bravo-pena/is-identifier-1.0

IS Identifier 1.2

IS Identifier 1.2 identifies institutional statements in regulatory sentences. It predicts AIM spans using a BIO token-classification head and derives the suggested AIM count from the decoded spans.

This repository keeps its original id (is-identifier-1.0) for continuity; it hosts the current model version, 1.2 (see training_config.json and the validation table below). Earlier revisions remain available in the repo history.

The model is intended to be used with the companion Python package (version 1.2, Paso 1). Given a PDF, Word .docx, Markdown, or TXT file, the package exports a reviewable Excel of structure-aware segments with AIM candidates, a substantive-context filter and human-review flags (needs_review / review_reason). Paso 1 proposes candidates for human coders — the taxonomic classification (TYPE / TAXON / LINK) belongs to a separate future tool (Paso 2) and is never part of this output.

Usage

from is_identifier import pipeline_paso1, write_paso1_excel
from is_identifier.model_v1 import ISIdentifierModel

model = ISIdentifierModel.from_pretrained("bravo-pena/is-identifier-1.0")
df, technical = pipeline_paso1("regulation.pdf", aim_model=model, language="es")
write_paso1_excel(df, "regulation_paso1.xlsx", technical)

Command line:

is-identifier regulation.pdf \
  --model bravo-pena/is-identifier-1.0 \
  --language es \
  --output regulation_paso1.xlsx

Outputs generated with package versions prior to the 2026-06-10 tokenizer fix must not be used to evaluate AIM candidates — regenerate them.

Expected Files In This Model Repository

• model.safetensors
• config.json
• tokenizer.json
• tokenizer_config.json
• training_config.json, if available

Validation

Version 1.2 is an interim retrain of the 1.0 recipe on the annotation base after the June 2026 correction round (verified label fixes from the coding team; same architecture and hyper-parameters). Figures are provisional: a further label-review round (double coding + list-article convention) is in progress and the metrics will be re-frozen with the final base.

Leave-one-regulation-out cross validation, 14 folds:

Metric	1.2 (provisional)	1.0 baseline
count_macro_f1	0.556	0.5345
span_f1_partial	0.6925	0.6723
recall_aim0	0.6448	0.5739
recall_aim_ge1	0.9316	0.9225

The validation data is private and is not included in this model repository.

Limitations

• Validated for Spanish and English regulatory-style text.
• Legacy .doc files should be converted to .docx before processing.
• The model does not assign institutional TYPE or TAXON.
• This model supports coding and audit workflows; it is not a legal advisor.

License

MIT.