R01G-SR

Research code for R01G-SR: Provenance-Preserving Selection over Complementary Evidence Views for Crisis Situation Report Generation, LT4CPR at AACL-IJCNLP 2026.

Two report candidates share a generator: R0 presents chronological evidence, while R1 filters and compresses evidence. A frozen logistic selector chooses one whole report; deterministic normalization, section-ownership repair, and evidence sanitation enforce structural constraints. Selection does not synthesize new claims or combine reports.

Release scope

This initial source release contains candidate generation, feature extraction, grouped selector training and application, structure repair, and submission sanitation. It is an extraction of existing research code, not a new implementation or a newly evaluated model. SOURCE_MANIFEST.json records the original source-file SHA-256 values.

Data, social-media messages, gold reports, generated reports, model weights, experiment logs, credentials, and historical evaluation environments are not redistributed. Frozen selector coefficients are not included in this initial release; train a selector using permitted training data and cached training scores. This is not yet a one-command reproduction of the submitted ZIP. Historical scoring helpers are retained as dependencies of selector training; for current benchmark results use the organizer's released evaluator, not the historical wrapper as a substitute.

Setup

Python 3.11 or newer is recommended. Install requirements.txt in a separate environment. The listed dependency ranges are compatibility requirements, not a lockfile of the historical environment.

Obtain task data and evaluation tools from the official task repository, respecting its terms. The recorded test-release snapshot is 78dbff586d391d97ec4d11310cf587f43c57ca2a; earlier research runs used earlier snapshots.

Obtain the Qwen3-4B-Instruct-2507 model separately and serve it with an OpenAI-compatible local endpoint. Principal experiments used a UD-Q4_K_XL GGUF conversion. No model download, server startup, training, or GPU work is performed by the smoke tests.

Candidate generation

From the repository root, with released cell files under data/dev and an already running local inference server:

python -m src.run --data data/dev --out outputs/dev --runs R0,R1 --base-url http://127.0.0.1:8080/v1 --model Qwen3-4B-Instruct-2507
python -m src.postprocess_outputs --source outputs/dev --out outputs/normalized --runs R0,R1
python -m src.build_selector_metadata --data data/dev --source outputs/normalized --r0-name R0 --r1-name R1 --out outputs/metadata

src.run also retains historical R2-R4 options; they are not the submitted primary method. Do not use its default route list when intending only R0/R1.

Selector

See CLI help for the training-score payload and metadata inputs:

python -m src.train_grouped_selector --help
python -m src.apply_grouped_selector --help

Train only on permitted training scores, grouping all windows and replicates from each crisis together. Never use hidden-test references to fit or select a model. Apply a frozen selector with:

python -m src.apply_grouped_selector --data data/dev --model artifacts/selector.json --selector-metadata outputs/metadata --source outputs/normalized --out outputs/selected

src.repair_section_ownership.repair(report) returns a repaired copy and a move log. It moves only subsection titles with exactly one legal owner; ambiguous titles stay unchanged. Apply it before final sanitation. Sanitation and validation are available through:

python -m src.prepare_submission sanitize --data data/dev --reports outputs/repaired --out outputs/clean
python -m src.prepare_submission validate --data data/dev --reports outputs/clean

The historical packaging subcommand is not a substitute for the latest official team-ZIP validator. Follow the organizer's current submission specification.

Checks

python -m unittest discover -s tests -v
python -m compileall -q src tests

These checks cover source compilation and synthetic structural invariants only, not model quality or end-to-end benchmark reproduction.

Official result

Team gogogo, primary system r01g-sr: official PRIMARY 0.5273, BERTScore F1 0.5018, BLEURT 0.5528; sixth of nine teams. PRIMARY is the mean of BERTScore F1 and BLEURT. These are organizer-reported historical scores, not scores produced by running this release's checks. See the official leaderboard.

Responsible use

Evidence IDs support traceability, not truth verification. This is research software, not an operational crisis-response system. Human review is required before consequential use. Third-party data and model licenses remain applicable. No additional software license is selected in this initial publication.