Hugging Face's logo

Datasets:
legexbenchmark
/
code

Languages:
English
ArXiv:
Tags:
legal
benchmark
llm-evaluation
information-extraction
reproducibility
License:
Dataset card Files
xet
 Community
code / HUGGINGFACE_UPLOAD.md
anonymous
[code] Reproduction bundle.
2e511b5
preview code
|
Raw
History Blame Contribute Delete
9.28 kB

How to publish LEGEX to Hugging Face anonymously

This walkthrough takes the contents of submission/ and publishes three public Hugging Face dataset repositories under the anonymous account legexbenchmark:

  1. legexbenchmark/goldensets — expert-coded gold annotations (19 jurisdictions, 1,548 judgments / 1,974 rows incl. secondary re-annotations).
  2. legexbenchmark/inference-results — Harvey (2 runs) / Gemini / GPT / Legora (2 runs) outputs (19 jurisdictions × 6 runs = 114 files, JSONL) plus the per-column prompts of the commercial runs (prompts/, 4 files).
  3. legexbenchmark/code — Python pipeline (scrape → process → infer → clean → evaluate).

All data files are JSONL (one JSON object per line). Inference was CSV in the first release and is now JSONL — see Updating an existing release below.

0. Updating an existing release (goldensets + inference-results)

The three repos already exist, and local clones live in ../hf-uploads/{goldensets,inference-results,code}. The staged bundle (submission/goldensets/, submission/inference-results/) mirrors the two data repos 1:1: data + the per-repo dataset card (README.md).

Rebuild the bundle only if the underlying data changed (convert_goldenset_to_jsonl.py needs the local-only data/reannotation/annotators.json + ANNOTATOR_SALT from .env):

uv run python submission/convert_goldenset_to_jsonl.py --data-dir data --out-dir submission/goldensets/data
(cd submission && uv run python build_inference_jsonl.py --data-dir ../data --gold-dir goldensets/data --out-dir inference-results/data)

When the working trees in ../hf-uploads/ are already synced from submission/ (data + README copied in, stale CSVs removed), pushing the update is:

# Log in once with the anonymous account's WRITE token.
pip install --upgrade huggingface_hub && git lfs install --skip-repo
huggingface-cli login                 # paste the hf_… write token
huggingface-cli whoami                # must print: legexbenchmark

# --- goldensets ---
cd ../hf-uploads/goldensets
git status && git diff --stat         # review: data/<cc>/*.jsonl + README.md
git config user.name anonymous && git config user.email anonymous@anonymous.invalid
git add -A
git commit -m "Update: re-annotations, Nepal/Singapore extensions, traceability fields, refreshed card"
git push
cd -

# --- inference-results (CSV -> JSONL) ---
cd ../hf-uploads/inference-results
git status                            # review: CSVs deleted, 57 JSONL added, README.md, .gitattributes
git config user.name anonymous && git config user.email anonymous@anonymous.invalid
git add -A
git commit -m "Switch inference to JSONL; 19 jurisdictions; cleaning provenance; refreshed card"
git push
cd -

Then verify usability (the old CSV raised ArrowInvalid):

from datasets import load_dataset
load_dataset("legexbenchmark/goldensets", split="brazil")
load_dataset("legexbenchmark/inference-results", "harvey", split="united_states")
load_dataset("legexbenchmark/inference-results", "gpt", split="switzerland")

Also confirm on the web UI that every commit author reads anonymous, the dataset viewer renders both repos, and no inference_*.csv remain in inference-results.

code is updated separately once the cleaned analysis package is staged — see item 3 of CHANGES_SINCE_SUBMISSION.md.

1. Create the anonymous Hugging Face account (once)

  1. Open a private / incognito browser window so cookies from any existing HF account are isolated.
  2. Visit https://huggingface.co/join.
  3. Sign up with a fresh email address that is not tied to any of the authors (a proton.me / tutanota.com mailbox works fine).
  4. Pick the username legexbenchmark so the public URLs match the names used in the paper.
  5. Verify the email and log in.

Why a fresh account? HF attaches your account's display name, avatar and email to commits in each repo's history. Reusing an existing account would leak your identity even if every file in the repo is clean.

2. Create a write-scoped access token (once)

  1. Visit https://huggingface.co/settings/tokens.
  2. Click New token, name it upload, role Write, click Generate token, copy the value (it starts with hf_).

