight stores.

Proposed workflow

  1. Run the merge once locally.
  2. Publish: 
    huggingface-cli login
huggingface-cli upload <org>/specter2_proximity_merged \
  ~/Library/Caches/org.R-project.R/R/openalexVectorComp/specter2_proximity_merged
  3. Users skip the merge entirely: 
    text-embeddings-router --model-id <org>/specter2_proximity_merged --port 8080
  4. inst/scripts/prepare_specter2_merged.py stays as a reproducibility artifact (how we built it) but is no longer required for setup.
  5. inst/scripts/start_tei_specter2.sh defaults --model-id to the published repo when no local merged model is found.
  6. backend_specter2_tei() default model arg updates to the published id.
  7. Vignette specter2-setup.qmd collapses to: install TEI, run one command, use from R.

Open questions

  • Which HF org/account hosts the published model? Personal vs project org.
  • License compatibility — SPECTER2 base + adapter licenses (verify Apache-2.0 on both before redistributing the merged weights).
  • Versioning: tag the HF repo on each re-merge; pin the tag in backend_specter2_tei() for reproducibility.

Middle-ground alternative

If publishing to HF Hub is undesirable, host the merged tarball on GitHub Releases (2 GB per-asset limit) or S3, and have an R function lazy-download into tools::R_user_dir("openalexVectorComp", "cache") on first use. Packages like piggyback automate the GitHub Releases case. Keeps weights out of the package source either way.