ARR Report Generator

HTML dashboards for OpenReview ARR/ACL venues for Senior Area Chairs, Area Chairs, and Program Chairs. Covers both the review phase and the commitment phase.

Runs entirely on your machine. Your OpenReview credentials are only used to fetch data and are never stored or transmitted elsewhere.

📄 See the doc page for a simple walkthrough, or use the README below for the full reference.

---

Screenshots

---

Installation

git clone https://github.com/mtreviso/arr-report-generator
cd arr-report-generator
pip install -r requirements.txt

Requires Python 3.9+.

---

Usage

Review Report (ARR venue)

python generate_review_report.py \
  --username "your@email.com" \
  --password "yourpassword" \
  --venue-id "aclweb.org/ACL/ARR/2026/January" \
  --me "~Your_Name1"

--role sac (default): only papers in your SAC batch.
--role ac: papers where you are AC. --role pc: all papers in the venue (PC/test mode, may take several minutes).

Commitment Report (ACL/EMNLP/etc. venue)

Add the --phase commitment flag to fetch the commitment phase data instead of the review phase. Note that SACs are now ACs, so the --role ac flag is required to see any papers at all.

python generate_review_report.py \
  --username "your@email.com" \
  --password "yourpassword" \
  --venue-id "aclweb.org/ACL/2026/Conference" \
  --me "~Your_Name1" \
  --phase "commitment" \
  --role "ac"

If you know the ARR venue ID that commitment papers link back to, pass --linked-venue-id for significantly faster data fetching (bulk pre-fetches linked ARR data instead of per-paper lookups):

python generate_review_report.py \
  ... \
  --phase "commitment" \
  --linked-venue-id "aclweb.org/ACL/ARR/2026/February"

PC Dashboard (ARR review phase)

Fetches all papers across all SAC batches. Intended for Program Chairs. May take several minutes for large venues (6000+ papers).

python generate_pc_report.py \
  --username "your@email.com" \
  --password "yourpassword" \
  --venue-id "aclweb.org/ACL/ARR/2026/January" \
  --me "~Your_Name1"

PC Commitment Dashboard (ACL/EMNLP/etc. commitment phase)

generate_pc_report.py also supports the commitment phase via --phase commitment. This generates the full PC dashboard across all linked commitment submissions and writes pc_commitment_report.html.

python generate_pc_report.py \
  --username "your@email.com" \
  --password "yourpassword" \
  --venue-id "aclweb.org/ACL/2026/Conference" \
  --me "~Your_Name1" \
  --phase "commitment" \
  --linked-venue-id "aclweb.org/ACL/ARR/2026/January"

Credentials via environment variables

All scripts read these env vars if flags are not supplied:

export OPENREVIEW_USERNAME="your@email.com"
export OPENREVIEW_PASSWORD="yourpassword"
export OPENREVIEW_ID="~Your_Name1"
python generate_review_report.py --venue-id aclweb.org/ACL/ARR/2025/February

---

Common options

Flag	Description
--username / --password	OpenReview credentials (or set via OPENREVIEW_USERNAME / OPENREVIEW_PASSWORD env vars).
--venue-id	Full venue group ID, e.g. aclweb.org/ACL/ARR/2025/February.
--me	Your OpenReview tilde ID, e.g. ~Your_Name1 (or set via OPENREVIEW_ID).
--role	sac (default for review) / ac / pc — scope of papers to fetch.
--phase	review (default) / commitment — which phase to generate a report for.
--linked-venue-id	ARR venue ID for bulk pre-fetching linked submissions (commitment phase only).
--comments-level	none / basic / full (default) — higher detail means slower runs and larger files.
--output-dir	Output directory for generated reports. Default: ./reports.
--append-date	Append today's date (YYYY-MM-DD) to the output filename to avoid overwriting.
--save-cache / --use-cache	Cache OpenReview data to disk; reload instantly on the next run.
--cache-dir	Custom directory for cache files. Auto-generated under .dev_cache/ if omitted.
--impersonate [GROUP_ID]	Lets PCs inspect another SAC's view. Requires PC-level permission.

Comments levels

All report scripts support --comments-level none|basic|full.

• none: omit comments tab and skip comment processing
• basic: direct replies only
• full (default): full reply threads — slowest / largest output

Output files

Script	Phase	Output
generate_review_report.py	review	reports/<venue>_review_report.html
generate_review_report.py	commitment	reports/<venue>_commitment_report.html
generate_pc_report.py	review	reports/<venue>_pc_report.html
generate_pc_report.py	commitment	reports/<venue>_pc_commitment_report.html

Open in any browser — fully self-contained, no server needed.

---

Developer Cache (skip the slow fetch)

For iterating on templates locally, you can cache all OpenReview data to disk after one full run and reload it instantly on subsequent runs.

# Step 1: full fetch + save cache (do this once)
python generate_review_report.py ... --save-cache

# Step 2: use cache for all subsequent runs (seconds, not minutes)
python generate_review_report.py ... --use-cache

--cache-dir controls where the pickle files go. If you omit it, the tool auto-generates a phase- and report-aware cache path under .dev_cache/ so review vs commitment and review vs PC runs do not collide. You can still maintain multiple named caches manually, e.g. one per SAC:

python generate_review_report.py ... --impersonate "~SAC_Name1" --save-cache --cache-dir .cache_sac1
python generate_review_report.py ... --use-cache --cache-dir .cache_sac1

The cache stores three files: submissions.pkl, group_index.pkl, and processed.pkl. By default, cache directories are automatically separated by report type, phase, venue, and role/impersonation where relevant.

--save-cache and --use-cache are mutually exclusive.

Impersonation (view as a SAC)

The review and commitment reports support impersonating another OpenReview user before fetching data. This requires PC-level permission on the venue and uses the OpenReview impersonate API.

python generate_review_report.py ... --phase "review" --me "~Target_SAC_Name1" --impersonate
python generate_review_report.py ... --phase "commitment" --me "~Target_SAC_Name1" --impersonate

You can pass a GROUP_ID after the --impersonate flag, which corresponds to the authorization group, e.g., aclweb.org/ACL/ARR/2025/October/Program_Chairs. If GROUP_ID is empty, it will assume to be <venue-id>/Program_Chairs. Your own credentials are used to authenticate, and all data is then fetched as the target user.

This is useful for verifying exactly what a SAC sees without asking them to run the script themselves. Can be combined with --save-cache / --use-cache for fast iteration.

Impersonation is intentionally not available in generate_pc_report.py, which already fetches all papers by design.

Configuring the low-confidence threshold

The low-confidence flag triggers when any reviewer on a paper has a confidence score at or below the threshold. The default is 2. To change it, edit the class variable near the top of arr_report_generator.py:

class ARRReportGenerator:
    LOW_CONF_THRESHOLD = 2   # ← change this

---

Credits

• Based on the ARR Tool by Yiming Cui

License

MIT