3. Install the tooling and log in (once) 

pip install --upgrade huggingface_hub
git lfs install --skip-repo   # macOS: brew install git-lfs first

# Paste the write token created in step 2 when prompted.
huggingface-cli login

4. Push the three repositories (initial release)

Run steps 4a → 4c in order. Each step is independent — if something breaks in 4b you can fix it without redoing 4a.

4a. Push legexbenchmark/goldensets 

# 1. Create the empty repo via the web UI:
#    https://huggingface.co/new-dataset
#    Owner: legexbenchmark | Name: goldensets | License: MIT | Public

# 2. Clone, anonymise local git, copy in the staged files.
git clone https://huggingface.co/datasets/legexbenchmark/goldensets hf-goldensets
cd hf-goldensets
git config user.name  "anonymous"
git config user.email "anonymous@anonymous.invalid"
git lfs track "*.jsonl"
git add .gitattributes

cp -R ../submission/goldensets/. .

# 3. Commit and push.
git add .
git commit -m "Initial release"
git push -u origin main
cd ..

Open https://huggingface.co/datasets/legexbenchmark/goldensets. Verify:

  • The dataset card renders with the jurisdiction tables.
  • data/<cc>/goldenset_<cc>.jsonl is browsable for all 19 jurisdictions.
  • The author next to every commit is anonymous.

4b. Push legexbenchmark/inference-results 

# 1. Create the empty repo:
#    https://huggingface.co/new-dataset
#    Owner: legexbenchmark | Name: inference-results | License: MIT | Public

# 2. Clone, anonymise, copy in.
git clone https://huggingface.co/datasets/legexbenchmark/inference-results hf-inference
cd hf-inference
git config user.name  "anonymous"
git config user.email "anonymous@anonymous.invalid"
git lfs track "*.jsonl"
git add .gitattributes

cp -R ../submission/inference-results/. .

# 3. Commit and push.
git add .
git commit -m "Initial release"
git push -u origin main
cd ..

Verify on https://huggingface.co/datasets/legexbenchmark/inference-results:

  • 19 data/<cc>/ folders (all jurisdictions).
  • Each folder has inference_harvey.jsonl, inference_gemini.jsonl, inference_gpt.jsonl (no .csv).

4c. Push legexbenchmark/code 

# 1. Create the empty repo:
#    https://huggingface.co/new-dataset
#    Owner: legexbenchmark | Name: code | License: MIT | Public
#    (HF dataset repos accept arbitrary files — we use "dataset" type as
#    the most permissive container for source code.)

# 2. Clone, anonymise, copy in the cleaned package (see
#    CHANGES_SINCE_SUBMISSION.md item 3 for what goes into the code repo).
git clone https://huggingface.co/datasets/legexbenchmark/code hf-code
cd hf-code
git config user.name  "anonymous"
git config user.email "anonymous@anonymous.invalid"

# 3. Commit and push.
git add .
git commit -m "Initial release"
git push -u origin main
cd ..

Verify on https://huggingface.co/datasets/legexbenchmark/code:

  • The repo card lists the CLI entrypoints and the workflow.
  • legex/scrapers/ contains 19 jurisdiction modules.

5. Cross-verify the org page

Open https://huggingface.co/legexbenchmark in a fresh tab — you should see three pinned datasets. Click each card and confirm:

  • The README renders correctly.
  • The "Files and versions" tab is browsable.
  • No identifying information appears next to commits (everything should read anonymous).

Also run one final string search per repo via the file viewer for your own surname, your affiliation, and your work email domain:

<your-surname>       # must return nothing
<your-affiliation>   # must return nothing
<your-email-domain>  # must return nothing

6. Share with reviewers

The single URL that ties the three repos together is https://huggingface.co/legexbenchmark. Add it to the supplementary material of your OpenReview submission; reviewers can browse all three datasets from that org page without logging into HF.

Troubleshooting

A commit shows your real name despite the local config. Run git log -1 --pretty='%an <%ae>' inside the clone. If it still shows your real identity, redo git config user.name / user.email and then git commit --amend --reset-author before pushing.

HF complains about LFS quota. Free accounts get 1 GB of LFS storage; this release uses ~30 MB so quota is not a concern.

A file exceeds 5 GB. None of the files in this release do. If a future addition does, git lfs track the path and recommit.

You want to delete a repo. Each repo's Settings page has "Delete this dataset repository" — that removes files and history.

You need to re-push after a fix. cd into the local clone, edit, commit with --reset-author if you also re-ran git config, then git push.

Reviewers ask for a single zip. HF's "Use this dataset" → "Repository size" → "Download repository" gives a .tar.gz of the whole repo.